repo
stringlengths
2
99
file
stringlengths
13
225
code
stringlengths
0
18.3M
file_length
int64
0
18.3M
avg_line_length
float64
0
1.36M
max_line_length
int64
0
4.26M
extension_type
stringclasses
1 value
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/CIF_2016/non_moving_window/without_stl_decomposition/create_o12_tfrecords.py
from tfrecords_handler.non_moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../../datasets/binary_data/CIF_2016/non_moving_window/without_stl_decomposition/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': tfrecord_writer = TFRecordWriter( output_size = 12, train_file_path='../../../../datasets/text_data/CIF_2016/non_moving_window/without_stl_decomposition/cif_12.txt', validate_file_path='../../../../datasets/text_data/CIF_2016/non_moving_window/without_stl_decomposition/cif_12v.txt', test_file_path='../../../../datasets/text_data/CIF_2016/non_moving_window/without_stl_decomposition/cif_test_12.txt', binary_train_file_path=output_path + 'cif_12.tfrecords', binary_validation_file_path=output_path + 'cif_12v.tfrecords', binary_test_file_path=output_path + 'cif_test_12.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
1,143
51
125
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/CIF_2016/non_moving_window/without_stl_decomposition/create_o6_tfrecords.py
from tfrecords_handler.non_moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../../datasets/binary_data/CIF_2016/non_moving_window/without_stl_decomposition/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': tfrecord_writer = TFRecordWriter( output_size = 6, train_file_path = '../../../../datasets/text_data/CIF_2016/non_moving_window/without_stl_decomposition/cif_6.txt', validate_file_path = '../../../../datasets/text_data/CIF_2016/non_moving_window/without_stl_decomposition/cif_6v.txt', test_file_path = '../../../../datasets/text_data/CIF_2016/non_moving_window/without_stl_decomposition/cif_test_6.txt', binary_train_file_path = output_path + 'cif_6.tfrecords', binary_validation_file_path = output_path + 'cif_6v.tfrecords', binary_test_file_path = output_path + 'cif_test_6.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
1,148
51.227273
126
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/NN5/moving_window/create_tfrecords.py
from tfrecords_handler.moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../datasets/binary_data/NN5/moving_window/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': tfrecord_writer = TFRecordWriter( input_size = 9, output_size = 56, train_file_path = '../../../datasets/text_data/NN5/moving_window/nn5_stl_56i9.txt', validate_file_path = '../../../datasets/text_data/NN5/moving_window/nn5_stl_56i9v.txt', test_file_path = '../../../datasets/text_data/NN5/moving_window/nn5_test_56i9.txt', binary_train_file_path = output_path + 'nn5_stl_56i9.tfrecords', binary_validation_file_path = output_path + 'nn5_stl_56i9v.tfrecords', binary_test_file_path = output_path + 'nn5_test_56i9.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
1,051
44.73913
95
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/NN5/moving_window/without_stl_decomposition/create_tfrecords.py
from tfrecords_handler.moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../../datasets/binary_data/NN5/moving_window/without_stl_decomposition/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': tfrecord_writer = TFRecordWriter( input_size = 9, output_size = 56, train_file_path = '../../../../datasets/text_data/NN5/moving_window/without_stl_decomposition/nn5_56i9.txt', validate_file_path = '../../../../datasets/text_data/NN5/moving_window/without_stl_decomposition/nn5_56i9v.txt', test_file_path = '../../../../datasets/text_data/NN5/moving_window/without_stl_decomposition/nn5_test_56i9.txt', binary_train_file_path = output_path + 'nn5_56i9.tfrecords', binary_validation_file_path = output_path + 'nn5_56i9v.tfrecords', binary_test_file_path = output_path + 'nn5_test_56i9.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
1,151
49.086957
120
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/NN5/non_moving_window/create_tfrecords.py
from tfrecords_handler.non_moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../datasets/binary_data/NN5/non_moving_window/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': tfrecord_writer = TFRecordWriter( output_size = 56, train_file_path = '../../../datasets/text_data/NN5/non_moving_window/nn5_stl_56.txt', validate_file_path = '../../../datasets/text_data/NN5/non_moving_window/nn5_stl_56v.txt', test_file_path = '../../../datasets/text_data/NN5/non_moving_window/nn5_test_56.txt', binary_train_file_path = output_path + 'nn5_stl_56.tfrecords', binary_validation_file_path = output_path + 'nn5_stl_56v.tfrecords', binary_test_file_path = output_path + 'nn5_test_56.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
1,035
46.090909
97
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/NN5/non_moving_window/without_stl_decomposition/create_tfrecords.py
from tfrecords_handler.non_moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../../datasets/binary_data/NN5/non_moving_window/without_stl_decomposition/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': tfrecord_writer = TFRecordWriter( output_size = 56, train_file_path = '../../../../datasets/text_data/NN5/non_moving_window/without_stl_decomposition/nn5_56.txt', validate_file_path = '../../../../datasets/text_data/NN5/non_moving_window/without_stl_decomposition/nn5_56v.txt', test_file_path = '../../../../datasets/text_data/NN5/non_moving_window/without_stl_decomposition/nn5_test_56.txt', binary_train_file_path = output_path + 'nn5_56.tfrecords', binary_validation_file_path = output_path + 'nn5_56v.tfrecords', binary_test_file_path = output_path + 'nn5_test_56.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
1,135
50.636364
122
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/kaggle_web_traffic/moving_window/create_tfrecords.py
from tfrecords_handler.moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../datasets/binary_data/kaggle_web_traffic/moving_window/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': tfrecord_writer = TFRecordWriter( input_size = 9, output_size = 59, train_file_path = '../../../datasets/text_data/kaggle_web_traffic/moving_window/kaggle_stl_59i9.txt', validate_file_path = '../../../datasets/text_data/kaggle_web_traffic/moving_window/kaggle_stl_59i9v.txt', test_file_path = '../../../datasets/text_data/kaggle_web_traffic/moving_window/kaggle_test_59i9.txt', binary_train_file_path = output_path + 'kaggle_stl_59i9.tfrecords', binary_validation_file_path = output_path + 'kaggle_stl_59i9v.tfrecords', binary_test_file_path = output_path + 'kaggle_test_59i9.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() tfrecord_writer = TFRecordWriter( input_size=74, output_size=59, train_file_path='../../../datasets/text_data/kaggle_web_traffic/moving_window/kaggle_stl_59i74.txt', validate_file_path='../../../datasets/text_data/kaggle_web_traffic/moving_window/kaggle_stl_59i74v.txt', test_file_path='../../../datasets/text_data/kaggle_web_traffic/moving_window/kaggle_test_59i74.txt', binary_train_file_path=output_path + 'kaggle_stl_59i74.tfrecords', binary_validation_file_path=output_path + 'kaggle_stl_59i74v.tfrecords', binary_test_file_path=output_path + 'kaggle_test_59i74.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
1,996
50.205128
113
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/kaggle_web_traffic/moving_window/without_stl_decomposition/create_tfrecords.py
from tfrecords_handler.moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../../datasets/binary_data/kaggle_web_traffic/moving_window/without_stl_decomposition/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': tfrecord_writer = TFRecordWriter( input_size = 9, output_size = 59, train_file_path = '../../../../datasets/text_data/kaggle_web_traffic/moving_window/without_stl_decomposition/kaggle_59i9.txt', validate_file_path = '../../../../datasets/text_data/kaggle_web_traffic/moving_window/without_stl_decomposition/kaggle_59i9v.txt', test_file_path = '../../../../datasets/text_data/kaggle_web_traffic/moving_window/without_stl_decomposition/kaggle_test_59i9.txt', binary_train_file_path = output_path + 'kaggle_59i9.tfrecords', binary_validation_file_path = output_path + 'kaggle_59i9v.tfrecords', binary_test_file_path = output_path + 'kaggle_test_59i9.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() tfrecord_writer = TFRecordWriter( input_size=74, output_size=59, train_file_path='../../../../datasets/text_data/kaggle_web_traffic/moving_window/without_stl_decomposition/kaggle_59i74.txt', validate_file_path='../../../../datasets/text_data/kaggle_web_traffic/moving_window/without_stl_decomposition/kaggle_59i74v.txt', test_file_path='../../../../datasets/text_data/kaggle_web_traffic/moving_window/without_stl_decomposition/kaggle_test_59i74.txt', binary_train_file_path=output_path + 'kaggle_59i74.tfrecords', binary_validation_file_path=output_path + 'kaggle_59i74v.tfrecords', binary_test_file_path=output_path + 'kaggle_test_59i74.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
2,163
54.487179
138
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/kaggle_web_traffic/non_moving_window/create_tfrecords.py
from tfrecords_handler.non_moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../datasets/binary_data/kaggle_web_traffic/non_moving_window/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': tfrecord_writer = TFRecordWriter( output_size = 59, train_file_path = '../../../datasets/text_data/kaggle_web_traffic/non_moving_window/kaggle_stl_59.txt', validate_file_path = '../../../datasets/text_data/kaggle_web_traffic/non_moving_window/kaggle_stl_59v.txt', test_file_path = '../../../datasets/text_data/kaggle_web_traffic/non_moving_window/kaggle_test_59.txt', binary_train_file_path = output_path + 'kaggle_stl_59.tfrecords', binary_validation_file_path = output_path + 'kaggle_stl_59v.tfrecords', binary_test_file_path = output_path + 'kaggle_test_59.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
1,113
49.636364
115
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/kaggle_web_traffic/non_moving_window/without_stl_decomposition/create_tfrecords.py
from tfrecords_handler.non_moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../../datasets/binary_data/kaggle_web_traffic/non_moving_window/without_stl_decomposition/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': tfrecord_writer = TFRecordWriter( output_size = 59, train_file_path = '../../../../datasets/text_data/kaggle_web_traffic/non_moving_window/without_stl_decomposition/kaggle_59.txt', validate_file_path = '../../../../datasets/text_data/kaggle_web_traffic/non_moving_window/without_stl_decomposition/kaggle_59v.txt', test_file_path = '../../../../datasets/text_data/kaggle_web_traffic/non_moving_window/without_stl_decomposition/kaggle_test_59.txt', binary_train_file_path = output_path + 'kaggle_59.tfrecords', binary_validation_file_path = output_path + 'kaggle_59v.tfrecords', binary_test_file_path = output_path + 'kaggle_test_59.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
1,217
54.363636
140
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/M4/moving_window/create_tfrecords.py
from tfrecords_handler.moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../datasets/binary_data/M4/moving_window/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': # macro data tfrecord_writer = TFRecordWriter( input_size = 15, output_size = 18, train_file_path = '../../../datasets/text_data/M4/moving_window/m4_stl_monthly_macro_18i15.txt', validate_file_path = '../../../datasets/text_data/M4/moving_window/m4_stl_monthly_macro_18i15v.txt', test_file_path = '../../../datasets/text_data/M4/moving_window/m4_test_monthly_macro_18i15.txt', binary_train_file_path = output_path + 'm4_stl_monthly_macro_18i15.tfrecords', binary_validation_file_path = output_path + 'm4_stl_monthly_macro_18i15v.tfrecords', binary_test_file_path = output_path + 'm4_test_monthly_macro_18i15.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # micro data tfrecord_writer = TFRecordWriter( input_size=15, output_size=18, train_file_path='../../../datasets/text_data/M4/moving_window/m4_stl_monthly_micro_18i15.txt', validate_file_path='../../../datasets/text_data/M4/moving_window/m4_stl_monthly_micro_18i15v.txt', test_file_path='../../../datasets/text_data/M4/moving_window/m4_test_monthly_micro_18i15.txt', binary_train_file_path=output_path + 'm4_stl_monthly_micro_18i15.tfrecords', binary_validation_file_path=output_path + 'm4_stl_monthly_micro_18i15v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_micro_18i15.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # industry data tfrecord_writer = TFRecordWriter( input_size=15, output_size=18, train_file_path='../../../datasets/text_data/M4/moving_window/m4_stl_monthly_industry_18i15.txt', validate_file_path='../../../datasets/text_data/M4/moving_window/m4_stl_monthly_industry_18i15v.txt', test_file_path='../../../datasets/text_data/M4/moving_window/m4_test_monthly_industry_18i15.txt', binary_train_file_path=output_path + 'm4_stl_monthly_industry_18i15.tfrecords', binary_validation_file_path=output_path + 'm4_stl_monthly_industry_18i15v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_industry_18i15.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # finance data tfrecord_writer = TFRecordWriter( input_size=15, output_size=18, train_file_path='../../../datasets/text_data/M4/moving_window/m4_stl_monthly_finance_18i15.txt', validate_file_path='../../../datasets/text_data/M4/moving_window/m4_stl_monthly_finance_18i15v.txt', test_file_path='../../../datasets/text_data/M4/moving_window/m4_test_monthly_finance_18i15.txt', binary_train_file_path=output_path + 'm4_stl_monthly_finance_18i15.tfrecords', binary_validation_file_path=output_path + 'm4_stl_monthly_finance_18i15v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_finance_18i15.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # other data tfrecord_writer = TFRecordWriter( input_size=5, output_size=18, train_file_path='../../../datasets/text_data/M4/moving_window/m4_stl_monthly_other_18i5.txt', validate_file_path='../../../datasets/text_data/M4/moving_window/m4_stl_monthly_other_18i5v.txt', test_file_path='../../../datasets/text_data/M4/moving_window/m4_test_monthly_other_18i5.txt', binary_train_file_path=output_path + 'm4_stl_monthly_other_18i5.tfrecords', binary_validation_file_path=output_path + 'm4_stl_monthly_other_18i5v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_other_18i5.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # demographic data tfrecord_writer = TFRecordWriter( input_size=15, output_size=18, train_file_path='../../../datasets/text_data/M4/moving_window/m4_stl_monthly_demo_18i15.txt', validate_file_path='../../../datasets/text_data/M4/moving_window/m4_stl_monthly_demo_18i15v.txt', test_file_path='../../../datasets/text_data/M4/moving_window/m4_test_monthly_demo_18i15.txt', binary_train_file_path=output_path + 'm4_stl_monthly_demo_18i15.tfrecords', binary_validation_file_path=output_path + 'm4_stl_monthly_demo_18i15v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_demo_18i15.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
5,637
50.724771
109
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/M4/moving_window/without_stl_decomposition/create_tfrecords.py
from tfrecords_handler.moving_window.tfrecord_writer import TFRecordWriter import os output_path = "'../../../../datasets/binary_data/M4/moving_window/without_stl_decomposition/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': # macro data tfrecord_writer = TFRecordWriter( input_size = 15, output_size = 18, train_file_path = '../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_monthly_macro_18i15.txt', validate_file_path = '../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_monthly_macro_18i15v.txt', test_file_path = '../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_test_monthly_macro_18i15.txt', binary_train_file_path = output_path + 'm4_monthly_macro_18i15.tfrecords', binary_validation_file_path = output_path + 'm4_monthly_macro_18i15v.tfrecords', binary_test_file_path = output_path + 'm4_test_monthly_macro_18i15.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # micro data tfrecord_writer = TFRecordWriter( input_size=15, output_size=18, train_file_path='../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_monthly_micro_18i15.txt', validate_file_path='../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_monthly_micro_18i15v.txt', test_file_path='../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_test_monthly_micro_18i15.txt', binary_train_file_path=output_path + 'm4_monthly_micro_18i15.tfrecords', binary_validation_file_path=output_path + 'm4_monthly_micro_18i15v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_micro_18i15.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # industry data tfrecord_writer = TFRecordWriter( input_size=15, output_size=18, train_file_path='../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_monthly_industry_18i15.txt', validate_file_path='../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_monthly_industry_18i15v.txt', test_file_path='../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_test_monthly_industry_18i15.txt', binary_train_file_path=output_path + 'm4_monthly_industry_18i15.tfrecords', binary_validation_file_path=output_path + 'm4_monthly_industry_18i15v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_industry_18i15.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # finance data tfrecord_writer = TFRecordWriter( input_size=15, output_size=18, train_file_path='../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_monthly_finance_18i15.txt', validate_file_path='../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_monthly_finance_18i15v.txt', test_file_path='../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_test_monthly_finance_18i15.txt', binary_train_file_path=output_path + 'm4_monthly_finance_18i15.tfrecords', binary_validation_file_path=output_path + 'm4_monthly_finance_18i15v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_finance_18i15.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # other data tfrecord_writer = TFRecordWriter( input_size=5, output_size=18, train_file_path='../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_monthly_other_18i5.txt', validate_file_path='../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_monthly_other_18i5v.txt', test_file_path='../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_test_monthly_other_18i5.txt', binary_train_file_path=output_path + 'm4_monthly_other_18i5.tfrecords', binary_validation_file_path=output_path + 'm4_monthly_other_18i5v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_other_18i5.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # demographic data tfrecord_writer = TFRecordWriter( input_size=15, output_size=18, train_file_path='../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_monthly_demo_18i15.txt', validate_file_path='../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_monthly_demo_18i15v.txt', test_file_path='../../../../datasets/text_data/M4/moving_window/without_stl_decomposition/m4_test_monthly_demo_18i15.txt', binary_train_file_path=output_path + 'm4_monthly_demo_18i15.tfrecords', binary_validation_file_path=output_path + 'm4_monthly_demo_18i15v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_demo_18i15.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
6,093
54.908257
134
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/M4/non_moving_window/create_tfrecords.py
from tfrecords_handler.non_moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../datasets/binary_data/M4/non_moving_window/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': # macro data tfrecord_writer = TFRecordWriter( output_size = 18, train_file_path = '../../../datasets/text_data/M4/non_moving_window/m4_stl_monthly_macro_18.txt', validate_file_path = '../../../datasets/text_data/M4/non_moving_window/m4_stl_monthly_macro_18v.txt', test_file_path = '../../../datasets/text_data/M4/non_moving_window/m4_test_monthly_macro_18.txt', binary_train_file_path = output_path + 'm4_stl_monthly_macro_18.tfrecords', binary_validation_file_path = output_path + 'm4_stl_monthly_macro_18v.tfrecords', binary_test_file_path = output_path + 'm4_test_monthly_macro_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # micro data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../datasets/text_data/M4/non_moving_window/m4_stl_monthly_micro_18.txt', validate_file_path='../../../datasets/text_data/M4/non_moving_window/m4_stl_monthly_micro_18v.txt', test_file_path='../../../datasets/text_data/M4/non_moving_window/m4_test_monthly_micro_18.txt', binary_train_file_path=output_path + 'm4_stl_monthly_micro_18.tfrecords', binary_validation_file_path=output_path + 'm4_stl_monthly_micro_18v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_micro_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # industry data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../datasets/text_data/M4/non_moving_window/m4_stl_monthly_industry_18.txt', validate_file_path='../../../datasets/text_data/M4/non_moving_window/m4_stl_monthly_industry_18v.txt', test_file_path='../../../datasets/text_data/M4/non_moving_window/m4_test_monthly_industry_18.txt', binary_train_file_path=output_path + 'm4_stl_monthly_industry_18.tfrecords', binary_validation_file_path=output_path + 'm4_stl_monthly_industry_18v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_industry_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # finance data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../datasets/text_data/M4/non_moving_window/m4_stl_monthly_finance_18.txt', validate_file_path='../../../datasets/text_data/M4/non_moving_window/m4_stl_monthly_finance_18v.txt', test_file_path='../../../datasets/text_data/M4/non_moving_window/m4_test_monthly_finance_18.txt', binary_train_file_path=output_path + 'm4_stl_monthly_finance_18.tfrecords', binary_validation_file_path=output_path + 'm4_stl_monthly_finance_18v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_finance_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # other data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../datasets/text_data/M4/non_moving_window/m4_stl_monthly_other_18.txt', validate_file_path='../../../datasets/text_data/M4/non_moving_window/m4_stl_monthly_other_18v.txt', test_file_path='../../../datasets/text_data/M4/non_moving_window/m4_test_monthly_other_18.txt', binary_train_file_path=output_path + 'm4_stl_monthly_other_18.tfrecords', binary_validation_file_path=output_path + 'm4_stl_monthly_other_18v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_other_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # demographic data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../datasets/text_data/M4/non_moving_window/m4_stl_monthly_demo_18.txt', validate_file_path='../../../datasets/text_data/M4/non_moving_window/m4_stl_monthly_demo_18v.txt', test_file_path='../../../datasets/text_data/M4/non_moving_window/m4_test_monthly_demo_18.txt', binary_train_file_path=output_path + 'm4_stl_monthly_demo_18.tfrecords', binary_validation_file_path=output_path + 'm4_stl_monthly_demo_18v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_demo_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
5,476
52.174757
110
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/M4/non_moving_window/without_stl_decomposition/create_tfrecords.py
from tfrecords_handler.non_moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../../datasets/binary_data/M4/non_moving_window/without_stl_decomposition/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': # macro data tfrecord_writer = TFRecordWriter( output_size = 18, train_file_path = '../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_monthly_macro_18.txt', validate_file_path = '../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_monthly_macro_18v.txt', test_file_path = '../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_test_monthly_macro_18.txt', binary_train_file_path = output_path + 'm4_monthly_macro_18.tfrecords', binary_validation_file_path = output_path + 'm4_monthly_macro_18v.tfrecords', binary_test_file_path = output_path + 'm4_test_monthly_macro_18.tfrecords', ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # micro data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_monthly_micro_18.txt', validate_file_path='../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_monthly_micro_18v.txt', test_file_path='../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_test_monthly_micro_18.txt', binary_train_file_path=output_path + 'm4_monthly_micro_18.tfrecords', binary_validation_file_path=output_path + 'm4_monthly_micro_18v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_micro_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # industry data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_monthly_industry_18.txt', validate_file_path='../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_monthly_industry_18v.txt', test_file_path='../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_test_monthly_industry_18.txt', binary_train_file_path=output_path + 'm4_monthly_industry_18.tfrecords', binary_validation_file_path=output_path + 'm4_monthly_industry_18v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_industry_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # finance data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_monthly_finance_18.txt', validate_file_path='../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_monthly_finance_18v.txt', test_file_path='../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_test_monthly_finance_18.txt', binary_train_file_path=output_path + 'm4_monthly_finance_18.tfrecords', binary_validation_file_path=output_path + 'm4_monthly_finance_18v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_finance_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # other data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_monthly_other_18.txt', validate_file_path='../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_monthly_other_18v.txt', test_file_path='../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_test_monthly_other_18.txt', binary_train_file_path=output_path + 'm4_monthly_other_18.tfrecords', binary_validation_file_path=output_path + 'm4_monthly_other_18v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_other_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # demographic data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_monthly_demo_18.txt', validate_file_path='../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_monthly_demo_18v.txt', test_file_path='../../../../datasets/text_data/M4/non_moving_window/without_stl_decomposition/m4_test_monthly_demo_18.txt', binary_train_file_path=output_path + 'm4_monthly_demo_18.tfrecords', binary_validation_file_path=output_path + 'm4_monthly_demo_18v.tfrecords', binary_test_file_path=output_path + 'm4_test_monthly_demo_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
5,932
56.601942
135
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/Tourism/moving_window/create_tfrecords.py
from tfrecords_handler.moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../datasets/binary_data/Tourism/moving_window/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': tfrecord_writer = TFRecordWriter( input_size = 15, output_size = 24, train_file_path = '../../../datasets/text_data/Tourism/moving_window/tourism_stl_24i15.txt', validate_file_path = '../../../datasets/text_data/Tourism/moving_window/tourism_stl_24i15v.txt', test_file_path = '../../../datasets/text_data/Tourism/moving_window/tourism_test_24i15.txt', binary_train_file_path = output_path + 'tourism_stl_24i15.tfrecords', binary_validation_file_path = output_path + 'tourism_stl_24i15v.tfrecords', binary_test_file_path = output_path + 'tourism_test_24i15.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
1,098
46.782609
104
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/Tourism/moving_window/without_stl_decomposition/create_tfrecords.py
from tfrecords_handler.moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../../datasets/binary_data/Tourism/moving_window/without_stl_decomposition/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': tfrecord_writer = TFRecordWriter( input_size = 15, output_size = 24, train_file_path = '../../../../datasets/text_data/Tourism/moving_window/without_stl_decomposition/tourism_24i15.txt', validate_file_path = '../../../../datasets/text_data/Tourism/moving_window/without_stl_decomposition/tourism_24i15v.txt', test_file_path = '../../../../datasets/text_data/Tourism/moving_window/without_stl_decomposition/tourism_test_24i15.txt', binary_train_file_path = output_path + 'tourism_24i15.tfrecords', binary_validation_file_path = output_path + 'tourism_24i15v.tfrecords', binary_test_file_path = output_path + 'tourism_test_24i15.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
1,198
51.130435
129
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/Tourism/non_moving_window/create_tfrecords.py
from tfrecords_handler.non_moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../datasets/binary_data/Tourism/non_moving_window/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': tfrecord_writer = TFRecordWriter( output_size = 24, train_file_path='../../../datasets/text_data/Tourism/non_moving_window/tourism_stl_24.txt', validate_file_path='../../../datasets/text_data/Tourism/non_moving_window/tourism_stl_24v.txt', test_file_path='../../../datasets/text_data/Tourism/non_moving_window/tourism_test_24.txt', binary_train_file_path=output_path + 'tourism_stl_24.tfrecords', binary_validation_file_path=output_path + 'tourism_stl_24v.tfrecords', binary_test_file_path=output_path + 'tourism_test_24.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
1,063
47.363636
103
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/Tourism/non_moving_window/without_stl_decomposition/create_tfrecords.py
from tfrecords_handler.non_moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../../datasets/binary_data/Tourism/non_moving_window/without_stl_decomposition/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': tfrecord_writer = TFRecordWriter( output_size = 24, train_file_path='../../../../datasets/text_data/Tourism/non_moving_window/without_stl_decomposition/tourism_24.txt', validate_file_path='../../../../datasets/text_data/Tourism/non_moving_window/without_stl_decomposition/tourism_24v.txt', test_file_path='../../../../datasets/text_data/Tourism/non_moving_window/without_stl_decomposition/tourism_test_24.txt', binary_train_file_path=output_path + 'tourism_24.tfrecords', binary_validation_file_path=output_path + 'tourism_24v.tfrecords', binary_test_file_path=output_path + 'tourism_test_24.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
1,163
51.909091
128
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/M3/train_test_data_splitter.py
import csv data_file = "../../datasets/text_data/M3/M3C.csv" train_data_file = "../../datasets/text_data/M3/Train_Dataset.csv" results_file = "../../datasets/text_data/M3/Test_Dataset.csv" with open(data_file, "r") as original_file, open(train_data_file, "w") as train_out, open(results_file, "w") as results_out: data_reader = csv.reader(original_file, delimiter=",") train_data_writer = csv.writer(train_out, delimiter=",") results_writer = csv.writer(results_out, delimiter=";") next(data_reader, None) for row in data_reader: number_of_columns = row[3] train_data = row[7 : 7 + int(number_of_columns)] train_data.insert(0, row[4]) train_data.insert(0, row[0]) results_data = row[7 + int(number_of_columns): len(row)] results_data.insert(0, row[4]) results_data.insert(0, row[0]) print(results_data) train_data_writer.writerow(train_data) results_writer.writerow(results_data)
986
35.555556
124
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/M3/moving_window/create_tfrecords.py
from tfrecords_handler.moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../datasets/binary_data/M3/moving_window/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': # macro data tfrecord_writer = TFRecordWriter( input_size = 12, output_size = 18, train_file_path = '../../../datasets/text_data/M3/moving_window/m3_stl_monthly_macro_18i12.txt', validate_file_path = '../../../datasets/text_data/M3/moving_window/m3_stl_monthly_macro_18i12v.txt', test_file_path = '../../../datasets/text_data/M3/moving_window/m3_test_monthly_macro_18i12.txt', binary_train_file_path = output_path + 'm3_stl_monthly_macro_18i12.tfrecords', binary_validation_file_path = output_path + 'm3_stl_monthly_macro_18i12v.tfrecords', binary_test_file_path = output_path + 'm3_test_monthly_macro_18i12.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # micro data tfrecord_writer = TFRecordWriter( input_size=13, output_size=18, train_file_path='../../../datasets/text_data/M3/moving_window/m3_stl_monthly_micro_18i13.txt', validate_file_path='../../../datasets/text_data/M3/moving_window/m3_stl_monthly_micro_18i13v.txt', test_file_path='../../../datasets/text_data/M3/moving_window/m3_test_monthly_micro_18i13.txt', binary_train_file_path=output_path + 'm3_stl_monthly_micro_18i13.tfrecords', binary_validation_file_path=output_path + 'm3_stl_monthly_micro_18i13v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_micro_18i13.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # industry data tfrecord_writer = TFRecordWriter( input_size=13, output_size=18, train_file_path='../../../datasets/text_data/M3/moving_window/m3_stl_monthly_industry_18i13.txt', validate_file_path='../../../datasets/text_data/M3/moving_window/m3_stl_monthly_industry_18i13v.txt', test_file_path='../../../datasets/text_data/M3/moving_window/m3_test_monthly_industry_18i13.txt', binary_train_file_path=output_path + 'm3_stl_monthly_industry_18i13.tfrecords', binary_validation_file_path=output_path + 'm3_stl_monthly_industry_18i13v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_industry_18i13.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # finance data tfrecord_writer = TFRecordWriter( input_size=13, output_size=18, train_file_path='../../../datasets/text_data/M3/moving_window/m3_stl_monthly_finance_18i13.txt', validate_file_path='../../../datasets/text_data/M3/moving_window/m3_stl_monthly_finance_18i13v.txt', test_file_path='../../../datasets/text_data/M3/moving_window/m3_test_monthly_finance_18i13.txt', binary_train_file_path=output_path + 'm3_stl_monthly_finance_18i13.tfrecords', binary_validation_file_path=output_path + 'm3_stl_monthly_finance_18i13v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_finance_18i13.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # other data tfrecord_writer = TFRecordWriter( input_size=13, output_size=18, train_file_path='../../../datasets/text_data/M3/moving_window/m3_stl_monthly_other_18i13.txt', validate_file_path='../../../datasets/text_data/M3/moving_window/m3_stl_monthly_other_18i13v.txt', test_file_path='../../../datasets/text_data/M3/moving_window/m3_test_monthly_other_18i13.txt', binary_train_file_path=output_path + 'm3_stl_monthly_other_18i13.tfrecords', binary_validation_file_path=output_path + 'm3_stl_monthly_other_18i13v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_other_18i13.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # demographic data tfrecord_writer = TFRecordWriter( input_size=13, output_size=18, train_file_path='../../../datasets/text_data/M3/moving_window/m3_stl_monthly_demo_18i13.txt', validate_file_path='../../../datasets/text_data/M3/moving_window/m3_stl_monthly_demo_18i13v.txt', test_file_path='../../../datasets/text_data/M3/moving_window/m3_test_monthly_demo_18i13.txt', binary_train_file_path=output_path + 'm3_stl_monthly_demo_18i13.tfrecords', binary_validation_file_path=output_path + 'm3_stl_monthly_demo_18i13v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_demo_18i13.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
5,644
50.788991
109
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/M3/moving_window/without_stl_decomposition/create_tfrecords.py
from tfrecords_handler.moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../../datasets/binary_data/M3/moving_window/without_stl_decomposition/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': # macro data tfrecord_writer = TFRecordWriter( input_size = 12, output_size = 18, train_file_path = '../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_monthly_macro_18i12.txt', validate_file_path = '../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_monthly_macro_18i12v.txt', test_file_path = '../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_test_monthly_macro_18i12.txt', binary_train_file_path = output_path + 'm3_monthly_macro_18i12.tfrecords', binary_validation_file_path = output_path + 'm3_monthly_macro_18i12v.tfrecords', binary_test_file_path = output_path + 'm3_test_monthly_macro_18i12.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # micro data tfrecord_writer = TFRecordWriter( input_size=13, output_size=18, train_file_path='../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_monthly_micro_18i13.txt', validate_file_path='../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_monthly_micro_18i13v.txt', test_file_path='../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_test_monthly_micro_18i13.txt', binary_train_file_path=output_path + 'm3_monthly_micro_18i13.tfrecords', binary_validation_file_path=output_path + 'm3_monthly_micro_18i13v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_micro_18i13.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # industry data tfrecord_writer = TFRecordWriter( input_size=13, output_size=18, train_file_path='../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_monthly_industry_18i13.txt', validate_file_path='../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_monthly_industry_18i13v.txt', test_file_path='../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_test_monthly_industry_18i13.txt', binary_train_file_path=output_path + 'm3_monthly_industry_18i13.tfrecords', binary_validation_file_path=output_path + 'm3_monthly_industry_18i13v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_industry_18i13.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # finance data tfrecord_writer = TFRecordWriter( input_size=13, output_size=18, train_file_path='../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_monthly_finance_18i13.txt', validate_file_path='../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_monthly_finance_18i13v.txt', test_file_path='../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_test_monthly_finance_18i13.txt', binary_train_file_path=output_path + 'm3_monthly_finance_18i13.tfrecords', binary_validation_file_path=output_path + 'm3_monthly_finance_18i13v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_finance_18i13.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # other data tfrecord_writer = TFRecordWriter( input_size=13, output_size=18, train_file_path='../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_monthly_other_18i13.txt', validate_file_path='../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_monthly_other_18i13v.txt', test_file_path='../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_test_monthly_other_18i13.txt', binary_train_file_path=output_path + 'm3_monthly_other_18i13.tfrecords', binary_validation_file_path=output_path + 'm3_monthly_other_18i13v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_other_18i13.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # demographic data tfrecord_writer = TFRecordWriter( input_size=13, output_size=18, train_file_path='../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_monthly_demo_18i13.txt', validate_file_path='../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_monthly_demo_18i13v.txt', test_file_path='../../../../datasets/text_data/M3/moving_window/without_stl_decomposition/m3_test_monthly_demo_18i13.txt', binary_train_file_path=output_path + 'm3_monthly_demo_18i13.tfrecords', binary_validation_file_path=output_path + 'm3_monthly_demo_18i13v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_demo_18i13.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
6,099
54.963303
134
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/M3/non_moving_window/create_tfrecords.py
from tfrecords_handler.non_moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../datasets/binary_data/M3/non_moving_window/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': # macro data tfrecord_writer = TFRecordWriter( output_size = 18, train_file_path = '../../../datasets/text_data/M3/non_moving_window/m3_stl_monthly_macro_18.txt', validate_file_path = '../../../datasets/text_data/M3/non_moving_window/m3_stl_monthly_macro_18v.txt', test_file_path = '../../../datasets/text_data/M3/non_moving_window/m3_test_monthly_macro_18.txt', binary_train_file_path = output_path + 'm3_stl_monthly_macro_18.tfrecords', binary_validation_file_path = output_path + 'm3_stl_monthly_macro_18v.tfrecords', binary_test_file_path = output_path + 'm3_test_monthly_macro_18.tfrecords', ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # micro data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../datasets/text_data/M3/non_moving_window/m3_stl_monthly_micro_18.txt', validate_file_path='../../../datasets/text_data/M3/non_moving_window/m3_stl_monthly_micro_18v.txt', test_file_path='../../../datasets/text_data/M3/non_moving_window/m3_test_monthly_micro_18.txt', binary_train_file_path=output_path + 'm3_stl_monthly_micro_18.tfrecords', binary_validation_file_path=output_path + 'm3_stl_monthly_micro_18v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_micro_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # industry data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../datasets/text_data/M3/non_moving_window/m3_stl_monthly_industry_18.txt', validate_file_path='../../../datasets/text_data/M3/non_moving_window/m3_stl_monthly_industry_18v.txt', test_file_path='../../../datasets/text_data/M3/non_moving_window/m3_test_monthly_industry_18.txt', binary_train_file_path=output_path + 'm3_stl_monthly_industry_18.tfrecords', binary_validation_file_path=output_path + 'm3_stl_monthly_industry_18v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_industry_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # finance data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../datasets/text_data/M3/non_moving_window/m3_stl_monthly_finance_18.txt', validate_file_path='../../../datasets/text_data/M3/non_moving_window/m3_stl_monthly_finance_18v.txt', test_file_path='../../../datasets/text_data/M3/non_moving_window/m3_test_monthly_finance_18.txt', binary_train_file_path=output_path + 'm3_stl_monthly_finance_18.tfrecords', binary_validation_file_path=output_path + 'm3_stl_monthly_finance_18v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_finance_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # other data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../datasets/text_data/M3/non_moving_window/m3_stl_monthly_other_18.txt', validate_file_path='../../../datasets/text_data/M3/non_moving_window/m3_stl_monthly_other_18v.txt', test_file_path='../../../datasets/text_data/M3/non_moving_window/m3_test_monthly_other_18.txt', binary_train_file_path=output_path + 'm3_stl_monthly_other_18.tfrecords', binary_validation_file_path=output_path + 'm3_stl_monthly_other_18v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_other_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # demographic data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../datasets/text_data/M3/non_moving_window/m3_stl_monthly_demo_18.txt', validate_file_path='../../../datasets/text_data/M3/non_moving_window/m3_stl_monthly_demo_18v.txt', test_file_path='../../../datasets/text_data/M3/non_moving_window/m3_test_monthly_demo_18.txt', binary_train_file_path=output_path + 'm3_stl_monthly_demo_18.tfrecords', binary_validation_file_path=output_path + 'm3_stl_monthly_demo_18v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_demo_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
5,477
52.184466
110
py
time-series-forecasting-release
time-series-forecasting-release/preprocess_scripts/M3/non_moving_window/without_stl_decomposition/create_tfrecords.py
from tfrecords_handler.non_moving_window.tfrecord_writer import TFRecordWriter import os output_path = "../../../../datasets/binary_data/M3/non_moving_window/without_stl_decomposition/" if not os.path.exists(output_path): os.makedirs(output_path) if __name__ == '__main__': # macro data tfrecord_writer = TFRecordWriter( output_size = 18, train_file_path = '../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_monthly_macro_18.txt', validate_file_path = '../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_monthly_macro_18v.txt', test_file_path = '../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_test_monthly_macro_18.txt', binary_train_file_path = output_path + 'm3_monthly_macro_18.tfrecords', binary_validation_file_path = output_path + 'm3_monthly_macro_18v.tfrecords', binary_test_file_path = output_path + 'm3_test_monthly_macro_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # micro data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_monthly_micro_18.txt', validate_file_path='../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_monthly_micro_18v.txt', test_file_path='../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_test_monthly_micro_18.txt', binary_train_file_path=output_path + 'm3_monthly_micro_18.tfrecords', binary_validation_file_path=output_path + 'm3_monthly_micro_18v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_micro_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # industry data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_monthly_industry_18.txt', validate_file_path='../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_monthly_industry_18v.txt', test_file_path='../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_test_monthly_industry_18.txt', binary_train_file_path=output_path + 'm3_monthly_industry_18.tfrecords', binary_validation_file_path=output_path + 'm3_monthly_industry_18v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_industry_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # finance data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_monthly_finance_18.txt', validate_file_path='../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_monthly_finance_18v.txt', test_file_path='../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_test_monthly_finance_18.txt', binary_train_file_path=output_path + 'm3_monthly_finance_18.tfrecords', binary_validation_file_path=output_path + 'm3_monthly_finance_18v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_finance_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # other data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_monthly_other_18.txt', validate_file_path='../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_monthly_other_18v.txt', test_file_path='../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_test_monthly_other_18.txt', binary_train_file_path=output_path + 'm3_monthly_other_18.tfrecords', binary_validation_file_path=output_path + 'm3_monthly_other_18v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_other_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file() # demographic data tfrecord_writer = TFRecordWriter( output_size=18, train_file_path='../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_monthly_demo_18.txt', validate_file_path='../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_monthly_demo_18v.txt', test_file_path='../../../../datasets/text_data/M3/non_moving_window/without_stl_decomposition/m3_test_monthly_demo_18.txt', binary_train_file_path=output_path + 'm3_monthly_demo_18.tfrecords', binary_validation_file_path=output_path + 'm3_monthly_demo_18v.tfrecords', binary_test_file_path=output_path + 'm3_test_monthly_demo_18.tfrecords' ) tfrecord_writer.read_text_data() tfrecord_writer.write_train_data_to_tfrecord_file() tfrecord_writer.write_validation_data_to_tfrecord_file() tfrecord_writer.write_test_data_to_tfrecord_file()
5,931
56.592233
135
py
time-series-forecasting-release
time-series-forecasting-release/graph_plotter/training_curve_plotter.py
import tensorflow as tf import numpy as np class CurvePlotter: def __init__(self, session, no_of_curves): self.__session = session self.__writer_train = tf.summary.FileWriter('./logs/plot_train') if no_of_curves == 2: self.__writer_val = tf.summary.FileWriter('./logs/plot_val') self.__loss_var = tf.Variable(0.0) tf.summary.scalar("loss", self.__loss_var) self.__write_op = tf.summary.merge_all() def plot_train(self, loss, epoch): summary = self.__session.run(self.__write_op, {self.__loss_var: np.mean(loss)}) self.__writer_train.add_summary(summary, epoch) self.__writer_train.flush() def plot_val(self, loss, epoch): summary = self.__session.run(self.__write_op, {self.__loss_var: np.mean(np.mean(loss))}) self.__writer_val.add_summary(summary, epoch) self.__writer_val.flush()
906
38.434783
96
py
time-series-forecasting-release
time-series-forecasting-release/configs/global_configs.py
# configs for the model training class model_training_configs: VALIDATION_ERRORS_DIRECTORY = 'results/validation_errors/' INFO_FREQ = 1 # configs for the model testing class model_testing_configs: RNN_FORECASTS_DIRECTORY = 'results/rnn_forecasts/' RNN_ERRORS_DIRECTORY = 'results/errors' PROCESSED_RNN_FORECASTS_DIRECTORY = '/results/processed_rnn_forecasts/' # configs for hyperparameter tuning(SMAC3) class hyperparameter_tuning_configs: SMAC_RUNCOUNT_LIMIT = 50 class gpu_configs: log_device_placement = False
544
29.277778
75
py
time-series-forecasting-release
time-series-forecasting-release/tfrecords_handler/moving_window/tfrecord_writer.py
import tensorflow as tf import numpy as np import pandas as pd class TFRecordWriter: def __init__(self, **kwargs): self.__input_size = kwargs['input_size'] self.__output_size = kwargs['output_size'] self.__train_file_path = kwargs['train_file_path'] self.__validate_file_path = kwargs['validate_file_path'] self.__test_file_path = kwargs['test_file_path'] self.__binary_train_file_path = kwargs['binary_train_file_path'] self.__binary_validation_file_path = kwargs['binary_validation_file_path'] self.__binary_test_file_path = kwargs['binary_test_file_path'] # read the text data from text files def read_text_data(self): self.__list_of_training_inputs = [] self.__list_of_training_outputs = [] self.__list_of_validation_inputs = [] self.__list_of_validation_outputs =[] self.__list_of_validation_metadata = [] self.__list_of_test_inputs = [] self.__list_of_test_metadata = [] # Reading the training dataset. train_df = pd.read_csv(self.__train_file_path, nrows=10) float_cols = [c for c in train_df if train_df[c].dtype == "float64"] float32_cols = {c: np.float32 for c in float_cols} train_df = pd.read_csv(self.__train_file_path, sep=" ", header=None, engine='c', dtype=float32_cols) train_df = train_df.rename(columns={0: 'series'}) # Returns unique number of time series in the dataset. series = pd.unique(train_df['series']) # Construct input and output training tuples for each time series. for ser in series: one_series_df = train_df[train_df['series'] == ser] inputs_df = one_series_df.iloc[:, range(1, (self.__input_size + 1))] outputs_df = one_series_df.iloc[:, range((self.__input_size + 2), (self.__input_size + self.__output_size + 2))] self.__list_of_training_inputs.append(np.ascontiguousarray(inputs_df, dtype=np.float32)) self.__list_of_training_outputs.append(np.ascontiguousarray(outputs_df, dtype=np.float32)) # Reading the validation dataset. val_df = pd.read_csv(self.__validate_file_path, nrows=10) float_cols = [c for c in val_df if val_df[c].dtype == "float64"] float32_cols = {c: np.float32 for c in float_cols} val_df = pd.read_csv(self.__validate_file_path, sep=" ", header=None, engine='c', dtype=float32_cols) val_df = val_df.rename(columns={0: 'series'}) series = pd.unique(val_df['series']) for ser in series: one_series_df = val_df[val_df['series'] == ser] inputs_df_test = one_series_df.iloc[:, range(1, (self.__input_size + 1))] metadata_df = one_series_df.iloc[:, range((self.__input_size + self.__output_size + 3), one_series_df.shape[1])] outputs_df_test = one_series_df.iloc[:, range((self.__input_size + 2), (self.__input_size + self.__output_size + 2))] self.__list_of_validation_inputs.append(np.ascontiguousarray(inputs_df_test, dtype=np.float32)) self.__list_of_validation_outputs.append(np.ascontiguousarray(outputs_df_test, dtype=np.float32)) self.__list_of_validation_metadata.append(np.ascontiguousarray(metadata_df, dtype=np.float32)) # Reading the test file. test_df = pd.read_csv(self.__test_file_path, nrows=10) float_cols = [c for c in test_df if test_df[c].dtype == "float64"] float32_cols = {c: np.float32 for c in float_cols} test_df = pd.read_csv(self.__test_file_path, sep=" ", header=None, engine='c', dtype=float32_cols) test_df = test_df.rename(columns={0: 'series'}) series = pd.unique(test_df['series']) for ser in series: test_series_df = test_df[test_df['series'] == ser] test_inputs_df = test_series_df.iloc[:, range(1, (self.__input_size + 1))] metadata_df = test_series_df.iloc[:, range((self.__input_size + 2), test_series_df.shape[1])] self.__list_of_test_inputs.append(np.ascontiguousarray(test_inputs_df, dtype=np.float32)) self.__list_of_test_metadata.append(np.ascontiguousarray(metadata_df, dtype=np.float32)) # write the train and validation text data into tfrecord file def write_train_data_to_tfrecord_file(self): writer = tf.python_io.TFRecordWriter(self.__binary_train_file_path, tf.python_io.TFRecordOptions(tf.python_io.TFRecordCompressionType.ZLIB)) # write the training data file in tfrecords format for input, output in zip(self.__list_of_training_inputs, self.__list_of_training_outputs): sequence_length = input.shape[0] sequence_example = tf.train.SequenceExample( context=tf.train.Features(feature={ "sequence_length" : tf.train.Feature(int64_list=tf.train.Int64List(value=[sequence_length])) }), feature_lists = tf.train.FeatureLists(feature_list={ "input" : tf.train.FeatureList(feature=[ tf.train.Feature(float_list=tf.train.FloatList(value=input_sequence)) for input_sequence in input ]), "output" : tf.train.FeatureList(feature=[ tf.train.Feature(float_list=tf.train.FloatList(value=output_sequence)) for output_sequence in output ]) }) ) writer.write(sequence_example.SerializeToString()) writer.close() # write the train and validation text data into tfrecord file def write_validation_data_to_tfrecord_file(self): writer = tf.python_io.TFRecordWriter(self.__binary_validation_file_path, tf.python_io.TFRecordOptions( tf.python_io.TFRecordCompressionType.ZLIB)) # write the training data file in tfrecords format for input, output, metadata in zip(self.__list_of_validation_inputs, self.__list_of_validation_outputs, self.__list_of_validation_metadata): sequence_length = input.shape[0] sequence_example = tf.train.SequenceExample( context=tf.train.Features(feature={ "sequence_length": tf.train.Feature(int64_list=tf.train.Int64List(value=[sequence_length])) }), feature_lists=tf.train.FeatureLists(feature_list={ "input": tf.train.FeatureList(feature=[ tf.train.Feature(float_list=tf.train.FloatList(value=input_sequence)) for input_sequence in input ]), "output": tf.train.FeatureList(feature=[ tf.train.Feature(float_list=tf.train.FloatList(value=output_sequence)) for output_sequence in output ]), "metadata": tf.train.FeatureList(feature=[ tf.train.Feature(float_list=tf.train.FloatList(value=metadata_sequence)) for metadata_sequence in metadata ]) }) ) writer.write(sequence_example.SerializeToString()) writer.close() # write the test text data into tfrecord file def write_test_data_to_tfrecord_file(self): writer = tf.python_io.TFRecordWriter(self.__binary_test_file_path, tf.python_io.TFRecordOptions(tf.python_io.TFRecordCompressionType.ZLIB)) # write the training data file in tfrecords format for input, metadata in zip(self.__list_of_test_inputs, self.__list_of_test_metadata): sequence_length = input.shape[0] sequence_example = tf.train.SequenceExample( context=tf.train.Features(feature={ "sequence_length" : tf.train.Feature(int64_list=tf.train.Int64List(value=[sequence_length])) }), feature_lists = tf.train.FeatureLists(feature_list={ "input" : tf.train.FeatureList(feature=[ tf.train.Feature(float_list=tf.train.FloatList(value=input_sequence)) for input_sequence in input ]), "metadata" : tf.train.FeatureList(feature=[ tf.train.Feature(float_list=tf.train.FloatList(value=metadata_sequence)) for metadata_sequence in metadata ]) }) ) writer.write(sequence_example.SerializeToString()) writer.close()
8,497
50.817073
148
py
time-series-forecasting-release
time-series-forecasting-release/tfrecords_handler/moving_window/tfrecord_reader.py
import tensorflow as tf class TFRecordReader: def __init__(self, input_size, output_size, metadata_size): self.__input_size = input_size self.__output_size = output_size self.__metadata_size = metadata_size def train_data_parser(self, serialized_example): context_parsed, sequence_parsed = tf.parse_single_sequence_example( serialized_example, context_features=({ "sequence_length": tf.FixedLenFeature([], dtype=tf.int64) }), sequence_features=({ "input": tf.FixedLenSequenceFeature([self.__input_size], dtype=tf.float32), "output": tf.FixedLenSequenceFeature([self.__output_size], dtype=tf.float32) }) ) return context_parsed["sequence_length"], sequence_parsed["input"], sequence_parsed["output"] def validation_data_parser(self, serialized_example): context_parsed, sequence_parsed = tf.parse_single_sequence_example( serialized_example, context_features=({ "sequence_length": tf.FixedLenFeature([], dtype=tf.int64) }), sequence_features=({ "input": tf.FixedLenSequenceFeature([self.__input_size], dtype=tf.float32), "output": tf.FixedLenSequenceFeature([self.__output_size], dtype=tf.float32), "metadata": tf.FixedLenSequenceFeature([self.__metadata_size], dtype=tf.float32) }) ) return context_parsed["sequence_length"], sequence_parsed["input"], sequence_parsed["output"], sequence_parsed[ "metadata"] def test_data_parser(self, serialized_example): context_parsed, sequence_parsed = tf.parse_single_sequence_example( serialized_example, context_features=({ "sequence_length": tf.FixedLenFeature([], dtype=tf.int64) }), sequence_features=({ "input": tf.FixedLenSequenceFeature([self.__input_size], dtype=tf.float32), "metadata": tf.FixedLenSequenceFeature([self.__metadata_size], dtype=tf.float32) }) ) return context_parsed["sequence_length"], sequence_parsed["input"], sequence_parsed["metadata"]
2,277
41.981132
119
py
time-series-forecasting-release
time-series-forecasting-release/tfrecords_handler/non_moving_window/tfrecord_writer.py
import tensorflow as tf import numpy as np import pandas as pd import csv class TFRecordWriter: def __init__(self, **kwargs): self.__output_size = kwargs['output_size'] self.__train_file_path = kwargs['train_file_path'] self.__validate_file_path = kwargs['validate_file_path'] self.__test_file_path = kwargs['test_file_path'] self.__binary_train_file_path = kwargs['binary_train_file_path'] self.__binary_validation_file_path = kwargs['binary_validation_file_path'] self.__binary_test_file_path = kwargs['binary_test_file_path'] # read the text data from text files def read_text_data(self): self.__list_of_training_inputs = [] self.__list_of_training_outputs = [] self.__list_of_validation_inputs = [] self.__list_of_validation_outputs = [] self.__list_of_validation_metadata = [] self.__list_of_test_inputs = [] self.__list_of_test_metadata = [] # Reading the training dataset. with open(self.__train_file_path) as train_file: train_data_reader = csv.reader(train_file, delimiter=" ") train_data_list = list(train_data_reader) for series in train_data_list: last_train_input_index = len(series) - self.__output_size - 2 train_input_data = series[1: last_train_input_index + 1] train_output_data = series[(last_train_input_index + 2): len(series)] self.__list_of_training_inputs.append(np.ascontiguousarray(train_input_data, dtype=np.float32)) self.__list_of_training_outputs.append(np.ascontiguousarray(train_output_data, dtype=np.float32)) # Reading the validation dataset with open(self.__validate_file_path) as validate_file: validate_file_reader = csv.reader(validate_file, delimiter=" ") validate_data_list = list(validate_file_reader) for series in validate_data_list: meta_data_index = series.index("|#") output_index = series.index("|o") validate_input_data = series[1: output_index] validate_output_data = series[output_index + 1: meta_data_index] validate_meta_data = series[meta_data_index + 1:] self.__list_of_validation_inputs.append(np.ascontiguousarray(validate_input_data, dtype=np.float32)) self.__list_of_validation_outputs.append(np.ascontiguousarray(validate_output_data, dtype=np.float32)) self.__list_of_validation_metadata.append(np.ascontiguousarray(validate_meta_data, dtype=np.float32)) # Reading the test file with open(self.__test_file_path) as test_file: test_file_reader = csv.reader(test_file, delimiter=" ") test_data_list = list(test_file_reader) for series in test_data_list: meta_data_index = series.index("|#") test_input_data = series[1: meta_data_index] test_meta_data = series[meta_data_index + 1:] self.__list_of_test_inputs.append(np.ascontiguousarray(test_input_data, dtype=np.float32)) self.__list_of_test_metadata.append(np.ascontiguousarray(test_meta_data, dtype=np.float32)) # write the train and validation text data into tfrecord file def write_train_data_to_tfrecord_file(self): writer = tf.python_io.TFRecordWriter(self.__binary_train_file_path, tf.python_io.TFRecordOptions(tf.python_io.TFRecordCompressionType.ZLIB)) # write the training data file in tfrecords format for input, output in zip(self.__list_of_training_inputs, self.__list_of_training_outputs): sequence_length = len(input) sequence_example = tf.train.SequenceExample( context=tf.train.Features(feature={ "sequence_length": tf.train.Feature(int64_list=tf.train.Int64List(value=[sequence_length])) }), feature_lists=tf.train.FeatureLists(feature_list={ "input": tf.train.FeatureList(feature=[ tf.train.Feature(float_list=tf.train.FloatList(value=[input_data_element])) for input_data_element in input ]), "output": tf.train.FeatureList(feature=[ tf.train.Feature(float_list=tf.train.FloatList(value=[output_data_element])) for output_data_element in output ]) }) ) writer.write(sequence_example.SerializeToString()) writer.close() # write the train and validation text data into tfrecord file def write_validation_data_to_tfrecord_file(self): writer = tf.python_io.TFRecordWriter(self.__binary_validation_file_path, tf.python_io.TFRecordOptions( tf.python_io.TFRecordCompressionType.ZLIB)) # write the training data file in tfrecords format for input, output, metadata in zip(self.__list_of_validation_inputs, self.__list_of_validation_outputs, self.__list_of_validation_metadata): sequence_length = input.shape[0] sequence_example = tf.train.SequenceExample( context=tf.train.Features(feature={ "sequence_length": tf.train.Feature(int64_list=tf.train.Int64List(value=[sequence_length])) }), feature_lists=tf.train.FeatureLists(feature_list={ "input": tf.train.FeatureList(feature=[ tf.train.Feature(float_list=tf.train.FloatList(value=[input_data_element])) for input_data_element in input ]), "output": tf.train.FeatureList(feature=[ tf.train.Feature(float_list=tf.train.FloatList(value=[output_data_element])) for output_data_element in output ]), "metadata": tf.train.FeatureList(feature=[ tf.train.Feature(float_list=tf.train.FloatList(value=[metadata_element])) for metadata_element in metadata ]) }) ) writer.write(sequence_example.SerializeToString()) writer.close() # write the test text data into tfrecord file def write_test_data_to_tfrecord_file(self): writer = tf.python_io.TFRecordWriter(self.__binary_test_file_path, tf.python_io.TFRecordOptions(tf.python_io.TFRecordCompressionType.ZLIB)) # write the training data file in tfrecords format for input, metadata in zip(self.__list_of_test_inputs, self.__list_of_test_metadata): sequence_length = input.shape[0] sequence_example = tf.train.SequenceExample( context=tf.train.Features(feature={ "sequence_length": tf.train.Feature(int64_list=tf.train.Int64List(value=[sequence_length])) }), feature_lists=tf.train.FeatureLists(feature_list={ "input": tf.train.FeatureList(feature=[ tf.train.Feature(float_list=tf.train.FloatList(value=[input_data_element])) for input_data_element in input ]), "metadata": tf.train.FeatureList(feature=[ tf.train.Feature(float_list=tf.train.FloatList(value=[metadata_element])) for metadata_element in metadata ]) }) ) writer.write(sequence_example.SerializeToString()) writer.close()
7,776
50.164474
124
py
time-series-forecasting-release
time-series-forecasting-release/tfrecords_handler/non_moving_window/tfrecord_reader.py
import tensorflow as tf class TFRecordReader: def train_data_parser(self, serialized_example): context_parsed, sequence_parsed = tf.parse_single_sequence_example( serialized_example, context_features=({ "sequence_length": tf.FixedLenFeature([], dtype=tf.int64) }), sequence_features=({ "input": tf.FixedLenSequenceFeature([1], dtype=tf.float32), "output": tf.FixedLenSequenceFeature([1], dtype=tf.float32) }) ) return context_parsed["sequence_length"], sequence_parsed["input"], sequence_parsed["output"] def validation_data_parser(self, serialized_example): context_parsed, sequence_parsed = tf.parse_single_sequence_example( serialized_example, context_features=({ "sequence_length": tf.FixedLenFeature([], dtype=tf.int64) }), sequence_features=({ "input": tf.FixedLenSequenceFeature([1], dtype=tf.float32), "output": tf.FixedLenSequenceFeature([1], dtype=tf.float32), "metadata": tf.FixedLenSequenceFeature([1], dtype=tf.float32) }) ) return context_parsed["sequence_length"], sequence_parsed["input"], sequence_parsed["output"], sequence_parsed[ "metadata"] def test_data_parser(self, serialized_example): context_parsed, sequence_parsed = tf.parse_single_sequence_example( serialized_example, context_features=({ "sequence_length": tf.FixedLenFeature([], dtype=tf.int64) }), sequence_features=({ "input": tf.FixedLenSequenceFeature([1], dtype=tf.float32), "metadata": tf.FixedLenSequenceFeature([1], dtype=tf.float32) }) ) return context_parsed["sequence_length"], sequence_parsed["input"], sequence_parsed["metadata"]
1,967
40
119
py
time-series-forecasting-release
time-series-forecasting-release/utility_scripts/persist_optimized_config_results.py
def persist_results(results, file): file_object = open(file, mode = 'w') for k, v in results.items(): file_object.write(str(k) + ' >>> ' + str(v) + '\n\n') file_object.close()
198
23.875
61
py
time-series-forecasting-release
time-series-forecasting-release/utility_scripts/time_series_length_calculator.py
import argparse import csv argument_parser = argparse.ArgumentParser("Calculate the time series length") argument_parser.add_argument('--data_file', required=True, help='The full name of the data file') argument_parser.add_argument('--output_file', required=True, help='The full name of the output file') args = argument_parser.parse_args() data_file = args.data_file output_file = args.output_file lengths_list =[] with open(data_file) as input: for line in input: length = line.count(",") + 1 lengths_list.append(length) with open(output_file, "w") as output: for item in lengths_list: output.write("%s\n" % item)
652
30.095238
101
py
time-series-forecasting-release
time-series-forecasting-release/utility_scripts/invoke_r_final_evaluation.py
import subprocess from configs.global_configs import model_testing_configs def invoke_r_script(args, moving_window): if moving_window: subprocess.call(["Rscript", "--vanilla", "error_calculator/moving_window/final_evaluation.R", args[0], model_testing_configs.RNN_ERRORS_DIRECTORY, model_testing_configs.PROCESSED_RNN_FORECASTS_DIRECTORY, args[1], args[2], args[3], args[4], args[5], args[6], args[7], args[8], args[9], args[10], args[11]]) else: subprocess.call(["Rscript", "--vanilla", "error_calculator/non_moving_window/final_evaluation.R", args[0], model_testing_configs.RNN_ERRORS_DIRECTORY, model_testing_configs.PROCESSED_RNN_FORECASTS_DIRECTORY, args[1], args[2], args[3], args[4], args[5], args[6], args[7], args[8], args[9], args[10]])
772
84.888889
313
py
time-series-forecasting-release
time-series-forecasting-release/utility_scripts/hyperparameter_scripts/hyperparameter_summary_generator.py
## This file concatenates the optimal hyperparameter configs across all the models for the same dataset and writes to a csv file import glob import argparse import pandas as pd import re from utility_scripts.hyperparameter_scripts.hyperparameter_config_reader import read_optimal_hyperparameter_values # get the different cluster names as external arguments argument_parser = argparse.ArgumentParser("Create hyperparameter summaries") argument_parser.add_argument('--dataset_name', required=True, help='Unique string for the name of the dataset') # parse the user arguments args = argument_parser.parse_args() dataset_name = args.dataset_name input_path = '../results/optimized_configurations/' output_path = '../results/optimized_configurations/aggregate_hyperparameter_configs/' output_file = output_path + dataset_name + ".csv" # get the list of all the files matching the regex hyperparameter_files = [filename for filename in glob.iglob(input_path + dataset_name + "_*")] hyperparameters_df = pd.DataFrame( columns=["Model_Name", "cell_dimension", "gaussian_noise_stdev", "l2_regularization", "max_epoch_size", "max_num_epochs", "minibatch_size", "num_hidden_layers", "random_normal_initializer_stdev", "rate_of_learning"]) # concat all the hyperparameters to data frames for config_file in sorted(hyperparameter_files): file_name_part = re.split(pattern=dataset_name + "_", string=config_file, maxsplit=1)[1] model_name = file_name_part.rsplit('_', 1)[0] print(model_name) hyperparameter_values_dic = read_optimal_hyperparameter_values(config_file) if "rate_of_learning" not in hyperparameter_values_dic.keys(): hyperparameter_values_dic["rate_of_learning"] = "-" hyperparameters_df.loc[-1] = [model_name, hyperparameter_values_dic["cell_dimension"], hyperparameter_values_dic["gaussian_noise_stdev"], hyperparameter_values_dic["l2_regularization"], hyperparameter_values_dic["max_epoch_size"], hyperparameter_values_dic["max_num_epochs"], hyperparameter_values_dic["minibatch_size"], hyperparameter_values_dic["num_hidden_layers"], hyperparameter_values_dic["random_normal_initializer_stdev"], hyperparameter_values_dic["rate_of_learning"]] hyperparameters_df.index = hyperparameters_df.index + 1 # write the errors to csv file hyperparameters_df.to_csv(output_file, index=False)
2,639
46.142857
128
py
time-series-forecasting-release
time-series-forecasting-release/utility_scripts/hyperparameter_scripts/hyperparameter_config_reader.py
import re def read_optimal_hyperparameter_values(file_name): # define dictionary to store the hyperparameter values hyperparameter_values_dic = {} with open(file_name) as configs_file: configs = configs_file.readlines() for config in configs: if not config.startswith('#') and config.strip(): values = [value.strip() for value in (re.split(">>>", config))] hyperparameter_values_dic[values[0]] = float(values[1]) configs_file.close() return hyperparameter_values_dic def read_initial_hyperparameter_values(initial_hyperparameter_values_file): # define dictionary to store the hyperparameter values hyperparameter_values_dic = {} with open(initial_hyperparameter_values_file) as configs_file: configs = configs_file.readlines() for config in configs: if not config.startswith('#') and config.strip(): values = [value.strip() for value in (re.split("-|,", config))] hyperparameter_values_dic[values[0]] = [float(values[1]), float(values[2])] configs_file.close() return hyperparameter_values_dic
1,169
36.741935
91
py
time-series-forecasting-release
time-series-forecasting-release/utility_scripts/error_summary_scripts/error_summary_generator.py
## This file concatenates the results across all the models for the same dataset and writes the mean SMAPE, median SMAPE, ranked SMAPE, # mean MASE, median MASE and ranked MASE across the time series to a csv file import glob import argparse import pandas as pd import re import numpy as np from collections import defaultdict # get the different cluster names as external arguments argument_parser = argparse.ArgumentParser("Create error summaries") argument_parser.add_argument('--dataset_name', required=True, help='Unique string for the name of the dataset') argument_parser.add_argument('--is_merged_cluster_result', required=True, help='1/0 denoting whether the results are merged from multiple clusters or not') # parse the user arguments args = argument_parser.parse_args() dataset_name = args.dataset_name is_merged_cluster_result = args.is_merged_cluster_result if int(is_merged_cluster_result): input_path = '../results/ensemble_errors/merged_cluster_results/' else: input_path = '../results/ensemble_errors/' output_path = '../results/ensemble_errors/aggregate_errors/' output_file = output_path + dataset_name + ".csv" # get the list of all the files matching the regex all_SMAPE_files = [filename for filename in glob.iglob(input_path + "all_smape_errors_" + dataset_name + "_*")] all_MASE_files = [filename for filename in glob.iglob(input_path + "all_mase_errors_" + dataset_name + "_*")] average_smape_errors_df = pd.DataFrame() average_mase_errors_df = pd.DataFrame() all_errors_df = pd.DataFrame(columns=["Model_Name", "Mean_SMAPE", "Median_SMAPE", "Mean_MASE", "Median_MASE"]) all_seeds_smape_errors_dic = {} all_seeds_mase_errors_dic = {} # concat all the errors to data frames for smape_errors_file, mase_errors_file in zip(sorted(all_SMAPE_files), sorted(all_MASE_files)): print(smape_errors_file) smape_errors_file_object = open(smape_errors_file, "r") mase_errors_file_object = open(mase_errors_file, "r") file_name_part = re.split(pattern="all_smape_errors_" + dataset_name + "_", string=smape_errors_file, maxsplit=1)[1] current_model_all_smape_errors = [] for num in smape_errors_file_object: if num == "NA\n" or num == "NA": current_model_all_smape_errors.append(np.nan) else: current_model_all_smape_errors.append(float(num)) current_model_all_mase_errors = [] for num in mase_errors_file_object: if num == "NA\n" or num == "NA": current_model_all_mase_errors.append(np.nan) else: current_model_all_mase_errors.append(float(num)) smape_errors = np.asarray(current_model_all_smape_errors) mase_errors = np.asarray(current_model_all_mase_errors) # store the errors to calculate ranked errors later average_smape_errors_df[file_name_part] = smape_errors average_mase_errors_df[file_name_part] = mase_errors # calculate the mean, median mean_SMAPE = np.mean(smape_errors) median_SMAPE = np.median(smape_errors) mean_MASE = np.mean(mase_errors) median_MASE = np.median(mase_errors) all_errors_df.loc[-1] = [file_name_part, mean_SMAPE, median_SMAPE, mean_MASE, median_MASE] all_errors_df.index = all_errors_df.index + 1 # calculate the ranked errors all_smape_ranks_df = np.mean(average_smape_errors_df.rank(axis=1), axis=0) all_mase_ranks_df = np.mean(average_mase_errors_df.rank(axis=1), axis=0) # add the ranked errors to all_errors_df all_errors_df["Ranked_SMAPE"] = all_smape_ranks_df.tolist() all_errors_df["Ranked_MASE"] = all_mase_ranks_df.tolist() # write the errors to csv file all_errors_df.to_csv(output_file, index=False)
3,726
38.648936
135
py
time-series-forecasting-release
time-series-forecasting-release/utility_scripts/error_summary_scripts/ensembling_forecasts.py
## This file concatenates the results across all the models for the same dataset, takes the median of errors for different seeds and writes the mean SMAPE, median SMAPE, ranked SMAPE, # mean MASE, median MASE and ranked MASE to a csv file import glob import argparse import pandas as pd import numpy as np from collections import defaultdict # get the different cluster names as external arguments argument_parser = argparse.ArgumentParser("Ensembling forecasts") argument_parser.add_argument('--dataset_name', required=True, help='Unique string for the name of the dataset') # parse the user arguments args = argument_parser.parse_args() dataset_name = args.dataset_name input_path = '../results/rnn_forecasts/' output_path = '../results/ensemble_rnn_forecasts/' all_forecast_files = [filename for filename in glob.iglob(input_path + dataset_name + "_*")] all_models_all_seeds_forecasts_dic = defaultdict(list) # concat all the errors to data frames for file_name in sorted(all_forecast_files): file_name_part = file_name.rsplit('_', 1)[0] file_name_part = file_name_part.split(input_path, 1)[1] # read the forecasts from the current file current_model_forecasts_df = pd.read_csv(file_name, header=None, dtype=np.float64) # append the forecasts to the dictionary all_models_all_seeds_forecasts_dic[file_name_part].append(current_model_forecasts_df) # iterate the dictionary for (model, forecasts_df_list) in all_models_all_seeds_forecasts_dic.items(): # convert the dataframe list to an array forecasts_array = np.stack(forecasts_df_list) # take the mean of forecasts across seeds for ensembling ensembled_forecasts = np.nanmedian(forecasts_array, axis=0) # write the ensembled forecasts to a file output_file = output_path + model np.savetxt(output_file, ensembled_forecasts, delimiter = ',')
1,864
34.865385
183
py
time-series-forecasting-release
time-series-forecasting-release/utility_scripts/error_summary_scripts/clusters_results_merger.py
## This file concatenates the error results of the different clusters for a given dataset ## The script requires that all the files(mean_median file, all_smape_errors and all_mase_errors files) are present for all the clusters subject to consideration import glob import numpy as np import argparse import re from collections import defaultdict import os input_path = '../results/ensemble_errors/' output_path = '../results/ensemble_errors/merged_cluster_results/' # get the different cluster names as external arguments argument_parser = argparse.ArgumentParser("Concatenate different cluster results of the same dataset") argument_parser.add_argument('--dataset_name', required=True, help='Unique string for the name of the dataset') # parse the user arguments args = argument_parser.parse_args() dataset_name = args.dataset_name output_file_mean_median = output_path + "mean_median_" + dataset_name + "_" output_file_all_smape_errors = output_path + "all_smape_errors_" + dataset_name + "_" output_file_all_mase_errors = output_path + "all_mase_errors_" + dataset_name + "_" # delete files if existing existing_mean_median_file_list = glob.glob(output_file_mean_median + "*", recursive = True) existing_all_smape_file_list = glob.glob(output_file_all_smape_errors + "*", recursive = True) existing_all_mase_file_list = glob.glob(output_file_all_mase_errors + "*", recursive = True) for file in existing_mean_median_file_list: try: os.remove(file) except: print("Error while deleting file : ", file) for file in existing_all_smape_file_list: try: os.remove(file) except: print("Error while deleting file : ", file) for file in existing_all_mase_file_list: try: os.remove(file) except: print("Error while deleting file : ", file) # get the list of all the files matching the regex # smape errors all_smape_errors_files = [filename for filename in glob.iglob(input_path + "all_smape_errors_" + dataset_name + "_*")] # mase errors all_mase_errors_files = [filename for filename in glob.iglob(input_path + "all_mase_errors_" + dataset_name + "_*")] all_mase_errors_dic = defaultdict(dict) all_smape_errors_dic = defaultdict(dict) category_order_dic = { "macro": 0, "micro": 1, "demo": 2, "industry": 3, "finance": 4, "other": 5 } # read the files one by one and merge the content for smape_errors_file, mase_errors_file in zip(sorted(all_smape_errors_files), sorted(all_mase_errors_files)): filename_smape_object = open(smape_errors_file) filename_mase_object = open(mase_errors_file) string_list = re.split(pattern="all_smape_errors_" + dataset_name + "_" + "[a-zA-Z0-9]+_", string=smape_errors_file, maxsplit=1) filename_part = string_list[1] category = re.search(dataset_name + '_([a-zA-Z0-9]+)_', smape_errors_file).group(1) # read the errors from both files all_smape_errors = [] for num in filename_smape_object: if num == "NA\n": all_smape_errors.append("NA") else: all_smape_errors.append(float(num)) all_mase_errors = [] for num in filename_mase_object: if num == "NA\n": all_mase_errors.append("NA") else: all_mase_errors.append(float(num)) all_mase_errors_dic[filename_part][category] = all_mase_errors all_smape_errors_dic[filename_part][category] = all_smape_errors # iterate the dictionaries and write to files for (smape_key, smape_errors_dic), (mase_key, mase_errors_dic) in zip(all_smape_errors_dic.items(), all_mase_errors_dic.items()): # open the all errors smape file output_file_all_smape_errors_object = open(output_file_all_smape_errors + smape_key, "a") # open the all errors mase file output_file_all_mase_errors_object = open(output_file_all_mase_errors + mase_key, "a") # sort the categories according to the given order for (category, index) in sorted(category_order_dic.items(), key=lambda item: item[1]): category_smape_errors = smape_errors_dic[category] category_mase_errors = mase_errors_dic[category] # write to files output_file_all_smape_errors_object.writelines('\n'.join(str(element) for element in category_smape_errors)) output_file_all_smape_errors_object.writelines('\n') output_file_all_mase_errors_object.writelines('\n'.join(str(element) for element in category_mase_errors)) output_file_all_mase_errors_object.writelines('\n') # close the files output_file_all_smape_errors_object.close() output_file_all_mase_errors_object.close()
4,761
38.032787
161
py
time-series-forecasting-release
time-series-forecasting-release/rnn_architectures/seq2seq_model/with_decoder/non_moving_window/unaccumulated_error/seq2seq_model_trainer.py
import numpy as np import tensorflow as tf from tensorflow.python.layers.core import Dense from tfrecords_handler.non_moving_window.tfrecord_reader import TFRecordReader from configs.global_configs import model_training_configs from configs.global_configs import training_data_configs from configs.global_configs import gpu_configs class Seq2SeqModelTrainer: def __init__(self, **kwargs): self.__use_bias = kwargs["use_bias"] self.__use_peepholes = kwargs["use_peepholes"] self.__output_size = kwargs["output_size"] self.__binary_train_file_path = kwargs["binary_train_file_path"] self.__binary_validation_file_path = kwargs["binary_validation_file_path"] self.__contain_zero_values = kwargs["contain_zero_values"] self.__address_near_zero_instability = kwargs["address_near_zero_instability"] self.__integer_conversion = kwargs["integer_conversion"] self.__seed = kwargs["seed"] self.__cell_type = kwargs["cell_type"] self.__without_stl_decomposition = kwargs['without_stl_decomposition'] def __l1_loss(self, z, t): loss = tf.reduce_mean(tf.abs(t - z)) return loss # Training the time series def train_model(self, **kwargs): num_hidden_layers = int(kwargs['num_hidden_layers']) cell_dimension = int(kwargs["cell_dimension"]) minibatch_size = int(kwargs["minibatch_size"]) max_epoch_size = int(kwargs["max_epoch_size"]) max_num_epochs = int(kwargs["max_num_epochs"]) l2_regularization = kwargs["l2_regularization"] gaussian_noise_stdev = kwargs["gaussian_noise_stdev"] random_normal_initializer_stdev = kwargs['random_normal_initializer_stdev'] optimizer_fn = kwargs["optimizer_fn"] tf.reset_default_graph() tf.set_random_seed(self.__seed) # adding noise to the input input = tf.placeholder(dtype=tf.float32, shape=[None, None, 1]) noise = tf.random_normal(shape=tf.shape(input), mean=0.0, stddev=gaussian_noise_stdev, dtype=tf.float32) training_input = input + noise validation_input = input training_target = tf.placeholder(dtype=tf.float32, shape=[None, self.__output_size, 1]) decoder_input = tf.placeholder(dtype=tf.float32, shape=[None, self.__output_size, 1]) # placeholder for the sequence lengths input_sequence_length = tf.placeholder(dtype=tf.int32, shape=[None]) output_sequence_length = tf.placeholder(dtype=tf.int32, shape=[None]) weight_initializer = tf.truncated_normal_initializer(stddev=random_normal_initializer_stdev) # create the model architecture # RNN with the layer of cells def cell(): if self.__cell_type == "LSTM": cell = tf.nn.rnn_cell.LSTMCell(num_units=int(cell_dimension), use_peepholes=self.__use_peepholes, initializer=weight_initializer) elif self.__cell_type == "GRU": cell = tf.nn.rnn_cell.GRUCell(num_units=int(cell_dimension), kernel_initializer=weight_initializer) elif self.__cell_type == "RNN": cell = tf.nn.rnn_cell.BasicRNNCell(num_units=int(cell_dimension)) return cell # building the encoder network multi_layered_encoder_cell = tf.nn.rnn_cell.MultiRNNCell( cells=[cell() for _ in range(int(num_hidden_layers))]) with tf.variable_scope('train_encoder_scope') as encoder_train_scope: training_encoder_outputs, training_encoder_state = tf.nn.dynamic_rnn(cell=multi_layered_encoder_cell, inputs=training_input, sequence_length=input_sequence_length, dtype=tf.float32) with tf.variable_scope(encoder_train_scope, reuse=tf.AUTO_REUSE) as encoder_inference_scope: inference_encoder_outputs, inference_encoder_states = tf.nn.dynamic_rnn(cell=multi_layered_encoder_cell, inputs=validation_input, sequence_length=input_sequence_length, dtype=tf.float32) # the final projection layer to convert the output to the desired dimension dense_layer = Dense(units=1, use_bias=self.__use_bias, kernel_initializer=weight_initializer) # decoder cell of the decoder network multi_layered_decoder_cell = tf.nn.rnn_cell.MultiRNNCell( cells=[cell() for _ in range(int(num_hidden_layers))]) # building the decoder network for training with tf.variable_scope('decoder_train_scope') as decoder_train_scope: # create the initial state for the decoder training_helper = tf.contrib.seq2seq.ScheduledOutputTrainingHelper(inputs=decoder_input, sequence_length=output_sequence_length, sampling_probability=0.0, name = "training_helper") training_decoder = tf.contrib.seq2seq.BasicDecoder(cell=multi_layered_decoder_cell, helper=training_helper, initial_state=training_encoder_state, output_layer=dense_layer) # perform the decoding training_decoder_outputs, _, _ = tf.contrib.seq2seq.dynamic_decode(decoder=training_decoder) # building the decoder network for inference with tf.variable_scope(decoder_train_scope, reuse=tf.AUTO_REUSE) as decoder_inference_scope: # create the initial state for the decoder inference_helper = tf.contrib.seq2seq.ScheduledOutputTrainingHelper(inputs=decoder_input, sequence_length=output_sequence_length, sampling_probability=1.0, name = "inference_helper") inference_decoder = tf.contrib.seq2seq.BasicDecoder(cell=multi_layered_decoder_cell, helper=inference_helper, initial_state=inference_encoder_states, output_layer=dense_layer) # perform the decoding inference_decoder_outputs, _, _ = tf.contrib.seq2seq.dynamic_decode(decoder=inference_decoder) # error that should be minimized in the training process error = self.__l1_loss(training_decoder_outputs[0], training_target) # l2 regularization of the trainable model parameters l2_loss = 0.0 for var in tf.trainable_variables(): l2_loss += tf.nn.l2_loss(var) l2_loss = tf.multiply(tf.cast(l2_regularization, dtype=tf.float64), tf.cast(l2_loss, dtype=tf.float64)) total_loss = tf.cast(error, dtype=tf.float64) + l2_loss # create the optimizer optimizer = optimizer_fn(total_loss) # create the training and validation datasets from the tfrecord files training_dataset = tf.data.TFRecordDataset(filenames = [self.__binary_train_file_path], compression_type = "ZLIB") validation_dataset = tf.data.TFRecordDataset(filenames = [self.__binary_validation_file_path], compression_type = "ZLIB") # parse the records tfrecord_reader = TFRecordReader() # define the expected shapes of data after padding train_padded_shapes = ([], [tf.Dimension(None), 1], [self.__output_size, 1]) validation_padded_shapes = ([], [tf.Dimension(None), 1], [self.__output_size, 1], [self.__meta_data_size, 1]) # preparing the training data shuffle_seed = tf.placeholder(dtype=tf.int64, shape=[]) # training_dataset = training_dataset.apply(tf.data.experimental.shuffle_and_repeat(buffer_size=training_data_configs.SHUFFLE_BUFFER_SIZE, # count=int(max_epoch_size), seed=shuffle_seed)) training_dataset = training_dataset.repeat(count=int(max_epoch_size)) training_dataset = training_dataset.map(tfrecord_reader.train_data_parser) padded_training_data_batches = training_dataset.padded_batch(batch_size=int(minibatch_size), padded_shapes=train_padded_shapes) training_data_batch_iterator = padded_training_data_batches.make_initializable_iterator() next_training_data_batch = training_data_batch_iterator.get_next() # preparing the validation data validation_dataset = validation_dataset.map(tfrecord_reader.validation_data_parser) # create a single batch from all the validation time series by padding the datasets to make the variable sequence lengths fixed padded_validation_dataset = validation_dataset.padded_batch(batch_size=int(minibatch_size), padded_shapes=validation_padded_shapes) # get an iterator to the validation data validation_data_iterator = padded_validation_dataset.make_initializable_iterator() # access the validation data using the iterator next_validation_data_batch = validation_data_iterator.get_next() # setup variable initialization init_op = tf.global_variables_initializer() # define the GPU options gpu_options = tf.GPUOptions(allow_growth=True) with tf.Session( config=tf.ConfigProto(log_device_placement=gpu_configs.log_device_placement, allow_soft_placement=True, gpu_options=gpu_options)) as session: session.run(init_op) smape_final = 0.0 smape_list = [] for epoch in range(max_num_epochs): print("Epoch->", epoch) session.run(training_data_batch_iterator.initializer, feed_dict={shuffle_seed:epoch}) while True: try: training_data_batch_value = session.run(next_training_data_batch, feed_dict={shuffle_seed:epoch}) decoder_input_value = np.hstack((np.expand_dims(training_data_batch_value[1][:, -1, :], axis=1), training_data_batch_value[2][:, :-1, :])) total_loss_value, _ = session.run([total_loss, optimizer], feed_dict={input: training_data_batch_value[1], training_target: training_data_batch_value[2], decoder_input: decoder_input_value, input_sequence_length: training_data_batch_value[0], output_sequence_length: [self.__output_size] * np.shape(training_data_batch_value[1])[0]}) except tf.errors.OutOfRangeError: break session.run(validation_data_iterator.initializer) while True: try: # get the batch of validation inputs validation_data_batch_value = session.run(next_validation_data_batch) # shape for the target data decoder_input_shape = [np.shape(validation_data_batch_value[1])[0], self.__output_size, 1] # get the output of the network for the validation input data batch validation_output = session.run(inference_decoder_outputs[0], feed_dict={input: validation_data_batch_value[1], decoder_input: np.zeros(decoder_input_shape), input_sequence_length: validation_data_batch_value[0], output_sequence_length: [self.__output_size] * np.shape(validation_data_batch_value[1])[0] }) # calculate the smape for the validation data using vectorization # convert the data to remove the preprocessing true_seasonality_values = validation_data_batch_value[3][:, 1:, 0] level_values = validation_data_batch_value[3][:, 0, 0] actual_values = validation_data_batch_value[2] if self.__without_stl_decomposition: converted_validation_output = np.exp(np.squeeze(validation_output, axis=2)) converted_actual_values = np.exp(np.squeeze(actual_values, axis=2)) else: converted_validation_output = np.exp( true_seasonality_values + level_values[:, np.newaxis] + np.squeeze(validation_output, axis=2)) converted_actual_values = np.exp( true_seasonality_values + level_values[:, np.newaxis] + np.squeeze(actual_values, axis=2)) if (self.__contain_zero_values): # to compensate for 0 values in data converted_validation_output = converted_validation_output - 1 converted_actual_values = converted_actual_values - 1 if self.__without_stl_decomposition: converted_validation_output = converted_validation_output * level_values[:, np.newaxis] converted_actual_values = converted_actual_values * level_values[:, np.newaxis] if self.__integer_conversion: converted_validation_output = np.round(converted_validation_output) converted_actual_values = np.round(converted_actual_values) converted_validation_output[converted_validation_output < 0] = 0 converted_actual_values[converted_actual_values < 0] = 0 if self.__address_near_zero_instability: # calculate the smape epsilon = 0.1 sum = np.maximum( np.abs(converted_validation_output) + np.abs(converted_actual_values) + epsilon, 0.5 + epsilon) smape_values = (np.abs(converted_validation_output - converted_actual_values) / sum) * 2 smape_values_per_series = np.mean(smape_values, axis=1) smape_list.extend(smape_values_per_series) else: # calculate the smape smape_values = (np.abs(converted_validation_output - converted_actual_values) / (np.abs(converted_validation_output) + np.abs(converted_actual_values))) * 2 smape_values_per_series = np.mean(smape_values, axis=1) smape_list.extend(smape_values_per_series) except tf.errors.OutOfRangeError: break smape_final = np.mean(smape_list) print("SMAPE value: {}".format(smape_final)) session.close() return float(smape_final), smape_list
16,182
54.611684
162
py
time-series-forecasting-release
time-series-forecasting-release/rnn_architectures/seq2seq_model/with_decoder/non_moving_window/unaccumulated_error/seq2seq_model_tester.py
import numpy as np import tensorflow as tf from tensorflow.python.layers.core import Dense from tfrecords_handler.non_moving_window.tfrecord_reader import TFRecordReader from configs.global_configs import training_data_configs from configs.global_configs import gpu_configs class Seq2SeqModelTester: def __init__(self, **kwargs): self.__use_bias = kwargs["use_bias"] self.__use_peepholes = kwargs["use_peepholes"] self.__output_size = kwargs["output_size"] self.__binary_train_file_path = kwargs["binary_train_file_path"] self.__binary_test_file_path = kwargs["binary_test_file_path"] self.__seed = kwargs["seed"] self.__cell_type = kwargs["cell_type"] def __l1_loss(self, z, t): loss = tf.reduce_mean(tf.abs(t - z)) return loss # Training the time series def test_model(self, **kwargs): # optimized hyperparameters num_hidden_layers = int(kwargs['num_hidden_layers']) max_num_epochs = int(kwargs['max_num_epochs']) max_epoch_size = int(kwargs['max_epoch_size']) cell_dimension = int(kwargs['cell_dimension']) l2_regularization = kwargs['l2_regularization'] minibatch_size = int(kwargs['minibatch_size']) gaussian_noise_stdev = kwargs['gaussian_noise_stdev'] random_normal_initializer_stdev = kwargs['random_normal_initializer_stdev'] optimizer_fn = kwargs['optimizer_fn'] # reset the tensorflow graph tf.reset_default_graph() tf.set_random_seed(self.__seed) # declare the input and output placeholders # adding noise to the input input = tf.placeholder(dtype=tf.float32, shape=[None, None, 1]) noise = tf.random_normal(shape=tf.shape(input), mean=0.0, stddev=gaussian_noise_stdev, dtype=tf.float32) training_input = input + noise testing_input = input training_target = tf.placeholder(dtype=tf.float32, shape=[None, self.__output_size, 1]) decoder_input = tf.placeholder(dtype=tf.float32, shape=[None, self.__output_size, 1]) # placeholder for the sequence lengths input_sequence_length = tf.placeholder(dtype=tf.int32, shape=[None]) output_sequence_length = tf.placeholder(dtype=tf.int32, shape=[None]) weight_initializer = tf.truncated_normal_initializer(stddev=random_normal_initializer_stdev) # create the model architecture # RNN with the layer of cells def cell(): if self.__cell_type == "LSTM": cell = tf.nn.rnn_cell.LSTMCell(num_units=int(cell_dimension), use_peepholes=self.__use_peepholes, initializer=weight_initializer) elif self.__cell_type == "GRU": cell = tf.nn.rnn_cell.GRUCell(num_units=int(cell_dimension), kernel_initializer=weight_initializer) elif self.__cell_type == "RNN": cell = tf.nn.rnn_cell.BasicRNNCell(num_units=int(cell_dimension)) return cell # building the encoder network multi_layered_encoder_cell = tf.nn.rnn_cell.MultiRNNCell( cells=[cell() for _ in range(int(num_hidden_layers))]) with tf.variable_scope('train_encoder_scope') as encoder_train_scope: training_encoder_outputs, training_encoder_state = tf.nn.dynamic_rnn(cell=multi_layered_encoder_cell, inputs=training_input, sequence_length=input_sequence_length, dtype=tf.float32) with tf.variable_scope(encoder_train_scope, reuse=tf.AUTO_REUSE) as encoder_inference_scope: inference_encoder_outputs, inference_encoder_state = tf.nn.dynamic_rnn(cell=multi_layered_encoder_cell, inputs=testing_input, sequence_length=input_sequence_length, dtype=tf.float32) # the final projection layer to convert the output to the desired dimension dense_layer = Dense(units=1, use_bias=self.__use_bias, kernel_initializer=weight_initializer) # decoder cell of the decoder network multi_layered_decoder_cell = tf.nn.rnn_cell.MultiRNNCell( cells=[cell() for _ in range(int(num_hidden_layers))]) # building the decoder network for training with tf.variable_scope('decoder_train_scope') as decoder_train_scope: # create the initial state for the decoder training_helper = tf.contrib.seq2seq.ScheduledOutputTrainingHelper(inputs=decoder_input, sequence_length=output_sequence_length, sampling_probability=0.0, name="training_helper") training_decoder = tf.contrib.seq2seq.BasicDecoder(cell=multi_layered_decoder_cell, helper=training_helper, initial_state=training_encoder_state, output_layer=dense_layer) # perform the decoding training_decoder_outputs, _, _ = tf.contrib.seq2seq.dynamic_decode(decoder=training_decoder) # building the decoder network for inference with tf.variable_scope(decoder_train_scope, reuse=tf.AUTO_REUSE) as decoder_inference_scope: # create the initial state for the decoder inference_helper = tf.contrib.seq2seq.ScheduledOutputTrainingHelper(inputs=decoder_input, sequence_length=output_sequence_length, sampling_probability=1.0, name="inference_helper") inference_decoder = tf.contrib.seq2seq.BasicDecoder(cell=multi_layered_decoder_cell, helper=inference_helper, initial_state=inference_encoder_state, output_layer=dense_layer) # perform the decoding inference_decoder_outputs, _, _ = tf.contrib.seq2seq.dynamic_decode(decoder=inference_decoder) # error that should be minimized in the training process error = self.__l1_loss(training_decoder_outputs[0], training_target) # l2 regularization of the trainable model parameters l2_loss = 0.0 for var in tf.trainable_variables(): l2_loss += tf.nn.l2_loss(var) l2_loss = tf.multiply(tf.cast(l2_regularization, dtype=tf.float64), tf.cast(l2_loss, dtype=tf.float64)) total_loss = tf.cast(error, dtype=tf.float64) + l2_loss # create the optimizer optimizer = optimizer_fn(total_loss) # create the Dataset objects for the training and test data training_dataset = tf.data.TFRecordDataset(filenames = [self.__binary_train_file_path], compression_type = "ZLIB") test_dataset = tf.data.TFRecordDataset([self.__binary_test_file_path], compression_type = "ZLIB") # parse the records tfrecord_reader = TFRecordReader() # preparing the training data # randomly shuffle the time series within the dataset shuffle_seed = tf.placeholder(dtype=tf.int64, shape=[]) # training_dataset = training_dataset.apply(tf.data.experimental.shuffle_and_repeat(buffer_size=training_data_configs.SHUFFLE_BUFFER_SIZE, # count=int(max_epoch_size), seed=shuffle_seed)) training_dataset = training_dataset.repeat(count=int(max_epoch_size)) training_dataset = training_dataset.map(tfrecord_reader.validation_data_parser) # create the batches by padding the datasets to make the variable sequence lengths fixed within the individual batches padded_training_data_batches = training_dataset.padded_batch(batch_size=int(minibatch_size), padded_shapes=([], [tf.Dimension(None), 1], [self.__output_size, 1], [self.__meta_data_size, 1])) # get an iterator to the batches training_data_batch_iterator = padded_training_data_batches.make_initializable_iterator() # access each batch using the iterator next_training_data_batch = training_data_batch_iterator.get_next() # preparing the test data test_dataset = test_dataset.map(tfrecord_reader.test_data_parser) # create a single batch from all the test time series by padding the datasets to make the variable sequence lengths fixed padded_test_input_data = test_dataset.padded_batch(batch_size=int(minibatch_size), padded_shapes=( [], [tf.Dimension(None), 1], [self.__meta_data_size, 1])) # get an iterator to the test input data batch test_input_iterator = padded_test_input_data.make_one_shot_iterator() # access the test input batch using the iterator test_input_data_batch = test_input_iterator.get_next() # setup variable initialization init_op = tf.global_variables_initializer() # define the GPU options gpu_options = tf.GPUOptions(allow_growth=True) with tf.Session( config=tf.ConfigProto(log_device_placement=gpu_configs.log_device_placement, allow_soft_placement=True, gpu_options=gpu_options)) as session: session.run(init_op) for epoch in range(int(max_num_epochs)): print("Epoch->", epoch) session.run(training_data_batch_iterator.initializer, feed_dict={shuffle_seed:epoch}) while True: try: next_training_batch_value = session.run(next_training_data_batch, feed_dict={shuffle_seed:epoch}) decoder_input_value = np.hstack((np.expand_dims(next_training_batch_value[1][:, -1, :], axis=1), next_training_batch_value[2][:, :-1, :])) # model training loss, _ = session.run([total_loss, optimizer], feed_dict={input: next_training_batch_value[1], training_target: next_training_batch_value[2], decoder_input: decoder_input_value, input_sequence_length: next_training_batch_value[0], output_sequence_length: [self.__output_size] * np.shape(next_training_batch_value[1])[0]}) except tf.errors.OutOfRangeError: break # applying the model to the test data list_of_forecasts = [] while True: try: # get the batch of test inputs test_input_batch_value = session.run(test_input_data_batch) # shape for the target data decoder_input_shape = [np.shape(test_input_batch_value[1])[0], self.__output_size, 1] # get the output of the network for the test input data batch test_output = session.run(inference_decoder_outputs[0], feed_dict={input: test_input_batch_value[1], decoder_input: np.zeros(decoder_input_shape), input_sequence_length: test_input_batch_value[0], output_sequence_length: [self.__output_size] * np.shape(test_input_batch_value[1])[0]}) forecasts = test_output list_of_forecasts.extend(forecasts.tolist()) except tf.errors.OutOfRangeError: break return np.squeeze(list_of_forecasts, axis = 2) #the third dimension is squeezed since it is one
12,870
53.538136
146
py
time-series-forecasting-release
time-series-forecasting-release/rnn_architectures/seq2seq_model/with_dense_layer/moving_window/unaccumulated_error/seq2seq_model_trainer.py
import numpy as np import tensorflow as tf from tfrecords_handler.moving_window.tfrecord_reader import TFRecordReader from configs.global_configs import model_training_configs from configs.global_configs import training_data_configs from configs.global_configs import gpu_configs class Seq2SeqModelTrainerWithDenseLayer: def __init__(self, **kwargs): self.__use_bias = kwargs["use_bias"] self.__use_peepholes = kwargs["use_peepholes"] self.__input_size = kwargs["input_size"] self.__output_size = kwargs["output_size"] self.__binary_train_file_path = kwargs["binary_train_file_path"] self.__binary_validation_file_path = kwargs["binary_validation_file_path"] self.__contain_zero_values = kwargs["contain_zero_values"] self.__address_near_zero_instability = kwargs["address_near_zero_instability"] self.__integer_conversion = kwargs["integer_conversion"] self.__seed = kwargs["seed"] self.__cell_type = kwargs["cell_type"] self.__without_stl_decomposition = kwargs['without_stl_decomposition'] # define the metadata size based on the usage of stl decomposition if self.__without_stl_decomposition: self.__meta_data_size = 1 else: self.__meta_data_size = self.__output_size + 1 def __l1_loss(self, z, t): loss = tf.reduce_mean(tf.abs(t - z)) return loss # Training the time series def train_model(self, **kwargs): num_hidden_layers = int(kwargs['num_hidden_layers']) cell_dimension = int(kwargs["cell_dimension"]) minibatch_size = int(kwargs["minibatch_size"]) max_epoch_size = int(kwargs["max_epoch_size"]) max_num_epochs = int(kwargs["max_num_epochs"]) l2_regularization = kwargs["l2_regularization"] gaussian_noise_stdev = kwargs["gaussian_noise_stdev"] random_normal_initializer_stdev = kwargs['random_normal_initializer_stdev'] optimizer_fn = kwargs["optimizer_fn"] tf.reset_default_graph() tf.set_random_seed(self.__seed) # adding noise to the input input = tf.placeholder(dtype=tf.float32, shape=[None, None, self.__input_size]) validation_input = input noise = tf.random_normal(shape=tf.shape(input), mean=0.0, stddev=gaussian_noise_stdev, dtype=tf.float32) training_input = input + noise target = tf.placeholder(dtype=tf.float32, shape=[None, None, self.__output_size]) # placeholder for the sequence lengths sequence_length = tf.placeholder(dtype=tf.int32, shape=[None]) weight_initializer = tf.truncated_normal_initializer(stddev=random_normal_initializer_stdev) # create a tensor array for the indices of the encoder outputs array and the target new_index_array = tf.range(start=0, limit=tf.shape(sequence_length)[0], delta=1) output_array_indices = tf.stack([new_index_array, sequence_length - 1], axis=-1) actual_targets = tf.gather_nd(params=target, indices=output_array_indices) actual_targets = tf.expand_dims(input=actual_targets, axis=1) # create the model architecture # RNN with the layer of cells def cell(): if self.__cell_type == "LSTM": cell = tf.nn.rnn_cell.LSTMCell(num_units=int(cell_dimension), use_peepholes=self.__use_peepholes, initializer=weight_initializer) elif self.__cell_type == "GRU": cell = tf.nn.rnn_cell.GRUCell(num_units=int(cell_dimension), kernel_initializer=weight_initializer) elif self.__cell_type == "RNN": cell = tf.nn.rnn_cell.BasicRNNCell(num_units=int(cell_dimension)) return cell # building the encoder network multi_layered_encoder_cell = tf.nn.rnn_cell.MultiRNNCell( cells=[cell() for _ in range(int(num_hidden_layers))]) with tf.variable_scope('train_encoder_scope') as encoder_train_scope: training_encoder_outputs, training_encoder_state = tf.nn.dynamic_rnn(cell=multi_layered_encoder_cell, inputs=training_input, sequence_length=sequence_length, dtype=tf.float32) with tf.variable_scope(encoder_train_scope, reuse=tf.AUTO_REUSE) as encoder_inference_scope: inference_encoder_outputs, inference_encoder_states = tf.nn.dynamic_rnn(cell=multi_layered_encoder_cell, inputs=validation_input, sequence_length=sequence_length, dtype=tf.float32) # building the decoder network for training with tf.variable_scope('dense_layer_train_scope') as dense_layer_train_scope: train_final_timestep_predictions = tf.gather_nd(params=training_encoder_outputs, indices=output_array_indices) # the final projection layer to convert the encoder_outputs to the desired dimension train_prediction_output = tf.layers.dense( inputs=tf.convert_to_tensor(value=train_final_timestep_predictions, dtype=tf.float32), units=self.__output_size, use_bias=self.__use_bias, kernel_initializer=weight_initializer) train_prediction_output = tf.expand_dims(input=train_prediction_output, axis=1) # building the decoder network for inference with tf.variable_scope(dense_layer_train_scope, reuse=tf.AUTO_REUSE) as dense_layer_inference_scope: inference_final_timestep_predictions = tf.gather_nd(params=inference_encoder_outputs, indices=output_array_indices) # the final projection layer to convert the encoder_outputs to the desired dimension inference_prediction_output = tf.layers.dense( inputs=tf.convert_to_tensor(value=inference_final_timestep_predictions, dtype=tf.float32), units=self.__output_size, use_bias=self.__use_bias, kernel_initializer=weight_initializer) inference_prediction_output = tf.expand_dims(input=inference_prediction_output, axis=1) # error that should be minimized in the training process error = self.__l1_loss(train_prediction_output, actual_targets) # l2 regularization of the trainable model parameters l2_loss = 0.0 for var in tf.trainable_variables(): l2_loss += tf.nn.l2_loss(var) l2_loss = tf.multiply(tf.cast(l2_regularization, dtype=tf.float64), tf.cast(l2_loss, dtype=tf.float64)) total_loss = tf.cast(error, dtype=tf.float64) + l2_loss # create the optimizer optimizer = optimizer_fn(total_loss) # create the training and validation datasets from the tfrecord files training_dataset = tf.data.TFRecordDataset(filenames=[self.__binary_train_file_path], compression_type="ZLIB") validation_dataset = tf.data.TFRecordDataset(filenames=[self.__binary_validation_file_path], compression_type="ZLIB") # parse the records tfrecord_reader = TFRecordReader(self.__input_size, self.__output_size, self.__meta_data_size) # define the expected shapes of data after padding train_padded_shapes = ([], [tf.Dimension(None), self.__input_size], [tf.Dimension(None), self.__output_size]) validation_padded_shapes = ( [], [tf.Dimension(None), self.__input_size], [tf.Dimension(None), self.__output_size], [tf.Dimension(None), self.__meta_data_size]) # preparing the training data shuffle_seed = tf.placeholder(dtype=tf.int64, shape=[]) # training_dataset = training_dataset.apply( # tf.data.experimental.shuffle_and_repeat(buffer_size=training_data_configs.SHUFFLE_BUFFER_SIZE, # count=int(max_epoch_size), seed=shuffle_seed)) training_dataset = training_dataset.repeat(count=int(max_epoch_size)) training_dataset = training_dataset.map(tfrecord_reader.train_data_parser) padded_training_data_batches = training_dataset.padded_batch(batch_size=int(minibatch_size), padded_shapes=train_padded_shapes) training_data_batch_iterator = padded_training_data_batches.make_initializable_iterator() next_training_data_batch = training_data_batch_iterator.get_next() # preparing the validation data validation_dataset = validation_dataset.map(tfrecord_reader.validation_data_parser) # create a single batch from all the validation time series by padding the datasets to make the variable sequence lengths fixed padded_validation_dataset = validation_dataset.padded_batch(batch_size=int(minibatch_size), padded_shapes=validation_padded_shapes) # get an iterator to the validation data validation_data_iterator = padded_validation_dataset.make_initializable_iterator() # access the validation data using the iterator next_validation_data_batch = validation_data_iterator.get_next() # setup variable initialization init_op = tf.global_variables_initializer() # define the GPU options gpu_options = tf.GPUOptions(allow_growth=True) with tf.Session( config=tf.ConfigProto(log_device_placement=gpu_configs.log_device_placement, allow_soft_placement=True, gpu_options=gpu_options)) as session: session.run(init_op) smape_final = 0.0 smape_list = [] for epoch in range(max_num_epochs): print("Epoch->", epoch) session.run(training_data_batch_iterator.initializer, feed_dict={shuffle_seed: epoch}) while True: try: training_data_batch_value = session.run(next_training_data_batch, feed_dict={shuffle_seed: epoch}) total_loss_value, _ = session.run([total_loss, optimizer], feed_dict={input: training_data_batch_value[1], target: training_data_batch_value[2], sequence_length: training_data_batch_value[0] }) except tf.errors.OutOfRangeError: break session.run(validation_data_iterator.initializer) while True: try: # get the batch of validation inputs validation_data_batch_value = session.run(next_validation_data_batch) # get the output of the network for the validation input data batch validation_output = session.run(inference_prediction_output, feed_dict={input: validation_data_batch_value[1], sequence_length: validation_data_batch_value[0] }) # calculate the smape for the validation data using vectorization last_indices = validation_data_batch_value[0] - 1 array_first_dimension = np.array(range(0, validation_data_batch_value[0].shape[0])) true_seasonality_values = validation_data_batch_value[3][array_first_dimension, last_indices, 1:] level_values = validation_data_batch_value[3][array_first_dimension, last_indices, 0] actual_values = validation_data_batch_value[2][array_first_dimension, last_indices, :] if self.__without_stl_decomposition: converted_actual_values = np.exp(actual_values) converted_validation_output = np.exp(validation_output) else: converted_actual_values = np.exp( true_seasonality_values + level_values[:, np.newaxis] + actual_values) converted_validation_output = np.exp( true_seasonality_values + level_values[:, np.newaxis] + validation_output) if (self.__contain_zero_values): # to compensate for 0 values in data converted_validation_output = converted_validation_output - 1 converted_actual_values = converted_actual_values - 1 if self.__without_stl_decomposition: converted_actual_values = converted_actual_values * level_values[:, np.newaxis] converted_validation_output = converted_validation_output * level_values[:, np.newaxis] if self.__integer_conversion: converted_validation_output = np.round(converted_validation_output) converted_actual_values = np.round(converted_actual_values) converted_validation_output[converted_validation_output < 0] = 0 converted_actual_values[converted_actual_values < 0] = 0 if self.__address_near_zero_instability: # calculate the smape epsilon = 0.1 sum = np.maximum( np.abs(converted_validation_output) + np.abs(converted_actual_values) + epsilon, 0.5 + epsilon) smape_values = (np.abs(converted_validation_output - converted_actual_values) / sum) * 2 smape_values_per_series = np.mean(smape_values, axis=1) smape_list.extend(smape_values_per_series) else: # calculate the smape smape_values = (np.abs(converted_validation_output - converted_actual_values) / (np.abs(converted_validation_output) + np.abs(converted_actual_values))) * 2 smape_values_per_series = np.mean(smape_values, axis=1) smape_list.extend(smape_values_per_series) except tf.errors.OutOfRangeError: break smape_final = np.mean(smape_list) print("SMAPE value: {}".format(smape_final)) session.close() return float(smape_final), smape_list
15,412
52.703833
135
py
time-series-forecasting-release
time-series-forecasting-release/rnn_architectures/seq2seq_model/with_dense_layer/moving_window/unaccumulated_error/seq2seq_model_tester.py
import numpy as np import tensorflow as tf from tfrecords_handler.moving_window.tfrecord_reader import TFRecordReader from configs.global_configs import training_data_configs from configs.global_configs import gpu_configs class Seq2SeqModelTesterWithDenseLayer: def __init__(self, **kwargs): self.__use_bias = kwargs["use_bias"] self.__use_peepholes = kwargs["use_peepholes"] self.__input_size = kwargs["input_size"] self.__output_size = kwargs["output_size"] self.__binary_train_file_path = kwargs["binary_train_file_path"] self.__binary_test_file_path = kwargs["binary_test_file_path"] self.__seed = kwargs["seed"] self.__cell_type = kwargs["cell_type"] self.__without_stl_decomposition = kwargs["without_stl_decomposition"] # define the metadata size based on the usage of stl decomposition if self.__without_stl_decomposition: self.__meta_data_size = 1 else: self.__meta_data_size = self.__output_size + 1 def __l1_loss(self, z, t): loss = tf.reduce_mean(tf.abs(t - z)) return loss # Training the time series def test_model(self, **kwargs): # optimized hyperparameters num_hidden_layers = int(kwargs['num_hidden_layers']) max_num_epochs = int(kwargs['max_num_epochs']) max_epoch_size = int(kwargs['max_epoch_size']) cell_dimension = int(kwargs['cell_dimension']) l2_regularization = kwargs['l2_regularization'] minibatch_size = int(kwargs['minibatch_size']) gaussian_noise_stdev = kwargs['gaussian_noise_stdev'] random_normal_initializer_stdev = kwargs['random_normal_initializer_stdev'] optimizer_fn = kwargs['optimizer_fn'] # reset the tensorflow graph tf.reset_default_graph() tf.set_random_seed(self.__seed) # declare the input and output placeholders # adding noise to the input input = tf.placeholder(dtype=tf.float32, shape=[None, None, self.__input_size]) testing_input = input noise = tf.random_normal(shape=tf.shape(input), mean=0.0, stddev=gaussian_noise_stdev, dtype=tf.float32) training_input = input + noise target = tf.placeholder(dtype=tf.float32, shape=[None, None, self.__output_size]) # placeholder for the sequence lengths sequence_length = tf.placeholder(dtype=tf.int32, shape=[None]) weight_initializer = tf.truncated_normal_initializer(stddev=random_normal_initializer_stdev) # create a tensor array for the indices of the encoder outputs array and the target new_index_array = tf.range(start=0, limit=tf.shape(sequence_length)[0], delta=1) output_array_indices = tf.stack([new_index_array, sequence_length - 1], axis=-1) actual_targets = tf.gather_nd(params=target, indices=output_array_indices) actual_targets = tf.expand_dims(input=actual_targets, axis=1) # create the model architecture # RNN with the layer of cells def cell(): if self.__cell_type == "LSTM": cell = tf.nn.rnn_cell.LSTMCell(num_units=int(cell_dimension), use_peepholes=self.__use_peepholes, initializer=weight_initializer) elif self.__cell_type == "GRU": cell = tf.nn.rnn_cell.GRUCell(num_units=int(cell_dimension), kernel_initializer=weight_initializer) elif self.__cell_type == "RNN": cell = tf.nn.rnn_cell.BasicRNNCell(num_units=int(cell_dimension)) return cell # building the encoder network multi_layered_encoder_cell = tf.nn.rnn_cell.MultiRNNCell( cells=[cell() for _ in range(int(num_hidden_layers))]) with tf.variable_scope('train_encoder_scope') as encoder_train_scope: training_encoder_outputs, training_encoder_state = tf.nn.dynamic_rnn(cell=multi_layered_encoder_cell, inputs=training_input, sequence_length=sequence_length, dtype=tf.float32) with tf.variable_scope(encoder_train_scope, reuse=tf.AUTO_REUSE) as encoder_inference_scope: inference_encoder_outputs, inference_encoder_states = tf.nn.dynamic_rnn(cell=multi_layered_encoder_cell, inputs=testing_input, sequence_length=sequence_length, dtype=tf.float32) # building the decoder network for training with tf.variable_scope('dense_layer_train_scope') as dense_layer_train_scope: train_final_timestep_predictions = tf.gather_nd(params=training_encoder_outputs, indices=output_array_indices) # the final projection layer to convert the encoder_outputs to the desired dimension train_prediction_output = tf.layers.dense( inputs=tf.convert_to_tensor(value=train_final_timestep_predictions, dtype=tf.float32), units=self.__output_size, use_bias=self.__use_bias, kernel_initializer=weight_initializer) train_prediction_output = tf.expand_dims(input=train_prediction_output, axis=1) # building the decoder network for inference with tf.variable_scope(dense_layer_train_scope, reuse=tf.AUTO_REUSE) as dense_layer_inference_scope: inference_final_timestep_predictions = tf.gather_nd(params=inference_encoder_outputs, indices=output_array_indices) # the final projection layer to convert the encoder_outputs to the desired dimension inference_prediction_output = tf.layers.dense( inputs=tf.convert_to_tensor(value=inference_final_timestep_predictions, dtype=tf.float32), units=self.__output_size, use_bias=self.__use_bias, kernel_initializer=weight_initializer) inference_prediction_output = tf.expand_dims(input=inference_prediction_output, axis=1) # error that should be minimized in the training process error = self.__l1_loss(train_prediction_output, actual_targets) # l2 regularization of the trainable model parameters l2_loss = 0.0 for var in tf.trainable_variables(): l2_loss += tf.nn.l2_loss(var) l2_loss = tf.multiply(tf.cast(l2_regularization, dtype=tf.float64), tf.cast(l2_loss, dtype=tf.float64)) total_loss = tf.cast(error, dtype=tf.float64) + l2_loss # create the optimizer optimizer = optimizer_fn(total_loss) # create the Dataset objects for the training and test data training_dataset = tf.data.TFRecordDataset(filenames = [self.__binary_train_file_path], compression_type = "ZLIB") test_dataset = tf.data.TFRecordDataset([self.__binary_test_file_path], compression_type = "ZLIB") # parse the records tfrecord_reader = TFRecordReader(self.__input_size, self.__output_size, self.__meta_data_size) # preparing the training data # randomly shuffle the time series within the dataset shuffle_seed = tf.placeholder(dtype=tf.int64, shape=[]) # training_dataset = training_dataset.apply(tf.data.experimental.shuffle_and_repeat(buffer_size=training_data_configs.SHUFFLE_BUFFER_SIZE, # count=int(max_epoch_size), seed=shuffle_seed)) training_dataset = training_dataset.repeat(count=int(max_epoch_size)) training_dataset = training_dataset.map(tfrecord_reader.validation_data_parser) # create the batches by padding the datasets to make the variable sequence lengths fixed within the individual batches padded_training_data_batches = training_dataset.padded_batch(batch_size=int(minibatch_size), padded_shapes=([], [tf.Dimension(None), self.__input_size], [tf.Dimension(None), self.__output_size], [tf.Dimension(None), self.__meta_data_size])) # get an iterator to the batches training_data_batch_iterator = padded_training_data_batches.make_initializable_iterator() # access each batch using the iterator next_training_data_batch = training_data_batch_iterator.get_next() # preparing the test data test_dataset = test_dataset.map(tfrecord_reader.test_data_parser) # create a single batch from all the test time series by padding the datasets to make the variable sequence lengths fixed padded_test_input_data = test_dataset.padded_batch(batch_size=int(minibatch_size), padded_shapes=( [], [tf.Dimension(None),self.__input_size], [tf.Dimension(None), self.__meta_data_size])) # get an iterator to the test input data batch test_input_iterator = padded_test_input_data.make_one_shot_iterator() # access the test input batch using the iterator test_input_data_batch = test_input_iterator.get_next() # setup variable initialization init_op = tf.global_variables_initializer() # define the GPU options gpu_options = tf.GPUOptions(allow_growth=True) with tf.Session( config=tf.ConfigProto(log_device_placement=gpu_configs.log_device_placement, allow_soft_placement=True, gpu_options=gpu_options)) as session: session.run(init_op) for epoch in range(int(max_num_epochs)): print("Epoch->", epoch) session.run(training_data_batch_iterator.initializer, feed_dict={shuffle_seed: epoch}) losses = [] while True: try: next_training_batch_value = session.run(next_training_data_batch, feed_dict={shuffle_seed: epoch}) # model training _, loss_val = session.run([optimizer, total_loss], feed_dict={input: next_training_batch_value[1], target: next_training_batch_value[2], sequence_length: next_training_batch_value[0], }) losses.append(loss_val) except tf.errors.OutOfRangeError: break # applying the model to the test data list_of_forecasts = [] while True: try: # get the batch of test inputs test_input_batch_value = session.run(test_input_data_batch) # get the output of the network for the test input data batch test_output = session.run(inference_prediction_output, feed_dict={input: test_input_batch_value[1], sequence_length: test_input_batch_value[0], }) forecasts = test_output list_of_forecasts.extend(forecasts.tolist()) except tf.errors.OutOfRangeError: break return np.squeeze(list_of_forecasts, axis = 1) #the second dimension is squeezed since it is one
11,874
51.083333
170
py
time-series-forecasting-release
time-series-forecasting-release/rnn_architectures/seq2seq_model/with_dense_layer/non_moving_window/unaccumulated_error/seq2seq_model_trainer.py
import numpy as np import tensorflow as tf from tfrecords_handler.non_moving_window.tfrecord_reader import TFRecordReader from configs.global_configs import model_training_configs from configs.global_configs import training_data_configs from configs.global_configs import gpu_configs class Seq2SeqModelTrainerWithDenseLayer: def __init__(self, **kwargs): self.__use_bias = kwargs["use_bias"] self.__use_peepholes = kwargs["use_peepholes"] self.__output_size = kwargs["output_size"] self.__binary_train_file_path = kwargs["binary_train_file_path"] self.__binary_validation_file_path = kwargs["binary_validation_file_path"] self.__contain_zero_values = kwargs["contain_zero_values"] self.__address_near_zero_instability = kwargs["address_near_zero_instability"] self.__integer_conversion = kwargs["integer_conversion"] self.__seed = kwargs["seed"] self.__cell_type = kwargs["cell_type"] self.__without_stl_decomposition = kwargs['without_stl_decomposition'] def __l1_loss(self, z, t): loss = tf.reduce_mean(tf.abs(t - z)) return loss # Training the time series def train_model(self, **kwargs): num_hidden_layers = int(kwargs['num_hidden_layers']) cell_dimension = int(kwargs["cell_dimension"]) minibatch_size = int(kwargs["minibatch_size"]) max_epoch_size = int(kwargs["max_epoch_size"]) max_num_epochs = int(kwargs["max_num_epochs"]) l2_regularization = kwargs["l2_regularization"] gaussian_noise_stdev = kwargs["gaussian_noise_stdev"] random_normal_initializer_stdev = kwargs['random_normal_initializer_stdev'] optimizer_fn = kwargs["optimizer_fn"] tf.reset_default_graph() tf.set_random_seed(self.__seed) # adding noise to the input input = tf.placeholder(dtype=tf.float32, shape=[None, None, 1]) validation_input = input noise = tf.random_normal(shape=tf.shape(input), mean=0.0, stddev=gaussian_noise_stdev, dtype=tf.float32) training_input = input + noise target = tf.placeholder(dtype=tf.float32, shape=[None, self.__output_size, 1]) # placeholder for the sequence lengths sequence_length = tf.placeholder(dtype=tf.int32, shape=[None]) weight_initializer = tf.truncated_normal_initializer(stddev=random_normal_initializer_stdev) # create the model architecture # RNN with the layer of cells def cell(): if self.__cell_type == "LSTM": cell = tf.nn.rnn_cell.LSTMCell(num_units=int(cell_dimension), use_peepholes=self.__use_peepholes, initializer=weight_initializer) elif self.__cell_type == "GRU": cell = tf.nn.rnn_cell.GRUCell(num_units=int(cell_dimension), kernel_initializer=weight_initializer) elif self.__cell_type == "RNN": cell = tf.nn.rnn_cell.BasicRNNCell(num_units=int(cell_dimension)) return cell # building the encoder network multi_layered_encoder_cell = tf.nn.rnn_cell.MultiRNNCell( cells=[cell() for _ in range(int(num_hidden_layers))]) with tf.variable_scope('train_encoder_scope') as encoder_train_scope: training_encoder_outputs, training_encoder_state = tf.nn.dynamic_rnn(cell=multi_layered_encoder_cell, inputs=training_input, sequence_length=sequence_length, dtype=tf.float32) with tf.variable_scope(encoder_train_scope, reuse=tf.AUTO_REUSE) as encoder_inference_scope: inference_encoder_outputs, inference_encoder_states = tf.nn.dynamic_rnn(cell=multi_layered_encoder_cell, inputs=validation_input, sequence_length=sequence_length, dtype=tf.float32) # create a tensor array for the indices of the encoder outputs array new_index_array = tf.range(start=0, limit=tf.shape(sequence_length)[0], delta=1) output_array_indices = tf.stack([new_index_array, sequence_length - 1], axis=-1) # building the decoder network for training with tf.variable_scope('dense_layer_train_scope') as dense_layer_train_scope: train_final_timestep_predictions = tf.gather_nd(params=training_encoder_outputs, indices=output_array_indices) # the final projection layer to convert the encoder_outputs to the desired dimension train_prediction_output = tf.layers.dense( inputs=tf.convert_to_tensor(value=train_final_timestep_predictions, dtype=tf.float32), units=self.__output_size, use_bias=self.__use_bias, kernel_initializer=weight_initializer) train_prediction_output = tf.expand_dims(input=train_prediction_output, axis=2) # building the decoder network for inference with tf.variable_scope(dense_layer_train_scope, reuse=tf.AUTO_REUSE) as dense_layer_inference_scope: inference_final_timestep_predictions = tf.gather_nd(params=inference_encoder_outputs, indices=output_array_indices) # the final projection layer to convert the encoder_outputs to the desired dimension inference_prediction_output = tf.layers.dense( inputs=tf.convert_to_tensor(value=inference_final_timestep_predictions, dtype=tf.float32), units=self.__output_size, use_bias=self.__use_bias, kernel_initializer=weight_initializer) inference_prediction_output = tf.expand_dims(input=inference_prediction_output, axis=2) # error that should be minimized in the training process error = self.__l1_loss(train_prediction_output, target) # l2 regularization of the trainable model parameters l2_loss = 0.0 for var in tf.trainable_variables(): l2_loss += tf.nn.l2_loss(var) l2_loss = tf.multiply(tf.cast(l2_regularization, dtype=tf.float64), tf.cast(l2_loss, dtype=tf.float64)) total_loss = tf.cast(error, dtype=tf.float64) + l2_loss # create the optimizer optimizer = optimizer_fn(total_loss) # create the training and validation datasets from the tfrecord files training_dataset = tf.data.TFRecordDataset(filenames = [self.__binary_train_file_path], compression_type = "ZLIB") validation_dataset = tf.data.TFRecordDataset(filenames = [self.__binary_validation_file_path], compression_type = "ZLIB") # parse the records tfrecord_reader = TFRecordReader() # define the expected shapes of data after padding train_padded_shapes = ([], [tf.Dimension(None), 1], [self.__output_size, 1]) validation_padded_shapes = ([], [tf.Dimension(None), 1], [self.__output_size, 1], [self.__meta_data_size, 1]) # preparing the training data shuffle_seed = tf.placeholder(dtype=tf.int64, shape=[]) # training_dataset = training_dataset.apply(tf.data.experimental.shuffle_and_repeat(buffer_size=training_data_configs.SHUFFLE_BUFFER_SIZE, # count=int(max_epoch_size), seed=shuffle_seed)) training_dataset = training_dataset.repeat(count=int(max_epoch_size)) training_dataset = training_dataset.map(tfrecord_reader.train_data_parser) padded_training_data_batches = training_dataset.padded_batch(batch_size=int(minibatch_size), padded_shapes=train_padded_shapes) training_data_batch_iterator = padded_training_data_batches.make_initializable_iterator() next_training_data_batch = training_data_batch_iterator.get_next() # preparing the validation data validation_dataset = validation_dataset.map(tfrecord_reader.validation_data_parser) # create a single batch from all the validation time series by padding the datasets to make the variable sequence lengths fixed padded_validation_dataset = validation_dataset.padded_batch(batch_size=int(minibatch_size), padded_shapes=validation_padded_shapes) # get an iterator to the validation data validation_data_iterator = padded_validation_dataset.make_initializable_iterator() # access the validation data using the iterator next_validation_data_batch = validation_data_iterator.get_next() # setup variable initialization init_op = tf.global_variables_initializer() # define the GPU options gpu_options = tf.GPUOptions(allow_growth=True) with tf.Session( config=tf.ConfigProto(log_device_placement=gpu_configs.log_device_placement, allow_soft_placement=True, gpu_options=gpu_options)) as session: session.run(init_op) smape_final = 0.0 smape_list = [] for epoch in range(max_num_epochs): print("Epoch->", epoch) session.run(training_data_batch_iterator.initializer, feed_dict={shuffle_seed:epoch}) losses = [] while True: try: training_data_batch_value = session.run(next_training_data_batch, feed_dict={shuffle_seed:epoch}) total_loss_value, _ = session.run([total_loss, optimizer], feed_dict={input: training_data_batch_value[1], target: training_data_batch_value[2], sequence_length: training_data_batch_value[0] }) losses.append(total_loss_value) except tf.errors.OutOfRangeError: break session.run(validation_data_iterator.initializer) while True: try: # get the batch of validation inputs validation_data_batch_value = session.run(next_validation_data_batch) # shape for the target data target_data_shape = [np.shape(validation_data_batch_value[1])[0], self.__output_size, 1] # get the output of the network for the validation input data batch validation_output = session.run(inference_prediction_output, feed_dict={input: validation_data_batch_value[1], target: np.zeros(target_data_shape), sequence_length: validation_data_batch_value[0] }) # calculate the smape for the validation data using vectorization # convert the data to remove the preprocessing true_seasonality_values = validation_data_batch_value[3][:, 1:, 0] level_values = validation_data_batch_value[3][:, 0, 0] actual_values = validation_data_batch_value[2] if self.__without_stl_decomposition: converted_validation_output = np.exp(np.squeeze(validation_output, axis=2)) converted_actual_values = np.exp(np.squeeze(actual_values, axis=2)) else: converted_validation_output = np.exp( true_seasonality_values + level_values[:, np.newaxis] + np.squeeze(validation_output, axis=2)) converted_actual_values = np.exp( true_seasonality_values + level_values[:, np.newaxis] + np.squeeze(actual_values, axis=2)) if (self.__contain_zero_values): # to compensate for 0 values in data converted_validation_output = converted_validation_output - 1 converted_actual_values = converted_actual_values - 1 if self.__without_stl_decomposition: converted_validation_output = converted_validation_output * level_values[:, np.newaxis] converted_actual_values = converted_actual_values * level_values[:, np.newaxis] if self.__integer_conversion: converted_validation_output = np.round(converted_validation_output) converted_actual_values = np.round(converted_actual_values) converted_validation_output[converted_validation_output < 0] = 0 converted_actual_values[converted_actual_values < 0] = 0 if self.__address_near_zero_instability: # calculate the smape epsilon = 0.1 sum = np.maximum( np.abs(converted_validation_output) + np.abs(converted_actual_values) + epsilon, 0.5 + epsilon) smape_values = (np.abs(converted_validation_output - converted_actual_values) / sum) * 2 smape_values_per_series = np.mean(smape_values, axis=1) smape_list.extend(smape_values_per_series) else: # calculate the smape smape_values = (np.abs(converted_validation_output - converted_actual_values) / (np.abs(converted_validation_output) + np.abs(converted_actual_values))) * 2 smape_values_per_series = np.mean(smape_values, axis=1) smape_list.extend(smape_values_per_series) except tf.errors.OutOfRangeError: break smape_final = np.mean(smape_list) print("SMAPE value: {}".format(smape_final)) session.close() return float(smape_final), smape_list
14,802
53.223443
146
py
time-series-forecasting-release
time-series-forecasting-release/rnn_architectures/seq2seq_model/with_dense_layer/non_moving_window/unaccumulated_error/seq2seq_model_tester.py
import numpy as np import tensorflow as tf from tfrecords_handler.non_moving_window.tfrecord_reader import TFRecordReader from configs.global_configs import training_data_configs from configs.global_configs import gpu_configs class Seq2SeqModelTesterWithDenseLayer: def __init__(self, **kwargs): self.__use_bias = kwargs["use_bias"] self.__use_peepholes = kwargs["use_peepholes"] self.__output_size = kwargs["output_size"] self.__binary_train_file_path = kwargs["binary_train_file_path"] self.__binary_test_file_path = kwargs["binary_test_file_path"] self.__seed = kwargs["seed"] self.__cell_type = kwargs["cell_type"] def __l1_loss(self, z, t): loss = tf.reduce_mean(tf.abs(t - z)) return loss # Training the time series def test_model(self, **kwargs): # optimized hyperparameters num_hidden_layers = int(kwargs['num_hidden_layers']) max_num_epochs = int(kwargs['max_num_epochs']) max_epoch_size = int(kwargs['max_epoch_size']) cell_dimension = int(kwargs['cell_dimension']) l2_regularization = kwargs['l2_regularization'] minibatch_size = int(kwargs['minibatch_size']) gaussian_noise_stdev = kwargs['gaussian_noise_stdev'] random_normal_initializer_stdev = kwargs['random_normal_initializer_stdev'] optimizer_fn = kwargs['optimizer_fn'] # reset the tensorflow graph tf.reset_default_graph() tf.set_random_seed(self.__seed) # declare the input and output placeholders # adding noise to the input input = tf.placeholder(dtype=tf.float32, shape=[None, None, 1]) testing_input = input noise = tf.random_normal(shape=tf.shape(input), mean=0.0, stddev=gaussian_noise_stdev, dtype=tf.float32) training_input = input + noise target = tf.placeholder(dtype=tf.float32, shape=[None, self.__output_size, 1]) # placeholder for the sequence lengths sequence_length = tf.placeholder(dtype=tf.int32, shape=[None]) weight_initializer = tf.truncated_normal_initializer(stddev=random_normal_initializer_stdev) # create the model architecture # RNN with the layer of cells def cell(): if self.__cell_type == "LSTM": cell = tf.nn.rnn_cell.LSTMCell(num_units=int(cell_dimension), use_peepholes=self.__use_peepholes, initializer=weight_initializer) elif self.__cell_type == "GRU": cell = tf.nn.rnn_cell.GRUCell(num_units=int(cell_dimension), kernel_initializer=weight_initializer) elif self.__cell_type == "RNN": cell = tf.nn.rnn_cell.BasicRNNCell(num_units=int(cell_dimension)) return cell # building the encoder network multi_layered_encoder_cell = tf.nn.rnn_cell.MultiRNNCell( cells=[cell() for _ in range(int(num_hidden_layers))]) with tf.variable_scope('train_encoder_scope') as encoder_train_scope: training_encoder_outputs, training_encoder_state = tf.nn.dynamic_rnn(cell=multi_layered_encoder_cell, inputs=training_input, sequence_length=sequence_length, dtype=tf.float32) with tf.variable_scope(encoder_train_scope, reuse=tf.AUTO_REUSE) as encoder_inference_scope: inference_encoder_outputs, inference_encoder_states = tf.nn.dynamic_rnn(cell=multi_layered_encoder_cell, inputs=testing_input, sequence_length=sequence_length, dtype=tf.float32) # create a tensor array for the indices of the encoder outputs array new_index_array = tf.range(start=0, limit=tf.shape(sequence_length)[0], delta=1) output_array_indices = tf.stack([new_index_array, sequence_length - 1], axis=-1) # building the decoder network for training with tf.variable_scope('dense_layer_train_scope') as dense_layer_train_scope: train_final_timestep_predictions = tf.gather_nd(params=training_encoder_outputs, indices=output_array_indices) # the final projection layer to convert the encoder_outputs to the desired dimension train_prediction_output = tf.layers.dense( inputs=tf.convert_to_tensor(value=train_final_timestep_predictions, dtype=tf.float32), units=self.__output_size, use_bias=self.__use_bias, kernel_initializer=weight_initializer) train_prediction_output = tf.expand_dims(input=train_prediction_output, axis=2) # building the decoder network for inference with tf.variable_scope(dense_layer_train_scope, reuse=tf.AUTO_REUSE) as dense_layer_inference_scope: inference_final_timestep_predictions = tf.gather_nd(params=inference_encoder_outputs, indices=output_array_indices) # the final projection layer to convert the encoder_outputs to the desired dimension inference_prediction_output = tf.layers.dense( inputs=tf.convert_to_tensor(value=inference_final_timestep_predictions, dtype=tf.float32), units=self.__output_size, use_bias=self.__use_bias, kernel_initializer=weight_initializer) inference_prediction_output = tf.expand_dims(input=inference_prediction_output, axis=2) # error that should be minimized in the training process error = self.__l1_loss(train_prediction_output, target) # l2 regularization of the trainable model parameters l2_loss = 0.0 for var in tf.trainable_variables(): l2_loss += tf.nn.l2_loss(var) l2_loss = tf.multiply(tf.cast(l2_regularization, dtype=tf.float64), tf.cast(l2_loss, dtype=tf.float64)) total_loss = tf.cast(error, dtype=tf.float64) + l2_loss # create the optimizer optimizer = optimizer_fn(total_loss) # create the Dataset objects for the training and test data training_dataset = tf.data.TFRecordDataset(filenames = [self.__binary_train_file_path], compression_type = "ZLIB") test_dataset = tf.data.TFRecordDataset([self.__binary_test_file_path], compression_type = "ZLIB") # parse the records tfrecord_reader = TFRecordReader() # preparing the training data # randomly shuffle the time series within the dataset shuffle_seed = tf.placeholder(dtype=tf.int64, shape=[]) # training_dataset = training_dataset.apply(tf.data.experimental.shuffle_and_repeat(buffer_size=training_data_configs.SHUFFLE_BUFFER_SIZE, # count=int(max_epoch_size), seed=shuffle_seed)) training_dataset = training_dataset.repeat(count=int(max_epoch_size)) training_dataset = training_dataset.map(tfrecord_reader.validation_data_parser) # create the batches by padding the datasets to make the variable sequence lengths fixed within the individual batches padded_training_data_batches = training_dataset.padded_batch(batch_size=int(minibatch_size), padded_shapes=([], [tf.Dimension(None), 1], [self.__output_size, 1], [self.__meta_data_size, 1])) # get an iterator to the batches training_data_batch_iterator = padded_training_data_batches.make_initializable_iterator() # access each batch using the iterator next_training_data_batch = training_data_batch_iterator.get_next() # preparing the test data test_dataset = test_dataset.map(tfrecord_reader.test_data_parser) # create a single batch from all the test time series by padding the datasets to make the variable sequence lengths fixed padded_test_input_data = test_dataset.padded_batch(batch_size=int(minibatch_size), padded_shapes=( [], [tf.Dimension(None), 1], [self.__meta_data_size, 1])) # get an iterator to the test input data batch test_input_iterator = padded_test_input_data.make_one_shot_iterator() # access the test input batch using the iterator test_input_data_batch = test_input_iterator.get_next() # setup variable initialization init_op = tf.global_variables_initializer() # define the GPU options gpu_options = tf.GPUOptions(allow_growth=True) with tf.Session( config=tf.ConfigProto(log_device_placement=gpu_configs.log_device_placement, allow_soft_placement=True, gpu_options=gpu_options)) as session: session.run(init_op) for epoch in range(int(max_num_epochs)): print("Epoch->", epoch) session.run(training_data_batch_iterator.initializer, feed_dict={shuffle_seed:epoch}) losses = [] while True: try: next_training_batch_value = session.run(next_training_data_batch, feed_dict={shuffle_seed:epoch}) # model training _, loss_val = session.run([optimizer, total_loss], feed_dict={input: next_training_batch_value[1], target: next_training_batch_value[2], sequence_length: next_training_batch_value[0], }) losses.append(loss_val) except tf.errors.OutOfRangeError: break # applying the model to the test data list_of_forecasts = [] while True: try: # get the batch of test inputs test_input_batch_value = session.run(test_input_data_batch) # shape for the target data target_data_shape = [np.shape(test_input_batch_value[1])[0], self.__output_size, 1] # get the output of the network for the test input data batch test_output = session.run(inference_prediction_output, feed_dict={input: test_input_batch_value[1], target: np.zeros(shape = target_data_shape), sequence_length: test_input_batch_value[0], }) forecasts = test_output list_of_forecasts.extend(forecasts.tolist()) except tf.errors.OutOfRangeError: break return np.squeeze(list_of_forecasts, axis = 2) #the third dimension is squeezed since it is one
11,488
50.986425
146
py
time-series-forecasting-release
time-series-forecasting-release/rnn_architectures/stacking_model/stacking_model_tester.py
import numpy as np import tensorflow as tf from tfrecords_handler.moving_window.tfrecord_reader import TFRecordReader from configs.global_configs import training_data_configs from configs.global_configs import gpu_configs class StackingModelTester: def __init__(self, **kwargs): self.__use_bias = kwargs["use_bias"] self.__use_peepholes = kwargs["use_peepholes"] self.__input_size = kwargs["input_size"] self.__output_size = kwargs["output_size"] self.__binary_train_file_path = kwargs["binary_train_file_path"] self.__binary_test_file_path = kwargs["binary_test_file_path"] self.__seed = kwargs["seed"] self.__cell_type = kwargs["cell_type"] self.__without_stl_decomposition = kwargs["without_stl_decomposition"] # define the metadata size based on the usage of stl decomposition if self.__without_stl_decomposition: self.__meta_data_size = 1 else: self.__meta_data_size = self.__output_size + 1 def __l1_loss(self, z, t): loss = tf.reduce_mean(tf.abs(t - z)) return loss def __l2_loss(selfself, z, t): loss = tf.losses.mean_squared_error(labels=t, predictions=z) return loss # Training the time series def test_model(self, **kwargs): # extract the parameters from the kwargs num_hidden_layers = kwargs['num_hidden_layers'] cell_dimension = kwargs['cell_dimension'] minibatch_size = kwargs['minibatch_size'] max_epoch_size = kwargs['max_epoch_size'] max_num_epochs = kwargs['max_num_epochs'] l2_regularization = kwargs['l2_regularization'] gaussian_noise_stdev = kwargs['gaussian_noise_stdev'] optimizer_fn = kwargs['optimizer_fn'] random_normal_initializer_stdev = kwargs['random_normal_initializer_stdev'] # reset the tensorflow graph tf.reset_default_graph() tf.set_random_seed(self.__seed) # declare the input and output placeholders input = tf.placeholder(dtype=tf.float32, shape=[None, None, self.__input_size]) noise = tf.random_normal(shape=tf.shape(input), mean=0.0, stddev=gaussian_noise_stdev, dtype=tf.float32) training_input = input + noise testing_input = input # output format [batch_size, sequence_length, dimension] true_output = tf.placeholder(dtype=tf.float32, shape=[None, None, self.__output_size]) sequence_lengths = tf.placeholder(dtype=tf.int64, shape=[None]) weight_initializer = tf.truncated_normal_initializer(stddev=random_normal_initializer_stdev) # RNN with the layer of cells def cell(): if self.__cell_type == "LSTM": cell = tf.nn.rnn_cell.LSTMCell(num_units=int(cell_dimension), use_peepholes=self.__use_peepholes, initializer=weight_initializer) elif self.__cell_type == "GRU": cell = tf.nn.rnn_cell.GRUCell(num_units=int(cell_dimension), kernel_initializer=weight_initializer) elif self.__cell_type == "RNN": cell = tf.nn.rnn_cell.BasicRNNCell(num_units=int(cell_dimension)) return cell multi_layered_cell = tf.nn.rnn_cell.MultiRNNCell(cells=[cell() for _ in range(int(num_hidden_layers))]) with tf.variable_scope('train_scope') as train_scope: training_rnn_outputs, training_rnn_states = tf.nn.dynamic_rnn(cell=multi_layered_cell, inputs=training_input, sequence_length=sequence_lengths, dtype=tf.float32) # connect the dense layer to the RNN training_prediction_output = tf.layers.dense( inputs=tf.convert_to_tensor(value=training_rnn_outputs, dtype=tf.float32), units=self.__output_size, use_bias=self.__use_bias, kernel_initializer=weight_initializer, name='dense_layer') with tf.variable_scope(train_scope, reuse=tf.AUTO_REUSE) as inference_scope: inference_rnn_outputs, inference_rnn_states = tf.nn.dynamic_rnn(cell=multi_layered_cell, inputs=testing_input, sequence_length=sequence_lengths, dtype=tf.float32) # connect the dense layer to the RNN inference_prediction_output = tf.layers.dense( inputs=tf.convert_to_tensor(value=inference_rnn_outputs, dtype=tf.float32), units=self.__output_size, use_bias=self.__use_bias, kernel_initializer=weight_initializer, name='dense_layer', reuse=True) # error that should be minimized in the training process error = self.__l1_loss(training_prediction_output, true_output) # l2 regularization of the trainable model parameters l2_loss = 0.0 for var in tf.trainable_variables(): l2_loss += tf.nn.l2_loss(var) l2_loss = tf.multiply(tf.cast(l2_regularization, dtype=tf.float64), tf.cast(l2_loss, dtype=tf.float64)) total_loss = tf.cast(error, dtype=tf.float64) + l2_loss # create the adagrad optimizer optimizer = optimizer_fn(total_loss) # create the Dataset objects for the training and test data training_dataset = tf.data.TFRecordDataset(filenames=[self.__binary_train_file_path], compression_type="ZLIB") test_dataset = tf.data.TFRecordDataset([self.__binary_test_file_path], compression_type="ZLIB") # parse the records tfrecord_reader = TFRecordReader(self.__input_size, self.__output_size, self.__meta_data_size) # prepare the training data into batches # randomly shuffle the time series within the dataset shuffle_seed = tf.placeholder(dtype=tf.int64, shape=[]) # training_dataset = training_dataset.apply( # tf.data.experimental.shuffle_and_repeat(buffer_size=training_data_configs.SHUFFLE_BUFFER_SIZE, # count=int(max_epoch_size), seed=shuffle_seed)) training_dataset = training_dataset.repeat(count=int(max_epoch_size)) training_dataset = training_dataset.map(tfrecord_reader.validation_data_parser) # create the batches by padding the datasets to make the variable sequence lengths fixed within the individual batches padded_training_data_batches = training_dataset.padded_batch(batch_size=int(minibatch_size), padded_shapes=( [], [tf.Dimension(None), self.__input_size], [tf.Dimension(None), self.__output_size], [tf.Dimension(None), self.__meta_data_size])) # get an iterator to the batches training_data_batch_iterator = padded_training_data_batches.make_initializable_iterator() # access each batch using the iterator next_training_data_batch = training_data_batch_iterator.get_next() # preparing the test data test_dataset = test_dataset.map(tfrecord_reader.test_data_parser) # create a single batch from all the test time series by padding the datasets to make the variable sequence lengths fixed padded_test_input_data = test_dataset.padded_batch(batch_size=int(minibatch_size), padded_shapes=([], [tf.Dimension(None), self.__input_size], [tf.Dimension(None), self.__meta_data_size])) # get an iterator to the test input data batch test_input_iterator = padded_test_input_data.make_one_shot_iterator() # access the test input batch using the iterator test_input_data_batch = test_input_iterator.get_next() # setup variable initialization init_op = tf.global_variables_initializer() # define the GPU options gpu_options = tf.GPUOptions(allow_growth=True) with tf.Session( config=tf.ConfigProto(log_device_placement=gpu_configs.log_device_placement, allow_soft_placement=True, gpu_options=gpu_options)) as session: session.run(init_op) for epoch in range(int(max_num_epochs)): print("Epoch->", epoch) session.run(training_data_batch_iterator.initializer, feed_dict={shuffle_seed: epoch}) while True: try: training_data_batch_value = session.run(next_training_data_batch, feed_dict={shuffle_seed: epoch}) session.run(optimizer, feed_dict={input: training_data_batch_value[1], true_output: training_data_batch_value[2], sequence_lengths: training_data_batch_value[0]}) except tf.errors.OutOfRangeError: break # applying the model to the test data list_of_forecasts = [] while True: try: # get the batch of test inputs test_input_batch_value = session.run(test_input_data_batch) # get the output of the network for the test input data batch test_output = session.run(inference_prediction_output, feed_dict={input: test_input_batch_value[1], sequence_lengths: test_input_batch_value[0]}) last_output_index = test_input_batch_value[0] - 1 array_first_dimension = np.array(range(0, test_input_batch_value[0].shape[0])) forecasts = test_output[array_first_dimension, last_output_index] list_of_forecasts.extend(forecasts.tolist()) except tf.errors.OutOfRangeError: break session.close() return list_of_forecasts
10,712
49.533019
129
py
time-series-forecasting-release
time-series-forecasting-release/rnn_architectures/stacking_model/stacking_model_trainer.py
import numpy as np import tensorflow as tf from tfrecords_handler.moving_window.tfrecord_reader import TFRecordReader from configs.global_configs import model_training_configs from configs.global_configs import training_data_configs from configs.global_configs import gpu_configs class StackingModelTrainer: def __init__(self, **kwargs): self.__use_bias = kwargs["use_bias"] self.__use_peepholes = kwargs["use_peepholes"] self.__input_size = kwargs["input_size"] self.__output_size = kwargs["output_size"] self.__binary_train_file_path = kwargs["binary_train_file_path"] self.__binary_validation_file_path = kwargs["binary_validation_file_path"] self.__contain_zero_values = kwargs["contain_zero_values"] self.__address_near_zero_instability = kwargs["address_near_zero_instability"] self.__integer_conversion = kwargs["integer_conversion"] self.__seed = kwargs["seed"] self.__cell_type = kwargs["cell_type"] self.__without_stl_decomposition = kwargs["without_stl_decomposition"] # define the metadata size based on the usage of stl decomposition if self.__without_stl_decomposition: self.__meta_data_size = 1 else: self.__meta_data_size = self.__output_size + 1 def __l1_loss(self, z, t): loss = tf.reduce_mean(tf.abs(t - z)) return loss def __l2_loss(self, z, t): loss = tf.losses.mean_squared_error(labels=t, predictions=z) return loss # Training the time series def train_model(self, **kwargs): # extract the parameters from the kwargs num_hidden_layers = kwargs['num_hidden_layers'] cell_dimension = kwargs['cell_dimension'] minibatch_size = kwargs['minibatch_size'] max_epoch_size = kwargs['max_epoch_size'] max_num_epochs = kwargs['max_num_epochs'] l2_regularization = kwargs['l2_regularization'] gaussian_noise_stdev = kwargs['gaussian_noise_stdev'] optimizer_fn = kwargs['optimizer_fn'] random_normal_initializer_stdev = kwargs['random_normal_initializer_stdev'] tf.reset_default_graph() tf.set_random_seed(self.__seed) # declare the input and output placeholders # input format [batch_size, sequence_length, dimension] input = tf.placeholder(dtype=tf.float32, shape=[None, None, self.__input_size]) noise = tf.random_normal(shape=tf.shape(input), mean=0.0, stddev=gaussian_noise_stdev, dtype=tf.float32) training_input = input + noise validation_input = input # output format [batch_size, sequence_length, dimension] true_output = tf.placeholder(dtype=tf.float32, shape=[None, None, self.__output_size]) sequence_lengths = tf.placeholder(dtype=tf.int64, shape=[None]) weight_initializer = tf.truncated_normal_initializer(stddev=random_normal_initializer_stdev) # RNN with the layer of cells def cell(): if self.__cell_type == "LSTM": cell = tf.nn.rnn_cell.LSTMCell(num_units=int(cell_dimension), use_peepholes=self.__use_peepholes, initializer=weight_initializer) elif self.__cell_type == "GRU": cell = tf.nn.rnn_cell.GRUCell(num_units=int(cell_dimension), kernel_initializer=weight_initializer) elif self.__cell_type == "RNN": cell = tf.nn.rnn_cell.BasicRNNCell(num_units=int(cell_dimension)) return cell multi_layered_cell = tf.nn.rnn_cell.MultiRNNCell(cells=[cell() for _ in range(int(num_hidden_layers))]) with tf.variable_scope('train_scope') as train_scope: training_rnn_outputs, training_rnn_states = tf.nn.dynamic_rnn(cell=multi_layered_cell, inputs=training_input, sequence_length=sequence_lengths, dtype=tf.float32) # connect the dense layer to the RNN training_prediction_output = tf.layers.dense( inputs=tf.convert_to_tensor(value=training_rnn_outputs, dtype=tf.float32), units=self.__output_size, use_bias=self.__use_bias, kernel_initializer=weight_initializer, name='dense_layer') with tf.variable_scope(train_scope, reuse=tf.AUTO_REUSE) as inference_scope: inference_rnn_outputs, inference_rnn_states = tf.nn.dynamic_rnn(cell=multi_layered_cell, inputs=validation_input, sequence_length=sequence_lengths, dtype=tf.float32) # connect the dense layer to the RNN inference_prediction_output = tf.layers.dense( inputs=tf.convert_to_tensor(value=inference_rnn_outputs, dtype=tf.float32), units=self.__output_size, use_bias=self.__use_bias, kernel_initializer=weight_initializer, name='dense_layer', reuse=True) error = self.__l1_loss(training_prediction_output, true_output) # l2 regularization of the trainable model parameters l2_loss = 0.0 for var in tf.trainable_variables(): l2_loss += tf.nn.l2_loss(var) l2_loss = tf.multiply(tf.cast(l2_regularization, dtype=tf.float64), tf.cast(l2_loss, dtype=tf.float64)) total_loss = tf.cast(error, dtype=tf.float64) + l2_loss # create the adagrad optimizer optimizer = optimizer_fn(total_loss) # create the training and validation datasets from the tfrecord files training_dataset = tf.data.TFRecordDataset(filenames=[self.__binary_train_file_path], compression_type="ZLIB") validation_dataset = tf.data.TFRecordDataset(filenames=[self.__binary_validation_file_path], compression_type="ZLIB") # parse the records tfrecord_reader = TFRecordReader(self.__input_size, self.__output_size, self.__meta_data_size) # define the expected shapes of data after padding train_padded_shapes = ([], [tf.Dimension(None), self.__input_size], [tf.Dimension(None), self.__output_size]) validation_padded_shapes = ( [], [tf.Dimension(None), self.__input_size], [tf.Dimension(None), self.__output_size], [tf.Dimension(None), self.__meta_data_size]) # prepare the training data into batches # randomly shuffle the time series within the dataset and repeat for the value of the epoch size shuffle_seed = tf.placeholder(dtype=tf.int64, shape=[]) # training_dataset = training_dataset.apply( # tf.data.experimental.shuffle_and_repeat(buffer_size=training_data_configs.SHUFFLE_BUFFER_SIZE, # count=int(max_epoch_size), seed=shuffle_seed)) training_dataset = training_dataset.repeat(count=int(max_epoch_size)) training_dataset = training_dataset.map(tfrecord_reader.train_data_parser) padded_training_data_batches = training_dataset.padded_batch(batch_size=int(minibatch_size), padded_shapes=train_padded_shapes) training_data_batch_iterator = padded_training_data_batches.make_initializable_iterator() next_training_data_batch = training_data_batch_iterator.get_next() # prepare the validation data into batches validation_dataset = validation_dataset.map(tfrecord_reader.validation_data_parser) # create a single batch from all the validation time series by padding the datasets to make the variable sequence lengths fixed padded_validation_dataset = validation_dataset.padded_batch(batch_size=int(minibatch_size), padded_shapes=validation_padded_shapes) # get an iterator to the validation data validation_data_iterator = padded_validation_dataset.make_initializable_iterator() # access the validation data using the iterator next_validation_data_batch = validation_data_iterator.get_next() # setup variable initialization init_op = tf.global_variables_initializer() # define the GPU options gpu_options = tf.GPUOptions(allow_growth=True) with tf.Session( config=tf.ConfigProto(log_device_placement=gpu_configs.log_device_placement, allow_soft_placement=True, gpu_options=gpu_options)) as session: session.run(init_op) smape_final = 0.0 smape_list = [] for epoch in range(int(max_num_epochs)): print("Epoch->", epoch) session.run(training_data_batch_iterator.initializer, feed_dict={ shuffle_seed: epoch}) # initialize the iterator to the beginning of the training dataset while True: try: training_data_batch_value = session.run(next_training_data_batch, feed_dict={shuffle_seed: epoch}) _, total_loss_value = session.run([optimizer, total_loss], feed_dict={training_input: training_data_batch_value[1], true_output: training_data_batch_value[2], sequence_lengths: training_data_batch_value[0]}) except tf.errors.OutOfRangeError: break session.run( validation_data_iterator.initializer) # initialize the iterator to the beginning of the training dataset while True: try: # get the batch of validation inputs validation_data_batch_value = session.run(next_validation_data_batch) # get the output of the network for the validation input data batch validation_output = session.run(inference_prediction_output, feed_dict={input: validation_data_batch_value[1], sequence_lengths: validation_data_batch_value[0] }) # calculate the smape for the validation data using vectorization # convert the data to remove the preprocessing last_indices = validation_data_batch_value[0] - 1 array_first_dimension = np.array(range(0, validation_data_batch_value[0].shape[0])) true_seasonality_values = validation_data_batch_value[3][array_first_dimension, last_indices, 1:] level_values = validation_data_batch_value[3][array_first_dimension, last_indices, 0] lambda_val = -0.7 last_validation_outputs = validation_output[array_first_dimension, last_indices] actual_values = validation_data_batch_value[2][array_first_dimension, last_indices, :] if self.__without_stl_decomposition: converted_validation_output = np.exp(last_validation_outputs) converted_actual_values = np.exp(actual_values) else: converted_validation_output = np.exp( true_seasonality_values + level_values[:, np.newaxis] + last_validation_outputs) converted_actual_values = np.exp(true_seasonality_values + level_values[:, np.newaxis] + actual_values) if self.__contain_zero_values: # to compensate for 0 values in data converted_validation_output = converted_validation_output - 1 converted_actual_values = converted_actual_values - 1 if self.__without_stl_decomposition: converted_validation_output = converted_validation_output * level_values[:, np.newaxis] converted_actual_values = converted_actual_values * level_values[:, np.newaxis] if self.__integer_conversion: converted_validation_output = np.round(converted_validation_output) converted_actual_values = np.round(converted_actual_values) converted_validation_output[converted_validation_output < 0] = 0 converted_actual_values[converted_actual_values < 0] = 0 if self.__address_near_zero_instability: # calculate the smape epsilon = 0.1 sum = np.maximum(np.abs(converted_validation_output) + np.abs(converted_actual_values) + epsilon, 0.5 + epsilon) smape_values = (np.abs(converted_validation_output - converted_actual_values) / sum) * 2 smape_values_per_series = np.mean(smape_values, axis=1) smape_list.extend(smape_values_per_series) else: # calculate the smape smape_values = (np.abs(converted_validation_output - converted_actual_values) / (np.abs(converted_validation_output) + np.abs(converted_actual_values))) * 2 smape_values_per_series = np.mean(smape_values, axis=1) smape_list.extend(smape_values_per_series) except tf.errors.OutOfRangeError: break smape_final = np.mean(smape_list) print("SMAPE value: {}".format(smape_final)) session.close() return float(smape_final), smape_list
14,326
51.479853
136
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noisier2Noise/models/unet_fastMRI.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ import torch import math from torch import nn from torch.nn import functional as F class unet_fastMRI(nn.Module): """ PyTorch implementation of a U-Net model. O. Ronneberger, P. Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015. """ def __init__( self, in_chans: int, chans: int = 32, num_pool_layers: int = 4, drop_prob: float = 0.0, residual_connection: bool = True, ): """ Args: in_chans: Number of channels in the input to the U-Net model. out_chans: Number of channels in the output to the U-Net model. chans: Number of output channels of the first convolution layer. num_pool_layers: Number of down-sampling and up-sampling layers. drop_prob: Dropout probability. residual_connection: Network outputs the residual between input and output. """ super().__init__() self.in_chans = in_chans self.out_chans = in_chans self.chans = chans self.num_pool_layers = num_pool_layers self.drop_prob = drop_prob self.residual_connection = residual_connection self.down_sample_layers = nn.ModuleList([ConvBlock(in_chans, chans, drop_prob)]) #first conv block followed by downsampling ch = chans for _ in range(num_pool_layers - 1): self.down_sample_layers.append(ConvBlock(ch, ch * 2, drop_prob)) #conv blocks that are followed by downsampling ch *= 2 self.conv = ConvBlock(ch, ch * 2, drop_prob) self.up_conv = nn.ModuleList() self.up_transpose_conv = nn.ModuleList() for _ in range(num_pool_layers - 1): self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch)) #transposed conv blocks self.up_conv.append(ConvBlock(ch * 2, ch, drop_prob)) ch //= 2 self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch)) #last transposed conv block self.up_conv.append( nn.Sequential(# ConvBlock(ch * 2, ch, drop_prob), nn.Conv2d(ch, self.out_chans, kernel_size=1, stride=1),# )# ) def add_args(parser): """Add model-specific arguments to the parser.""" parser.add_argument("--bias", default=True, help="use residual bias") parser.add_argument("--residual", default=True, help="use residual connection") parser.add_argument("--in-chans", default=3, help="Either color (3) or grey (1)") parser.add_argument("--chans", default=32, help="Number of channels in outer most layer") parser.add_argument("--num-pool-layers", default=3, help="Number of layers per down- and up-sampling path.") parser.add_argument("--no-pooling", default=False, help="No downsampling. Use the no_unet_fastMRI module.") return parser def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H, W)`. """ stack = [] output = image.detach().clone() # apply down-sampling layers for layer in self.down_sample_layers: output = layer(output) stack.append(output) output = F.avg_pool2d(output, kernel_size=2, stride=2, padding=0) output = self.conv(output) # apply up-sampling layers for transpose_conv, conv in zip(self.up_transpose_conv, self.up_conv): downsample_layer = stack.pop() output = transpose_conv(output) # reflect pad on the right/botton if needed to handle odd input dimensions padding = [0, 0, 0, 0] if output.shape[-1] != downsample_layer.shape[-1]: padding[1] = 1 # padding right if output.shape[-2] != downsample_layer.shape[-2]: padding[3] = 1 # padding bottom if torch.sum(torch.tensor(padding)) != 0: output = F.pad(output, padding, "reflect") output = torch.cat([output, downsample_layer], dim=1) output = conv(output) if self.residual_connection: output = image - output; return output class ConvBlock(nn.Module): """ A Convolutional Block that consists of two convolution layers each followed by instance normalization, LeakyReLU activation and dropout. """ def __init__(self, in_chans: int, out_chans: int, drop_prob: float): """ Args: in_chans: Number of channels in the input. out_chans: Number of channels in the output. drop_prob: Dropout probability. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.drop_prob = drop_prob self.layers = nn.Sequential( nn.Conv2d(in_chans, out_chans, kernel_size=3, padding=1, bias=False), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Dropout2d(drop_prob), nn.Conv2d(out_chans, out_chans, kernel_size=3, padding=1, bias=False), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Dropout2d(drop_prob), ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H, W)`. """ return self.layers(image) class TransposeConvBlock(nn.Module): """ A Transpose Convolutional Block that consists of one convolution transpose layers followed by instance normalization and LeakyReLU activation. """ def __init__(self, in_chans: int, out_chans: int): """ Args: in_chans: Number of channels in the input. out_chans: Number of channels in the output. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.layers = nn.Sequential( nn.ConvTranspose2d( in_chans, out_chans, kernel_size=2, stride=2, bias=False ), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H*2, W*2)`. """ return self.layers(image)
7,138
34.517413
131
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noisier2Noise/models/__init__.py
from .unet_fastMRI import *
29
9
27
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noisier2Noise/utils/progress_bar.py
from collections import OrderedDict from numbers import Number from tqdm import tqdm from .meters import AverageMeter, RunningAverageMeter, TimeMeter class ProgressBar: def __init__(self, iterable, epoch=None, prefix=None, quiet=False): self.epoch = epoch self.quiet = quiet self.prefix = prefix + ' | ' if prefix is not None else '' if epoch is not None: self.prefix += f"epoch {epoch:02d}" self.iterable = iterable if self.quiet else tqdm(iterable, self.prefix, leave=False) def __iter__(self): return iter(self.iterable) def log(self, stats, verbose=False): if not self.quiet: self.iterable.set_postfix(self.format_stats(stats, verbose), refresh=True) def format_stats(self, stats, verbose=False): postfix = OrderedDict(stats) for key, value in postfix.items(): if isinstance(value, Number): fmt = "{:.3f}" if value > 0.001 else "{:.1e}" postfix[key] = fmt.format(value) elif isinstance(value, AverageMeter) or isinstance(value, RunningAverageMeter): if verbose: postfix[key] = f"{value.avg:.5f} ({value.val:.5f})" else: postfix[key] = f"{value.avg:.5f}" elif isinstance(value, TimeMeter): postfix[key] = f"{value.elapsed_time:.1f}s" elif not isinstance(postfix[key], str): postfix[key] = str(value) return postfix def print(self, stats, verbose=False): postfix = " | ".join(key + " " + value.strip() for key, value in self.format_stats(stats, verbose).items()) return f"{self.prefix + ' | ' if self.epoch is not None else ''}{postfix}"
1,789
37.913043
115
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noisier2Noise/utils/utils_image.py
import os import math import random import numpy as np import torch import cv2 from torchvision.utils import make_grid def modcrop(img, scale): # img_in: BCHW or CHW or HW #img = np.copy(img_in) if img.ndim == 2: H, W = img.shape H_r, W_r = H % scale, W % scale img = img[:H - H_r, :W - W_r] elif img.ndim == 3: C, H, W = img.shape H_r, W_r = H % scale, W % scale img = img[:, :H - H_r, :W - W_r] elif img.ndim == 4: B, C, H, W = img.shape H_r, W_r = H % scale, W % scale img = img[:, :, :H - H_r, :W - W_r] else: raise ValueError('Wrong img ndim: [{:d}].'.format(img.ndim)) return img def modcrop_pil(image, modulo): w = image.width - image.width % modulo h = image.height - image.height % modulo return image.crop((0, 0, w, h)) def crop_center(pil_img, crop_width, crop_height): # Perform center crop on a PIL image img_width, img_height = pil_img.size return pil_img.crop(((img_width - crop_width) // 2, (img_height - crop_height) // 2, (img_width + crop_width) // 2, (img_height + crop_height) // 2)) ''' # -------------------------------------------- # matlab's bicubic imresize (numpy and torch) [0, 1] # -------------------------------------------- # from https://github.com/cszn/KAIR/blob/master/utils/utils_image.py ''' # matlab 'imresize' function, now only support 'bicubic' def cubic(x): absx = torch.abs(x) absx2 = absx**2 absx3 = absx**3 return (1.5*absx3 - 2.5*absx2 + 1) * ((absx <= 1).type_as(absx)) + \ (-0.5*absx3 + 2.5*absx2 - 4*absx + 2) * (((absx > 1)*(absx <= 2)).type_as(absx)) def calculate_weights_indices(in_length, out_length, scale, kernel, kernel_width, antialiasing): if (scale < 1) and (antialiasing): # Use a modified kernel to simultaneously interpolate and antialias- larger kernel width kernel_width = kernel_width / scale # Output-space coordinates x = torch.linspace(1, out_length, out_length) # Input-space coordinates. Calculate the inverse mapping such that 0.5 # in output space maps to 0.5 in input space, and 0.5+scale in output # space maps to 1.5 in input space. u = x / scale + 0.5 * (1 - 1 / scale) # What is the left-most pixel that can be involved in the computation? left = torch.floor(u - kernel_width / 2) # What is the maximum number of pixels that can be involved in the # computation? Note: it's OK to use an extra pixel here; if the # corresponding weights are all zero, it will be eliminated at the end # of this function. P = math.ceil(kernel_width) + 2 # The indices of the input pixels involved in computing the k-th output # pixel are in row k of the indices matrix. indices = left.view(out_length, 1).expand(out_length, P) + torch.linspace(0, P - 1, P).view( 1, P).expand(out_length, P) # The weights used to compute the k-th output pixel are in row k of the # weights matrix. distance_to_center = u.view(out_length, 1).expand(out_length, P) - indices # apply cubic kernel if (scale < 1) and (antialiasing): weights = scale * cubic(distance_to_center * scale) else: weights = cubic(distance_to_center) # Normalize the weights matrix so that each row sums to 1. weights_sum = torch.sum(weights, 1).view(out_length, 1) weights = weights / weights_sum.expand(out_length, P) # If a column in weights is all zero, get rid of it. only consider the first and last column. weights_zero_tmp = torch.sum((weights == 0), 0) if not math.isclose(weights_zero_tmp[0], 0, rel_tol=1e-6): indices = indices.narrow(1, 1, P - 2) weights = weights.narrow(1, 1, P - 2) if not math.isclose(weights_zero_tmp[-1], 0, rel_tol=1e-6): indices = indices.narrow(1, 0, P - 2) weights = weights.narrow(1, 0, P - 2) weights = weights.contiguous() indices = indices.contiguous() sym_len_s = -indices.min() + 1 sym_len_e = indices.max() - in_length indices = indices + sym_len_s - 1 return weights, indices, int(sym_len_s), int(sym_len_e) # -------------------------------------------- # imresize for tensor image [0, 1] # -------------------------------------------- def imresize(img, scale, antialiasing=True): # Now the scale should be the same for H and W # input: img: pytorch tensor, CHW or HW [0,1] # output: CHW or HW [0,1] w/o round need_squeeze = True if img.dim() == 2 else False if need_squeeze: img.unsqueeze_(0) in_C, in_H, in_W = img.size() out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale) kernel_width = 4 kernel = 'cubic' # Return the desired dimension order for performing the resize. The # strategy is to perform the resize first along the dimension with the # smallest scale factor. # Now we do not support this. # get weights and indices weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices( in_H, out_H, scale, kernel, kernel_width, antialiasing) weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices( in_W, out_W, scale, kernel, kernel_width, antialiasing) # process H dimension # symmetric copying img_aug = torch.FloatTensor(in_C, in_H + sym_len_Hs + sym_len_He, in_W) img_aug.narrow(1, sym_len_Hs, in_H).copy_(img) sym_patch = img[:, :sym_len_Hs, :] inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(1, inv_idx) img_aug.narrow(1, 0, sym_len_Hs).copy_(sym_patch_inv) sym_patch = img[:, -sym_len_He:, :] inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(1, inv_idx) img_aug.narrow(1, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv) out_1 = torch.FloatTensor(in_C, out_H, in_W) kernel_width = weights_H.size(1) for i in range(out_H): idx = int(indices_H[i][0]) for j in range(out_C): out_1[j, i, :] = img_aug[j, idx:idx + kernel_width, :].transpose(0, 1).mv(weights_H[i]) # process W dimension # symmetric copying out_1_aug = torch.FloatTensor(in_C, out_H, in_W + sym_len_Ws + sym_len_We) out_1_aug.narrow(2, sym_len_Ws, in_W).copy_(out_1) sym_patch = out_1[:, :, :sym_len_Ws] inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(2, inv_idx) out_1_aug.narrow(2, 0, sym_len_Ws).copy_(sym_patch_inv) sym_patch = out_1[:, :, -sym_len_We:] inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(2, inv_idx) out_1_aug.narrow(2, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv) out_2 = torch.FloatTensor(in_C, out_H, out_W) kernel_width = weights_W.size(1) for i in range(out_W): idx = int(indices_W[i][0]) for j in range(out_C): out_2[j, :, i] = out_1_aug[j, :, idx:idx + kernel_width].mv(weights_W[i]) if need_squeeze: out_2.squeeze_() return out_2 # -------------------------------------------- # imresize for numpy image [0, 1] # -------------------------------------------- def imresize_np(img, scale, antialiasing=True): # Now the scale should be the same for H and W # input: img: Numpy, HWC or HW [0,1] # output: HWC or HW [0,1] w/o round img = torch.from_numpy(img) need_squeeze = True if img.dim() == 2 else False if need_squeeze: img.unsqueeze_(2) in_H, in_W, in_C = img.size() out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale) kernel_width = 4 kernel = 'cubic' # Return the desired dimension order for performing the resize. The # strategy is to perform the resize first along the dimension with the # smallest scale factor. # Now we do not support this. # get weights and indices weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices( in_H, out_H, scale, kernel, kernel_width, antialiasing) weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices( in_W, out_W, scale, kernel, kernel_width, antialiasing) # process H dimension # symmetric copying img_aug = torch.FloatTensor(in_H + sym_len_Hs + sym_len_He, in_W, in_C) img_aug.narrow(0, sym_len_Hs, in_H).copy_(img) sym_patch = img[:sym_len_Hs, :, :] inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(0, inv_idx) img_aug.narrow(0, 0, sym_len_Hs).copy_(sym_patch_inv) sym_patch = img[-sym_len_He:, :, :] inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(0, inv_idx) img_aug.narrow(0, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv) out_1 = torch.FloatTensor(out_H, in_W, in_C) kernel_width = weights_H.size(1) for i in range(out_H): idx = int(indices_H[i][0]) for j in range(out_C): out_1[i, :, j] = img_aug[idx:idx + kernel_width, :, j].transpose(0, 1).mv(weights_H[i]) # process W dimension # symmetric copying out_1_aug = torch.FloatTensor(out_H, in_W + sym_len_Ws + sym_len_We, in_C) out_1_aug.narrow(1, sym_len_Ws, in_W).copy_(out_1) sym_patch = out_1[:, :sym_len_Ws, :] inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(1, inv_idx) out_1_aug.narrow(1, 0, sym_len_Ws).copy_(sym_patch_inv) sym_patch = out_1[:, -sym_len_We:, :] inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(1, inv_idx) out_1_aug.narrow(1, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv) out_2 = torch.FloatTensor(out_H, out_W, in_C) kernel_width = weights_W.size(1) for i in range(out_W): idx = int(indices_W[i][0]) for j in range(out_C): out_2[:, i, j] = out_1_aug[:, idx:idx + kernel_width, j].mv(weights_W[i]) if need_squeeze: out_2.squeeze_() return out_2.numpy()
10,256
37.852273
99
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noisier2Noise/utils/train_utils.py
import argparse import os import logging import numpy as np import random import sys import torch from datetime import datetime from torch.serialization import default_restore_location def add_logging_arguments(parser): parser.add_argument("--seed", default=0, type=int, help="random number generator seed") parser.add_argument("--output-dir", default="experiments", help="path to experiment directories") parser.add_argument("--experiment", default=None, help="experiment name to be used with Tensorboard") parser.add_argument("--resume-training", action="store_true", help="whether to resume training") parser.add_argument("--restore-mode", default=None, help="Either 'best' 'last' or '\path\to\checkpoint\dir'") parser.add_argument("--restore-file", default=None, help="filename to load checkpoint") parser.add_argument("--test-mode", default=None, help="Evaluate on which test set.") parser.add_argument("--no-save", action="store_true", help="don't save models or checkpoints") parser.add_argument("--step-checkpoints", action="store_true", help="store all step checkpoints") parser.add_argument("--no-log", action="store_true", help="don't save logs to file or Tensorboard directory") parser.add_argument("--log-interval", type=int, default=100, help="log every N steps") parser.add_argument("--no-visual", action="store_true", help="don't use Tensorboard") parser.add_argument("--visual-interval", type=int, default=100, help="log every N steps") parser.add_argument("--no-progress", action="store_true", help="don't use progress bar") parser.add_argument("--draft", action="store_true", help="save experiment results to draft directory") parser.add_argument("--dry-run", action="store_true", help="no log, no save, no visualization") return parser def init_logging(args): for handler in logging.root.handlers[:]: logging.root.removeHandler(handler) handlers = [logging.StreamHandler()] if not args.no_log and args.log_file is not None: mode = "a" if args.resume_training else "w" handlers.append(logging.FileHandler(args.log_file, mode=mode)) logging.basicConfig(handlers=handlers, format="[%(asctime)s] %(message)s", datefmt="%Y-%m-%d %H:%M:%S", level=logging.INFO) logging.info("Arguments: {}".format(vars(args))) def setup_experiment(args): torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = True torch.manual_seed(args.seed) np.random.seed(args.seed) random.seed(args.seed) if args.dry_run: args.no_save = args.no_log = args.no_visual = True return args.experiment = args.experiment or f"{args.model.replace('_', '-')}" #unet args.experiment = "-".join([args.experiment, 'std'+str(args.noise_std)]) if not args.resume_training: args.experiment = "-".join([args.experiment, datetime.now().strftime("%b-%d-%H:%M:%S")]) args.experiment = "-".join([args.experiment, 'tr'+str(args.train_size)]) args.experiment_dir = os.path.join(args.output_dir, args.experiment) os.makedirs(args.experiment_dir, exist_ok=True) #dir is only created if it not already exists. If it already exists no error is raised if not args.no_save: args.checkpoint_dir = os.path.join(args.experiment_dir, "checkpoints") os.makedirs(args.checkpoint_dir, exist_ok=True) if not args.no_log: args.log_dir = os.path.join(args.experiment_dir, "logs") os.makedirs(args.log_dir, exist_ok=True) args.log_file = os.path.join(args.log_dir, "train.log") def save_checkpoint(args, step, epoch, model, optimizer=None, scheduler=None, score=None, mode="min"): assert mode == "min" or mode == "max" last_step = getattr(save_checkpoint, "last_step", -1) #-1 as default argument that is given if attribute does not exist save_checkpoint.last_step = max(last_step, step) default_score = float("inf") if mode == "min" else float("-inf") best_score = getattr(save_checkpoint, "best_score", default_score) if (score < best_score and mode == "min") or (score > best_score and mode == "max"): save_checkpoint.best_step = step save_checkpoint.best_epoch = epoch save_checkpoint.best_score = score if not args.no_save and step % args.save_interval == 0: os.makedirs(args.checkpoint_dir, exist_ok=True) model = [model] if model is not None and not isinstance(model, list) else model optimizer = [optimizer] if optimizer is not None and not isinstance(optimizer, list) else optimizer scheduler = [scheduler] if scheduler is not None and not isinstance(scheduler, list) else scheduler state_dict = { "step": step, "epoch": epoch, "score": score, "last_step": save_checkpoint.last_step, "best_step": save_checkpoint.best_step, "best_epoch": save_checkpoint.best_epoch, "best_score": getattr(save_checkpoint, "best_score", None), "model": [m.state_dict() for m in model] if model is not None else None, "optimizer": [o.state_dict() for o in optimizer] if optimizer is not None else None, "scheduler": [s.state_dict() for s in scheduler] if scheduler is not None else None, "args": argparse.Namespace(**{k: v for k, v in vars(args).items() if not callable(v)}), } if args.step_checkpoints: torch.save(state_dict, os.path.join(args.checkpoint_dir, "checkpoint{}.pt".format(step))) if (score < best_score and mode == "min") or (score > best_score and mode == "max"): torch.save(state_dict, os.path.join(args.checkpoint_dir, "checkpoint_best.pt")) if step > last_step: torch.save(state_dict, os.path.join(args.checkpoint_dir, "checkpoint_last.pt")) def load_checkpoint(args, model=None, optimizer=None, scheduler=None): if args.restore_file is not None and os.path.isfile(args.restore_file): print('restoring model..') state_dict = torch.load(args.restore_file, map_location=lambda s, l: default_restore_location(s, "cpu")) model = [model] if model is not None and not isinstance(model, list) else model optimizer = [optimizer] if optimizer is not None and not isinstance(optimizer, list) else optimizer scheduler = [scheduler] if scheduler is not None and not isinstance(scheduler, list) else scheduler if "best_score" in state_dict: save_checkpoint.best_score = state_dict["best_score"] save_checkpoint.best_step = state_dict["best_step"] if "last_step" in state_dict: save_checkpoint.last_step = state_dict["last_step"] if model is not None and state_dict.get("model", None) is not None: for m, state in zip(model, state_dict["model"]): m.load_state_dict(state) if optimizer is not None and state_dict.get("optimizer", None) is not None: for o, state in zip(optimizer, state_dict["optimizer"]): o.load_state_dict(state) if scheduler is not None and state_dict.get("scheduler", None) is not None: for s, state in zip(scheduler, state_dict["scheduler"]): milestones = s.milestones state['milestones'] = milestones s.load_state_dict(state) s.milestones = milestones logging.info("Loaded checkpoint {}".format(args.restore_file)) return state_dict
7,573
52.716312
138
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noisier2Noise/utils/main_function_helpers.py
import torch import argparse import os import yaml import pathlib import pickle import logging import sys import time from torch.utils.tensorboard import SummaryWriter import torch.nn.functional as F import torchvision import glob from torch.serialization import default_restore_location from torch.utils.data import DataLoader import matplotlib.pyplot as plt import numpy as np import utils import models from utils.data_helpers.load_datasets_helpers import * from utils.meters import * from utils.progress_bar import * from utils.noise_model import get_noise from utils.metrics import ssim,psnr from utils.util_calculate_psnr_ssim import calculate_psnr,calculate_ssim from utils.test_metrics import * def load_model(args): USE_CUDA = True device = torch.device('cuda') if (torch.cuda.is_available() and USE_CUDA) else torch.device('cpu') checkpoint_path = glob.glob(args.output_dir +'/unet*') if len(checkpoint_path) != 1: raise ValueError("There is either no or more than one model to load") checkpoint_path = pathlib.Path(checkpoint_path[0] + f"/checkpoints/checkpoint_{args.restore_mode}.pt") state_dict = torch.load(checkpoint_path, map_location=lambda s, l: default_restore_location(s, "cpu")) args = argparse.Namespace(**{ **vars(state_dict["args"]), "no_log": True}) model = models.unet_fastMRI( in_chans=args.in_chans, chans = args.chans, num_pool_layers = args.num_pool_layers, drop_prob = 0.0, residual_connection = args.residual, ).to(device) model.load_state_dict(state_dict["model"][0]) model.eval() return model def cli_main_test(args): USE_CUDA = True device = torch.device('cuda') if (torch.cuda.is_available() and USE_CUDA) else torch.device('cpu') model = load_model(args) # evaluate test performance over following noise range noise_std_range = np.linspace(args.test_noise_std_min, args.test_noise_std_max, ((args.test_noise_std_max-args.test_noise_std_min)//args.test_noise_stepsize)+1,dtype=int)/255. metrics_path = os.path.join(args.output_dir, args.test_mode + '_' + str(args.test_noise_std_min)+'-'+str(args.test_noise_std_max)+f'_metrics_{args.restore_mode}.p') metrics_dict = metrics_avg_on_noise_range(model, args, noise_std_range, device = device) pickle.dump( metrics_dict, open(metrics_path, "wb" ) ) def cli_main(args): available_models = glob.glob(f'{args.output_dir}/*') if not args.resume_training and available_models: raise ValueError('There exists already a trained model and resume_training is set False') if args.resume_training: f_restore_file(args) # reset the attributes of the function save_checkpoint mode = "max" default_score = float("inf") if mode == "min" else float("-inf") utils.save_checkpoint.best_score = default_score utils.save_checkpoint.best_step = -1 utils.save_checkpoint.best_epoch = -1 utils.save_checkpoint.last_step = -1 utils.save_checkpoint.current_lr = args.lr device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') # Set the name of the directory for saving results utils.setup_experiment(args) utils.init_logging(args) # Build data loaders, a model and an optimizer model = models.unet_fastMRI( in_chans=args.in_chans, chans = args.chans, num_pool_layers = args.num_pool_layers, drop_prob = 0.0, residual_connection = args.residual, ).to(device) optimizer = torch.optim.Adam(model.parameters(), lr=args.lr) scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau( optimizer, mode='max', factor=args.lr_gamma, patience=args.lr_patience, threshold=args.lr_threshold, threshold_mode='abs', cooldown=0, min_lr=args.lr_min, eps=1e-08, verbose=True ) logging.info(f"Built a model consisting of {sum(p.numel() for p in model.parameters()):,} parameters") trainset = ImagenetSubdataset(args.train_size,args.path_to_ImageNet_train,mode='train',patch_size=args.patch_size,val_crop=args.val_crop) train_loader = DataLoader(trainset, batch_size=args.batch_size, shuffle=True, num_workers=8, pin_memory=True,generator=torch.Generator().manual_seed(args.seed)) valset = ImagenetSubdataset(args.val_size,args.path_to_ImageNet_train,mode='val',patch_size=args.patch_size,val_crop=args.val_crop) val_loader = DataLoader(valset, batch_size=1, shuffle=False, num_workers=4, pin_memory=True,generator=torch.Generator().manual_seed(args.seed)) print(optimizer.param_groups[0]["lr"]) if args.resume_training: state_dict = utils.load_checkpoint(args, model, optimizer, scheduler) global_step = state_dict['last_step'] start_epoch = int(state_dict['last_step']/(len(train_loader)))+1 start_decay = True elif args.no_annealing: global_step = -1 start_epoch = 0 start_decay = True else: global_step = -1 start_epoch = 0 start_decay = False print(optimizer.param_groups[0]["lr"]) args.log_interval = min(len(trainset), 100) # len(train_loader)=log once per epoch args.no_visual = False # True for not logging to tensorboard # Track moving average of loss values train_meters = {"train_loss":RunningAverageMeter(0.98)} valid_meters = {name: AverageMeter() for name in (["valid_psnr", "valid_ssim", "valid_psnr_self_supervised", "valid_ssim_self_supervised"])} # Create tensorflow event file writer = SummaryWriter(log_dir=args.experiment_dir) if not args.no_visual else None break_counter = 0 # store the best val performance from lr-interval before the last lr decay best_val_last = 0 # track the best val performance for the current lr-inerval best_val_current = 0 # count for how many lr intervals there was no improvement and break only if there was no improvement for 2 lr_interval_counter = 0 # if best_val_current at the end of the current lr interval is smaller than best_val_last we perform early stopping for epoch in range(start_epoch, args.num_epochs): start = time.process_time() train_bar = ProgressBar(train_loader, epoch) # At beginning of each epoch reset the train meters for meter in train_meters.values(): meter.reset() for inputs, noise_seed in train_bar: model.train() #Sets the module in training mode. global_step += 1 inputs = inputs.to(device) noise = get_noise(inputs,noise_seed, fix_noise = args.fix_noise, noise_std = args.noise_std/255.) noisier_noise = get_noise(inputs,torch.mul(noise_seed,10),fix_noise = args.fix_noisier, noise_std = args.noise_std_noisier/255.) noisy_inputs = noise + inputs noisier_image = noisy_inputs + noisier_noise outputs = model(noisier_image) loss = F.mse_loss(outputs, noisy_inputs, reduction="sum") / torch.prod(torch.tensor(inputs.size())) #(inputs.size(0) * 2) model.zero_grad() loss.backward() optimizer.step() train_meters["train_loss"].update(loss.item()) train_bar.log(dict(**train_meters, lr=optimizer.param_groups[0]["lr"]), verbose=True) # Add to tensorflow event file: if writer is not None: writer.add_scalar("lr", optimizer.param_groups[0]["lr"], global_step) writer.add_scalar("loss/train", train_meters["train_loss"].avg, global_step) sys.stdout.flush() if epoch % args.valid_interval == 0: model.eval() gen_val = torch.Generator() gen_val = gen_val.manual_seed(10) for meter in valid_meters.values(): meter.reset() valid_bar = ProgressBar(val_loader) for sample, noise_seed in valid_bar: with torch.no_grad(): sample = sample.to(device) # Self-supervised validation with fixed noise noise_self_supervised = get_noise(sample,noise_seed, fix_noise = args.fix_noise, noise_std = args.noise_std/255.) noisier_noise = get_noise(sample,torch.mul(noise_seed,10),fix_noise = args.fix_noisier, noise_std = args.noise_std_noisier/255.) noisy_input_fixed = sample + noise_self_supervised noisier_input = noisy_input_fixed + noisier_noise model_output = model(noisier_input) prediction = 2*model_output - noisier_input valid_psnr_self_supervised = psnr(model_output, noisy_input_fixed) valid_ssim_self_supervised = ssim(model_output, noisy_input_fixed) valid_meters["valid_psnr_self_supervised"].update(valid_psnr_self_supervised.item()) valid_meters["valid_ssim_self_supervised"].update(valid_ssim_self_supervised.item()) # Ground truth validation wit fixed noise # It uses the same input and output as in the self-supervised case since the noise seed is fixed valid_psnr = psnr(prediction, sample) valid_ssim = ssim(prediction, sample) valid_meters["valid_psnr"].update(valid_psnr.item()) valid_meters["valid_ssim"].update(valid_ssim.item()) if writer is not None: # Average is correct valid_meters['valid_psnr'].avg since .val would be just the psnr of last sample in val set. writer.add_scalar("psnr/valid", valid_meters['valid_psnr'].avg, global_step) writer.add_scalar("ssim/valid", valid_meters['valid_ssim'].avg, global_step) writer.add_scalar("psnr_selfsupervised/valid", valid_meters['valid_psnr_self_supervised'].avg, global_step) writer.add_scalar("ssim_selfsupervised/valid", valid_meters["valid_ssim_self_supervised"].avg, global_step) writer.add_scalar("lr", optimizer.param_groups[0]["lr"], global_step) sys.stdout.flush() if args.val_flag == 0: # if we do self-supervised validation val_loss = valid_meters["valid_psnr_self_supervised"].avg else: # if we do supervised validation val_loss = valid_meters["valid_psnr"].avg if utils.save_checkpoint.best_score < val_loss and not start_decay: utils.save_checkpoint(args, global_step, epoch, model, optimizer, score=val_loss, mode="max") current_lr = utils.save_checkpoint.current_lr optimizer.param_groups[0]["lr"] = current_lr*args.lr_beta utils.save_checkpoint.current_lr = current_lr*args.lr_beta annealing_counter = 0 elif not start_decay: annealing_counter += 1 current_lr = utils.save_checkpoint.current_lr if annealing_counter == args.lr_patience_annealing: available_models = glob.glob(f'{args.output_dir}/*') if not available_models: raise ValueError('No file to restore') elif len(available_models)>1: raise ValueError('Too many files to restore from') model_path = os.path.join(available_models[0], "checkpoints/checkpoint_best.pt") state_dict = torch.load(model_path, map_location=lambda s, l: default_restore_location(s, "cpu")) model = [model] if model is not None and not isinstance(model, list) else model for m, state in zip(model, state_dict["model"]): m.load_state_dict(state) model = model[0] optimizer.param_groups[0]["lr"] = current_lr/(args.lr_beta*args.inital_decay_factor) start_decay = True scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau( optimizer, mode='max', factor=args.lr_gamma, patience=args.lr_patience, threshold=args.lr_threshold, threshold_mode='abs', cooldown=0, min_lr=args.lr_min, eps=1e-08, verbose=True ) else: utils.save_checkpoint(args, global_step, epoch, model, optimizer, score=val_loss, mode="max") current_lr = optimizer.param_groups[0]["lr"] if val_loss > best_val_current: best_val_current = val_loss if writer is not None: writer.add_scalar("epoch", epoch, global_step) sys.stdout.flush() if start_decay: current_lr = optimizer.param_groups[0]["lr"] scheduler.step(val_loss) new_lr = optimizer.param_groups[0]["lr"] #At every lr decay check if the model did not improve during the current or the previous lr interval and break if it didn't. if new_lr < current_lr: if best_val_current < best_val_last and lr_interval_counter==1: logging.info('Break training due to convergence of val loss!') break elif best_val_current < best_val_last and lr_interval_counter==0: lr_interval_counter += 1 logging.info('Do not yet break due to convergence of val loss!') else: best_val_last = best_val_current best_val_current = 0 lr_interval_counter = 0 end = time.process_time() - start logging.info(train_bar.print(dict(**train_meters, **valid_meters, lr=current_lr, time=np.round(end/60,3)))) if optimizer.param_groups[0]["lr"] == args.lr_min and start_decay: break_counter += 1 if break_counter == args.break_counter: print('Break training due to minimal learning rate constraint!') break logging.info(f"Done training! Best PSNR {utils.save_checkpoint.best_score:.3f} obtained after step {utils.save_checkpoint.best_step} (epoch {utils.save_checkpoint.best_epoch}).") def get_args(hp,ee,rr): parser = argparse.ArgumentParser(allow_abbrev=False) # Add data arguments parser.add_argument("--train-size", default=None, help="number of examples in training set") parser.add_argument("--val-size", default=40, help="number of examples in validation set") parser.add_argument("--test-size", default=100, help="number of examples in test set") parser.add_argument("--val-crop", default=True, type=bool, help="Crop validation images to train size.") parser.add_argument("--patch-size", default=128, help="size of the center cropped HR image") parser.add_argument("--batch-size", default=128, type=int, help="train batch size") # Add model arguments parser.add_argument("--model", default="unet", help="model architecture") # Add noise arguments parser.add_argument('--noise_std', default = 15, type = float, help = 'noise level') parser.add_argument('--test_noise_std_min', default = 15, type = float, help = 'minimal noise level for testing') parser.add_argument('--test_noise_std_max', default = 15, type = float, help = 'maximal noise level for testing') parser.add_argument('--test_noise_stepsize', default = 5, type = float, help = 'Stepsize between test_noise_std_min and test_noise_std_max') # Add optimization arguments parser.add_argument("--lr", default=1e-3, type=float, help="learning rate") parser.add_argument("--lr-gamma", default=0.5, type=float, help="factor by which to reduce learning rate") parser.add_argument("--lr-beta", default=2, type=float, help="factor by which to increase learning rate") parser.add_argument("--lr-patience", default=5, type=int, help="epochs without improvement before lr decay") parser.add_argument("--no_annealing", default=True, type=bool, help="Use lr annealing or not.") parser.add_argument("--lr-patience-annealing", default=3, type=int, help="epochs without improvement before lr annealing stops") parser.add_argument("--lr-min", default=1e-5, type=float, help="Once we reach this learning rate continue for break_counter many epochs then stop.") parser.add_argument("--lr-threshold", default=0.003, type=float, help="Improvements by less than this threshold are not counted for decay patience.") parser.add_argument("--break-counter", default=9, type=int, help="Once smallest learning rate is reached, continue for so many epochs before stopping.") parser.add_argument("--inital-decay-factor", default=2, type=int, help="After annealing found a lr for which val loss does not improve, go back initial_decay_factor many lrs") parser.add_argument("--num-epochs", default=100, type=int, help="force stop training at specified epoch") parser.add_argument("--valid-interval", default=1, type=int, help="evaluate every N epochs") parser.add_argument("--save-interval", default=1, type=int, help="save a checkpoint every N steps") # Add model arguments parser = models.unet_fastMRI.add_args(parser) parser = utils.add_logging_arguments(parser) #args = parser.parse_args() args, _ = parser.parse_known_args() # Set arguments specific for this experiment dargs = vars(args) for key in hp.keys(): dargs[key] = hp[key][ee] args.seed = int(42 + 10*rr) return args def f_restore_file(args): #available_models = glob.glob(f'{args.output_dir}/{args.experiment}-*') available_models = glob.glob(f'{args.output_dir}/*') if not available_models: raise ValueError('No file to restore') if not args.restore_mode: raise ValueError("Pick restore mode either 'best' 'last' or '\path\to\checkpoint\dir'") if args.restore_mode=='best': mode = "max" best_score = float("inf") if mode == "min" else float("-inf") best_model = None for modelp in available_models: model_path = os.path.join(modelp, "checkpoints/checkpoint_best.pt") if os.path.isfile(model_path): state_dict = torch.load(model_path, map_location=lambda s, l: default_restore_location(s, "cpu")) score = state_dict["best_score"] if (score < best_score and mode == "min") or (score > best_score and mode == "max"): best_score = score best_model = model_path best_modelp = modelp best_step = state_dict["best_step"] best_epoch = state_dict["best_epoch"] args.restore_file = best_model args.experiment_dir = best_modelp #logging.info(f"Prepare to restore best model {best_model} with PSNR {best_score} at step {best_step}, epoch {best_epoch}") elif args.restore_mode=='last': last_step = -1 last_model = None for modelp in available_models: model_path = os.path.join(modelp, "checkpoints/checkpoint_last.pt") if os.path.isfile(model_path): state_dict = torch.load(model_path, map_location=lambda s, l: default_restore_location(s, "cpu")) step = state_dict["last_step"] if step > last_step: last_step = step last_model = model_path last_modelp = modelp score = state_dict["score"] last_epoch = state_dict["epoch"] args.restore_file = last_model args.experiment_dir = last_modelp #logging.info(f"Prepare to restore last model {last_model} with PSNR {score} at step {last_step}, epoch {last_epoch}") else: args.restore_file = args.restore_mode args.experiment_dir = args.restore_mode[:args.restore_mode.find('/checkpoints')] def infer_images(args): USE_CUDA = True device = torch.device('cuda') if (torch.cuda.is_available() and USE_CUDA) else torch.device('cpu') net = load_model(args) # the denoiser seed_dict = { "val":10, "test":20, "cbsd68":30, "urban100":40, "mcmaster18":50, "kodak24":60, "CBSD68":70, } gen = torch.Generator() gen = gen.manual_seed(seed_dict[args.test_mode]) # Load the test images load_path = '../training_set_lists/' if args.test_mode == 'test': files_source = torch.load(load_path+f'ImageNetTest{args.test_size}_filepaths.pt') #files_source.sort() elif args.test_mode == 'val': files_source = torch.load(load_path+f'ImageNetVal{args.val_size}_filepaths.pt') #files_source.sort() else: files_source = torch.load(load_path+f'{args.test_mode}_filepaths.pt') if not os.path.isdir(args.output_dir+'/test_images'): os.mkdir(args.output_dir+'/test_images') counter = 0 transformT = transforms.ToTensor() transformIm = transforms.ToPILImage() for f in files_source: counter = counter + 1 if counter > 3: break # Create noise ISource = torch.unsqueeze(transformT(Image.open(f).convert("RGB")),0).to(device) noise = torch.randn(ISource.shape,generator = gen) * args.noise_std/255. INoisy = noise.to(device) + ISource out = torch.clamp(net(INoisy), 0., 1.).cpu() out = torch.squeeze(out,0) # Get rid of the 1 in dim 0. im = transformIm(out) INoisy = torch.clamp(torch.squeeze(INoisy,0), 0., 1.).cpu() #INoisy = torch.squeeze(INoisy,0).cpu() INoisy = transformIm(INoisy) clean_image = Image.open(f).convert("RGB") im.save(args.output_dir+f'/test_images/im{counter}_denoised_notclamped.png') clean_image.save(args.output_dir+f'/test_images/im{counter}_ground_truth_notclamped.png') INoisy.save(args.output_dir+f'/test_images/im{counter}_noisy_notclamped.png') im.save(args.output_dir+f'/test_images/im{counter}_denoised_notclamped.pdf') clean_image.save(args.output_dir+f'/test_images/im{counter}_ground_truth_notclamped.pdf') INoisy.save(args.output_dir+f'/test_images/im{counter}_noisy_notclamped.pdf')
22,955
45.563895
182
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noisier2Noise/utils/util_calculate_psnr_ssim.py
import cv2 import numpy as np import torch # from https://github.com/JingyunLiang/SwinIR/blob/328dda0f4768772e6d8c5aa3d5aa8e24f1ad903b/utils/util_calculate_psnr_ssim.py#L80 def calculate_psnr(img1, img2, crop_border, input_order='HWC', test_y_channel=False): """Calculate PSNR (Peak Signal-to-Noise Ratio). Ref: https://en.wikipedia.org/wiki/Peak_signal-to-noise_ratio Args: img1 (ndarray): Images with range [0, 255]. img2 (ndarray): Images with range [0, 255]. crop_border (int): Cropped pixels in each edge of an image. These pixels are not involved in the PSNR calculation. input_order (str): Whether the input order is 'HWC' or 'CHW'. Default: 'HWC'. test_y_channel (bool): Test on Y channel of YCbCr. Default: False. Returns: float: psnr result. """ assert img1.shape == img2.shape, (f'Image shapes are differnet: {img1.shape}, {img2.shape}.') if input_order not in ['HWC', 'CHW']: raise ValueError(f'Wrong input_order {input_order}. Supported input_orders are ' '"HWC" and "CHW"') img1 = reorder_image(img1, input_order=input_order) img2 = reorder_image(img2, input_order=input_order) img1 = img1.astype(np.float64) img2 = img2.astype(np.float64) if crop_border != 0: img1 = img1[crop_border:-crop_border, crop_border:-crop_border, ...] img2 = img2[crop_border:-crop_border, crop_border:-crop_border, ...] if test_y_channel: img1 = to_y_channel(img1) img2 = to_y_channel(img2) mse = np.mean((img1 - img2) ** 2) if mse == 0: return float('inf') return 20. * np.log10(255. / np.sqrt(mse)) def _ssim(img1, img2): """Calculate SSIM (structural similarity) for one channel images. It is called by func:`calculate_ssim`. Args: img1 (ndarray): Images with range [0, 255] with order 'HWC'. img2 (ndarray): Images with range [0, 255] with order 'HWC'. Returns: float: ssim result. """ C1 = (0.01 * 255) ** 2 C2 = (0.03 * 255) ** 2 img1 = img1.astype(np.float64) img2 = img2.astype(np.float64) kernel = cv2.getGaussianKernel(11, 1.5) window = np.outer(kernel, kernel.transpose()) mu1 = cv2.filter2D(img1, -1, window)[5:-5, 5:-5] mu2 = cv2.filter2D(img2, -1, window)[5:-5, 5:-5] mu1_sq = mu1 ** 2 mu2_sq = mu2 ** 2 mu1_mu2 = mu1 * mu2 sigma1_sq = cv2.filter2D(img1 ** 2, -1, window)[5:-5, 5:-5] - mu1_sq sigma2_sq = cv2.filter2D(img2 ** 2, -1, window)[5:-5, 5:-5] - mu2_sq sigma12 = cv2.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2 ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) * (sigma1_sq + sigma2_sq + C2)) return ssim_map.mean() def calculate_ssim(img1, img2, crop_border, input_order='HWC', test_y_channel=False): """Calculate SSIM (structural similarity). Ref: Image quality assessment: From error visibility to structural similarity The results are the same as that of the official released MATLAB code in https://ece.uwaterloo.ca/~z70wang/research/ssim/. For three-channel images, SSIM is calculated for each channel and then averaged. Args: img1 (ndarray): Images with range [0, 255]. img2 (ndarray): Images with range [0, 255]. crop_border (int): Cropped pixels in each edge of an image. These pixels are not involved in the SSIM calculation. input_order (str): Whether the input order is 'HWC' or 'CHW'. Default: 'HWC'. test_y_channel (bool): Test on Y channel of YCbCr. Default: False. Returns: float: ssim result. """ assert img1.shape == img2.shape, (f'Image shapes are differnet: {img1.shape}, {img2.shape}.') if input_order not in ['HWC', 'CHW']: raise ValueError(f'Wrong input_order {input_order}. Supported input_orders are ' '"HWC" and "CHW"') img1 = reorder_image(img1, input_order=input_order) img2 = reorder_image(img2, input_order=input_order) img1 = img1.astype(np.float64) img2 = img2.astype(np.float64) if crop_border != 0: img1 = img1[crop_border:-crop_border, crop_border:-crop_border, ...] img2 = img2[crop_border:-crop_border, crop_border:-crop_border, ...] if test_y_channel: img1 = to_y_channel(img1) img2 = to_y_channel(img2) ssims = [] for i in range(img1.shape[2]): ssims.append(_ssim(img1[..., i], img2[..., i])) return np.array(ssims).mean() def reorder_image(img, input_order='HWC'): """Reorder images to 'HWC' order. If the input_order is (h, w), return (h, w, 1); If the input_order is (c, h, w), return (h, w, c); If the input_order is (h, w, c), return as it is. Args: img (ndarray): Input image. input_order (str): Whether the input order is 'HWC' or 'CHW'. If the input image shape is (h, w), input_order will not have effects. Default: 'HWC'. Returns: ndarray: reordered image. """ if input_order not in ['HWC', 'CHW']: raise ValueError(f'Wrong input_order {input_order}. Supported input_orders are ' "'HWC' and 'CHW'") if len(img.shape) == 2: img = img[..., None] if input_order == 'CHW': img = img.transpose(1, 2, 0) return img def to_y_channel(img): """Change to Y channel of YCbCr. Args: img (ndarray): Images with range [0, 255]. Returns: (ndarray): Images with range [0, 255] (float type) without round. """ img = img.astype(np.float32) / 255. if img.ndim == 3 and img.shape[2] == 3: img = bgr2ycbcr(img, y_only=True) img = img[..., None] return img * 255. def _convert_input_type_range(img): """Convert the type and range of the input image. It converts the input image to np.float32 type and range of [0, 1]. It is mainly used for pre-processing the input image in colorspace convertion functions such as rgb2ycbcr and ycbcr2rgb. Args: img (ndarray): The input image. It accepts: 1. np.uint8 type with range [0, 255]; 2. np.float32 type with range [0, 1]. Returns: (ndarray): The converted image with type of np.float32 and range of [0, 1]. """ img_type = img.dtype img = img.astype(np.float32) if img_type == np.float32: pass elif img_type == np.uint8: img /= 255. else: raise TypeError('The img type should be np.float32 or np.uint8, ' f'but got {img_type}') return img def _convert_output_type_range(img, dst_type): """Convert the type and range of the image according to dst_type. It converts the image to desired type and range. If `dst_type` is np.uint8, images will be converted to np.uint8 type with range [0, 255]. If `dst_type` is np.float32, it converts the image to np.float32 type with range [0, 1]. It is mainly used for post-processing images in colorspace convertion functions such as rgb2ycbcr and ycbcr2rgb. Args: img (ndarray): The image to be converted with np.float32 type and range [0, 255]. dst_type (np.uint8 | np.float32): If dst_type is np.uint8, it converts the image to np.uint8 type with range [0, 255]. If dst_type is np.float32, it converts the image to np.float32 type with range [0, 1]. Returns: (ndarray): The converted image with desired type and range. """ if dst_type not in (np.uint8, np.float32): raise TypeError('The dst_type should be np.float32 or np.uint8, ' f'but got {dst_type}') if dst_type == np.uint8: img = img.round() else: img /= 255. return img.astype(dst_type) def bgr2ycbcr(img, y_only=False): """Convert a BGR image to YCbCr image. The bgr version of rgb2ycbcr. It implements the ITU-R BT.601 conversion for standard-definition television. See more details in https://en.wikipedia.org/wiki/YCbCr#ITU-R_BT.601_conversion. It differs from a similar function in cv2.cvtColor: `BGR <-> YCrCb`. In OpenCV, it implements a JPEG conversion. See more details in https://en.wikipedia.org/wiki/YCbCr#JPEG_conversion. Args: img (ndarray): The input image. It accepts: 1. np.uint8 type with range [0, 255]; 2. np.float32 type with range [0, 1]. y_only (bool): Whether to only return Y channel. Default: False. Returns: ndarray: The converted YCbCr image. The output image has the same type and range as input image. """ img_type = img.dtype img = _convert_input_type_range(img) if y_only: out_img = np.dot(img, [24.966, 128.553, 65.481]) + 16.0 else: out_img = np.matmul( img, [[24.966, 112.0, -18.214], [128.553, -74.203, -93.786], [65.481, -37.797, 112.0]]) + [16, 128, 128] out_img = _convert_output_type_range(out_img, img_type) return out_img
9,023
37.564103
129
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noisier2Noise/utils/metrics.py
import numpy as np from skimage.metrics import peak_signal_noise_ratio, structural_similarity def ssim(clean, noisy, normalized=True): """Use skimage.meamsure.compare_ssim to calculate SSIM Args: clean (Tensor): (B, C, H, W) noisy (Tensor): (B, C, H, W) normalized (bool): If True, the range of tensors are [0., 1.] else [0, 255] Returns: SSIM per image: (B, ) """ if len(clean.shape)!=4 or len(noisy.shape)!=4: raise ValueError("ssim expects clean (Tensor): (B, C, H, W) noisy (Tensor): (B, C, H, W)") if normalized: clean = clean.mul(255).clamp(0, 255) noisy = noisy.mul(255).clamp(0, 255) clean = clean.cpu().detach().numpy().astype(np.float32) noisy = noisy.cpu().detach().numpy().astype(np.float32) clean = np.moveaxis(clean,1,-1) noisy = np.moveaxis(noisy,1,-1) return np.array([structural_similarity(c, n, data_range=255, multichannel=True) for c, n in zip(clean, noisy)]).mean() def psnr(clean, noisy, normalized=True): """Use skimage.meamsure.compare_ssim to calculate SSIM Args: clean (Tensor): (B, C, H, W) noisy (Tensor): (B, C, H, W) normalized (bool): If True, the range of tensors are [0., 1.] else [0, 255] Returns: SSIM per image: (B, ) """ if len(clean.shape)!=4 or len(noisy.shape)!=4: raise ValueError("psnr expects clean (Tensor): (B, C, H, W) noisy (Tensor): (B, C, H, W)") if normalized: clean = clean.mul(255).clamp(0, 255) noisy = noisy.mul(255).clamp(0, 255) clean = clean.cpu().detach().numpy().astype(np.float32) noisy = noisy.cpu().detach().numpy().astype(np.float32) return np.array([peak_signal_noise_ratio(c, n, data_range=255) for c, n in zip(clean, noisy)]).mean()
1,811
36.75
122
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noisier2Noise/utils/__init__.py
from .train_utils import * from .main_function_helpers import *
63
31
36
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noisier2Noise/utils/test_metrics.py
import torch import numpy as np import matplotlib.pyplot as plt import glob import os #import cv2 from utils.noise_model import get_noise from utils.metrics import ssim,psnr from utils.util_calculate_psnr_ssim import calculate_psnr,calculate_ssim from skimage import color import PIL.Image as Image import torchvision.transforms as transforms from utils.utils_image import * metrics_key = ['psnr_m', 'psnr_s', 'psnr_delta_m', 'psnr_delta_s', 'ssim_m', 'ssim_s', 'ssim_delta_m', 'ssim_delta_s']; def tensor_to_image(torch_image, low=0.0, high = 1.0, clamp = True): if clamp: torch_image = torch.clamp(torch_image, low, high); return torch_image[0,0].cpu().data.numpy() def normalize(data): return data/255. def convert_dict_to_string(metrics): return_string = ''; for x in metrics.keys(): return_string += x+': '+str(round(metrics[x], 3))+' '; return return_string def get_all_comparison_metrics(denoised, source, noisy = None, scale=None, return_title_string = False, clamp = True): metrics = {}; metrics['psnr'] = np.zeros(len(denoised)) metrics['ssim'] = np.zeros(len(denoised)) if noisy is not None: metrics['psnr_delta'] = np.zeros(len(denoised)) metrics['ssim_delta'] = np.zeros(len(denoised)) if clamp: denoised = torch.clamp(denoised, 0.0, 1.0) metrics['psnr'] = psnr(source, denoised); metrics['ssim'] = ssim(source, denoised); if noisy is not None: metrics['psnr_delta'] = metrics['psnr'] - psnr(source, noisy); metrics['ssim_delta'] = metrics['ssim'] - ssim(source, noisy); if return_title_string: return convert_dict_to_string(metrics) else: return metrics def average_on_folder(args, net, noise_std, verbose=True, device = torch.device('cuda')): #if verbose: #print('Loading data info ...\n') print(f'\n Dataset: {args.test_mode}, Restore mode: {args.restore_mode}') load_path = '../training_set_lists/' seed_dict = { "val":10, "test":20, } gen = torch.Generator() gen = gen.manual_seed(seed_dict[args.test_mode]) gen2 = torch.Generator() gen2 = gen2.manual_seed(seed_dict[args.test_mode]*10) if args.test_mode == 'test': files_source = torch.load(load_path+f'ImageNetTest{args.test_size}_filepaths.pt') #files_source.sort() elif args.test_mode == 'val': files_source = torch.load(load_path+f'ImageNetVal{args.val_size}_filepaths.pt') #files_source.sort() avreage_metrics_key = ['psnr', 'psnr_delta', 'ssim', 'ssim_delta'] avg_metrics = {}; for x in avreage_metrics_key: avg_metrics[x] = []; psnr_list = [] ssim_list = [] for f in files_source: transformT = transforms.ToTensor() ISource = torch.unsqueeze(transformT(Image.open(args.path_to_ImageNet_train + f).convert("RGB")),0).to(device) if args.test_mode == 'val': noise_seed = int(f[f.find('train/')+17:-5].replace('_','')) gen = gen.manual_seed(noise_seed) gen2 = gen2.manual_seed(noise_seed*10) noise = torch.randn(ISource.shape,generator = gen) * args.noise_std/255. noisier_noise = torch.randn(ISource.shape,generator = gen2) * args.noise_std_noisier/255. INoisy = noise.to(device) + ISource INoisier = INoisy + noisier_noise.to(device) out = torch.clamp(((1 + args.alpha**2)*net(INoisier) - INoisier) / (args.alpha ** 2), 0., 1.) ind_metrics = get_all_comparison_metrics(out, ISource, INoisy, return_title_string = False); for x in avreage_metrics_key: avg_metrics[x].append(ind_metrics[x]) if(verbose): print("%s %s" % (f, convert_dict_to_string(ind_metrics))) metrics = {} for x in avreage_metrics_key: metrics[x+'_m'] = np.mean(avg_metrics[x]) metrics[x+'_s'] = np.std(avg_metrics[x]) if verbose: print("\n Average %s" % (convert_dict_to_string(metrics))) #if(not verbose): return metrics def metrics_avg_on_noise_range(net, args, noise_std_array, device = torch.device('cuda')): array_metrics = {} for x in metrics_key: array_metrics[x] = np.zeros(len(noise_std_array)) for j, noise_std in enumerate(noise_std_array): metric_list = average_on_folder(args, net, noise_std = noise_std, verbose=False, device=device); for x in metrics_key: array_metrics[x][j] += metric_list[x] print('noise: ', int(noise_std*255), ' ', x, ': ', str(array_metrics[x][j])) return array_metrics
4,784
31.331081
119
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noisier2Noise/utils/noise_model.py
import torch def get_noise(data, noise_seed, fix_noise, noise_std = float(25)/255.0): if fix_noise: device = torch.device('cuda') gen = torch.Generator(device=device) batch_size = data.size(dim=0) tensor_dim = list(data.size())[1:] for i in range(0,batch_size): gen = gen.manual_seed(noise_seed[i].item()) noise = torch.randn(tensor_dim,generator = gen, device=device) * noise_std noise = torch.unsqueeze(noise,0) if i == 0: noise_tensor = noise else: noise_tensor = torch.cat((noise_tensor, noise),0) noise = noise_tensor #noise = torch.randn(data.shape,generator = gen, device=device) * noise_std else: noise = torch.randn_like(data) noise.data = noise.data * noise_std return noise
880
31.62963
87
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noisier2Noise/utils/meters.py
import time import torch class AverageMeter(object): def __init__(self): self.reset() def reset(self): self.val = 0 self.avg = 0 self.sum = 0 self.count = 0 def update(self, val, n=1): if isinstance(val, torch.Tensor): val = val.item() self.val = val / n self.sum += val self.count += n self.avg = self.sum / self.count class RunningAverageMeter(object): def __init__(self, momentum=0.98): self.momentum = momentum self.reset() def reset(self): self.val = None self.avg = 0 def update(self, val): if isinstance(val, torch.Tensor): val = val.item() if self.val is None: self.avg = val else: self.avg = self.avg * self.momentum + val * (1 - self.momentum) self.val = val class TimeMeter(object): def __init__(self, init=0): self.reset(init) def reset(self, init=0): self.init = init self.start = time.time() self.n = 0 def update(self, val=1): self.n += val @property def avg(self): return self.n / self.elapsed_time @property def elapsed_time(self): return self.init + (time.time() - self.start)
1,321
20.322581
75
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noisier2Noise/utils/data_helpers/load_datasets_helpers.py
import os import os.path import numpy as np import h5py import torch import torchvision.transforms as transforms import PIL.Image as Image from utils.utils_image import * class ImagenetSubdataset(torch.utils.data.Dataset): def __init__(self, size, path_to_ImageNet_train, mode='train', patch_size='128', val_crop=True): super().__init__() load_path = '../training_set_lists/' self.path_to_ImageNet_train = path_to_ImageNet_train if mode=='train': self.files = torch.load(load_path+f'trsize{size}_filepaths.pt') self.transform = transforms.Compose([ transforms.CenterCrop(patch_size), transforms.ToTensor(), ]) elif mode=='val': self.files = torch.load(load_path+f'ImageNetVal{size}_filepaths.pt') #print(self.files) if val_crop: self.transform = transforms.Compose([ transforms.CenterCrop(patch_size), transforms.ToTensor(), ]) else: self.transform = transforms.Compose([ transforms.ToTensor(), ]) self.noise_seeds = {} for i, file in enumerate(self.files): key = file[file.find('train/')+16:-5] number = int(file[file.find('train/')+17:-5].replace('_','')) self.noise_seeds[key] = number def __len__(self): return len(self.files) def __getitem__(self, index): file = self.files[index] key = file[file.find('train/')+16:-5] noise_seed = self.noise_seeds[key] image = Image.open(self.path_to_ImageNet_train + self.files[index]).convert("RGB") #ImageNet contains some grayscale images data = self.transform(image) return data, noise_seed
1,949
32.62069
131
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Neighbor2Neighbor/models/unet_fastMRI.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ import torch import math from torch import nn from torch.nn import functional as F class unet_fastMRI(nn.Module): """ PyTorch implementation of a U-Net model. O. Ronneberger, P. Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015. """ def __init__( self, in_chans: int, chans: int = 32, num_pool_layers: int = 4, drop_prob: float = 0.0, residual_connection: bool = True, ): """ Args: in_chans: Number of channels in the input to the U-Net model. out_chans: Number of channels in the output to the U-Net model. chans: Number of output channels of the first convolution layer. num_pool_layers: Number of down-sampling and up-sampling layers. drop_prob: Dropout probability. residual_connection: Network outputs the residual between input and output. """ super().__init__() self.in_chans = in_chans self.out_chans = in_chans self.chans = chans self.num_pool_layers = num_pool_layers self.drop_prob = drop_prob self.residual_connection = residual_connection self.down_sample_layers = nn.ModuleList([ConvBlock(in_chans, chans, drop_prob)]) #first conv block followed by downsampling ch = chans for _ in range(num_pool_layers - 1): self.down_sample_layers.append(ConvBlock(ch, ch * 2, drop_prob)) #conv blocks that are followed by downsampling ch *= 2 self.conv = ConvBlock(ch, ch * 2, drop_prob) self.up_conv = nn.ModuleList() self.up_transpose_conv = nn.ModuleList() for _ in range(num_pool_layers - 1): self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch)) #transposed conv blocks self.up_conv.append(ConvBlock(ch * 2, ch, drop_prob)) ch //= 2 self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch)) #last transposed conv block self.up_conv.append( nn.Sequential(# ConvBlock(ch * 2, ch, drop_prob), nn.Conv2d(ch, self.out_chans, kernel_size=1, stride=1),# )# ) def add_args(parser): """Add model-specific arguments to the parser.""" parser.add_argument("--bias", default=True, help="use residual bias") parser.add_argument("--residual", default=True, help="use residual connection") parser.add_argument("--in-chans", default=3, help="Either color (3) or grey (1)") parser.add_argument("--chans", default=32, help="Number of channels in outer most layer") parser.add_argument("--num-pool-layers", default=3, help="Number of layers per down- and up-sampling path.") parser.add_argument("--no-pooling", default=False, help="No downsampling. Use the no_unet_fastMRI module.") return parser def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H, W)`. """ stack = [] output = image.detach().clone() # apply down-sampling layers for layer in self.down_sample_layers: output = layer(output) stack.append(output) output = F.avg_pool2d(output, kernel_size=2, stride=2, padding=0) output = self.conv(output) # apply up-sampling layers for transpose_conv, conv in zip(self.up_transpose_conv, self.up_conv): downsample_layer = stack.pop() output = transpose_conv(output) # reflect pad on the right/botton if needed to handle odd input dimensions padding = [0, 0, 0, 0] if output.shape[-1] != downsample_layer.shape[-1]: padding[1] = 1 # padding right if output.shape[-2] != downsample_layer.shape[-2]: padding[3] = 1 # padding bottom if torch.sum(torch.tensor(padding)) != 0: output = F.pad(output, padding, "reflect") output = torch.cat([output, downsample_layer], dim=1) output = conv(output) if self.residual_connection: output = image - output; return output class ConvBlock(nn.Module): """ A Convolutional Block that consists of two convolution layers each followed by instance normalization, LeakyReLU activation and dropout. """ def __init__(self, in_chans: int, out_chans: int, drop_prob: float): """ Args: in_chans: Number of channels in the input. out_chans: Number of channels in the output. drop_prob: Dropout probability. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.drop_prob = drop_prob self.layers = nn.Sequential( nn.Conv2d(in_chans, out_chans, kernel_size=3, padding=1, bias=False), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Dropout2d(drop_prob), nn.Conv2d(out_chans, out_chans, kernel_size=3, padding=1, bias=False), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Dropout2d(drop_prob), ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H, W)`. """ return self.layers(image) class TransposeConvBlock(nn.Module): """ A Transpose Convolutional Block that consists of one convolution transpose layers followed by instance normalization and LeakyReLU activation. """ def __init__(self, in_chans: int, out_chans: int): """ Args: in_chans: Number of channels in the input. out_chans: Number of channels in the output. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.layers = nn.Sequential( nn.ConvTranspose2d( in_chans, out_chans, kernel_size=2, stride=2, bias=False ), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H*2, W*2)`. """ return self.layers(image)
7,138
34.517413
131
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Neighbor2Neighbor/models/__init__.py
from .unet_fastMRI import *
28
13.5
27
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Neighbor2Neighbor/utils/progress_bar.py
from collections import OrderedDict from numbers import Number from tqdm import tqdm from .meters import AverageMeter, RunningAverageMeter, TimeMeter class ProgressBar: def __init__(self, iterable, epoch=None, prefix=None, quiet=False): self.epoch = epoch self.quiet = quiet self.prefix = prefix + ' | ' if prefix is not None else '' if epoch is not None: self.prefix += f"epoch {epoch:02d}" self.iterable = iterable if self.quiet else tqdm(iterable, self.prefix, leave=False) def __iter__(self): return iter(self.iterable) def log(self, stats, verbose=False): if not self.quiet: self.iterable.set_postfix(self.format_stats(stats, verbose), refresh=True) def format_stats(self, stats, verbose=False): postfix = OrderedDict(stats) for key, value in postfix.items(): if isinstance(value, Number): fmt = "{:.3f}" if value > 0.001 else "{:.1e}" postfix[key] = fmt.format(value) elif isinstance(value, AverageMeter) or isinstance(value, RunningAverageMeter): if verbose: postfix[key] = f"{value.avg:.5f} ({value.val:.5f})" else: postfix[key] = f"{value.avg:.5f}" elif isinstance(value, TimeMeter): postfix[key] = f"{value.elapsed_time:.1f}s" elif not isinstance(postfix[key], str): postfix[key] = str(value) return postfix def print(self, stats, verbose=False): postfix = " | ".join(key + " " + value.strip() for key, value in self.format_stats(stats, verbose).items()) return f"{self.prefix + ' | ' if self.epoch is not None else ''}{postfix}"
1,789
37.913043
115
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Neighbor2Neighbor/utils/utils_image.py
import os import math import random import numpy as np import torch import cv2 from torchvision.utils import make_grid def modcrop(img, scale): # img_in: BCHW or CHW or HW #img = np.copy(img_in) if img.ndim == 2: H, W = img.shape H_r, W_r = H % scale, W % scale img = img[:H - H_r, :W - W_r] elif img.ndim == 3: C, H, W = img.shape H_r, W_r = H % scale, W % scale img = img[:, :H - H_r, :W - W_r] elif img.ndim == 4: B, C, H, W = img.shape H_r, W_r = H % scale, W % scale img = img[:, :, :H - H_r, :W - W_r] else: raise ValueError('Wrong img ndim: [{:d}].'.format(img.ndim)) return img def modcrop_pil(image, modulo): w = image.width - image.width % modulo h = image.height - image.height % modulo return image.crop((0, 0, w, h)) def crop_center(pil_img, crop_width, crop_height): # Perform center crop on a PIL image img_width, img_height = pil_img.size return pil_img.crop(((img_width - crop_width) // 2, (img_height - crop_height) // 2, (img_width + crop_width) // 2, (img_height + crop_height) // 2)) ''' # -------------------------------------------- # matlab's bicubic imresize (numpy and torch) [0, 1] # -------------------------------------------- # from https://github.com/cszn/KAIR/blob/master/utils/utils_image.py ''' # matlab 'imresize' function, now only support 'bicubic' def cubic(x): absx = torch.abs(x) absx2 = absx**2 absx3 = absx**3 return (1.5*absx3 - 2.5*absx2 + 1) * ((absx <= 1).type_as(absx)) + \ (-0.5*absx3 + 2.5*absx2 - 4*absx + 2) * (((absx > 1)*(absx <= 2)).type_as(absx)) def calculate_weights_indices(in_length, out_length, scale, kernel, kernel_width, antialiasing): if (scale < 1) and (antialiasing): # Use a modified kernel to simultaneously interpolate and antialias- larger kernel width kernel_width = kernel_width / scale # Output-space coordinates x = torch.linspace(1, out_length, out_length) # Input-space coordinates. Calculate the inverse mapping such that 0.5 # in output space maps to 0.5 in input space, and 0.5+scale in output # space maps to 1.5 in input space. u = x / scale + 0.5 * (1 - 1 / scale) # What is the left-most pixel that can be involved in the computation? left = torch.floor(u - kernel_width / 2) # What is the maximum number of pixels that can be involved in the # computation? Note: it's OK to use an extra pixel here; if the # corresponding weights are all zero, it will be eliminated at the end # of this function. P = math.ceil(kernel_width) + 2 # The indices of the input pixels involved in computing the k-th output # pixel are in row k of the indices matrix. indices = left.view(out_length, 1).expand(out_length, P) + torch.linspace(0, P - 1, P).view( 1, P).expand(out_length, P) # The weights used to compute the k-th output pixel are in row k of the # weights matrix. distance_to_center = u.view(out_length, 1).expand(out_length, P) - indices # apply cubic kernel if (scale < 1) and (antialiasing): weights = scale * cubic(distance_to_center * scale) else: weights = cubic(distance_to_center) # Normalize the weights matrix so that each row sums to 1. weights_sum = torch.sum(weights, 1).view(out_length, 1) weights = weights / weights_sum.expand(out_length, P) # If a column in weights is all zero, get rid of it. only consider the first and last column. weights_zero_tmp = torch.sum((weights == 0), 0) if not math.isclose(weights_zero_tmp[0], 0, rel_tol=1e-6): indices = indices.narrow(1, 1, P - 2) weights = weights.narrow(1, 1, P - 2) if not math.isclose(weights_zero_tmp[-1], 0, rel_tol=1e-6): indices = indices.narrow(1, 0, P - 2) weights = weights.narrow(1, 0, P - 2) weights = weights.contiguous() indices = indices.contiguous() sym_len_s = -indices.min() + 1 sym_len_e = indices.max() - in_length indices = indices + sym_len_s - 1 return weights, indices, int(sym_len_s), int(sym_len_e) # -------------------------------------------- # imresize for tensor image [0, 1] # -------------------------------------------- def imresize(img, scale, antialiasing=True): # Now the scale should be the same for H and W # input: img: pytorch tensor, CHW or HW [0,1] # output: CHW or HW [0,1] w/o round need_squeeze = True if img.dim() == 2 else False if need_squeeze: img.unsqueeze_(0) in_C, in_H, in_W = img.size() out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale) kernel_width = 4 kernel = 'cubic' # Return the desired dimension order for performing the resize. The # strategy is to perform the resize first along the dimension with the # smallest scale factor. # Now we do not support this. # get weights and indices weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices( in_H, out_H, scale, kernel, kernel_width, antialiasing) weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices( in_W, out_W, scale, kernel, kernel_width, antialiasing) # process H dimension # symmetric copying img_aug = torch.FloatTensor(in_C, in_H + sym_len_Hs + sym_len_He, in_W) img_aug.narrow(1, sym_len_Hs, in_H).copy_(img) sym_patch = img[:, :sym_len_Hs, :] inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(1, inv_idx) img_aug.narrow(1, 0, sym_len_Hs).copy_(sym_patch_inv) sym_patch = img[:, -sym_len_He:, :] inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(1, inv_idx) img_aug.narrow(1, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv) out_1 = torch.FloatTensor(in_C, out_H, in_W) kernel_width = weights_H.size(1) for i in range(out_H): idx = int(indices_H[i][0]) for j in range(out_C): out_1[j, i, :] = img_aug[j, idx:idx + kernel_width, :].transpose(0, 1).mv(weights_H[i]) # process W dimension # symmetric copying out_1_aug = torch.FloatTensor(in_C, out_H, in_W + sym_len_Ws + sym_len_We) out_1_aug.narrow(2, sym_len_Ws, in_W).copy_(out_1) sym_patch = out_1[:, :, :sym_len_Ws] inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(2, inv_idx) out_1_aug.narrow(2, 0, sym_len_Ws).copy_(sym_patch_inv) sym_patch = out_1[:, :, -sym_len_We:] inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(2, inv_idx) out_1_aug.narrow(2, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv) out_2 = torch.FloatTensor(in_C, out_H, out_W) kernel_width = weights_W.size(1) for i in range(out_W): idx = int(indices_W[i][0]) for j in range(out_C): out_2[j, :, i] = out_1_aug[j, :, idx:idx + kernel_width].mv(weights_W[i]) if need_squeeze: out_2.squeeze_() return out_2 # -------------------------------------------- # imresize for numpy image [0, 1] # -------------------------------------------- def imresize_np(img, scale, antialiasing=True): # Now the scale should be the same for H and W # input: img: Numpy, HWC or HW [0,1] # output: HWC or HW [0,1] w/o round img = torch.from_numpy(img) need_squeeze = True if img.dim() == 2 else False if need_squeeze: img.unsqueeze_(2) in_H, in_W, in_C = img.size() out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale) kernel_width = 4 kernel = 'cubic' # Return the desired dimension order for performing the resize. The # strategy is to perform the resize first along the dimension with the # smallest scale factor. # Now we do not support this. # get weights and indices weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices( in_H, out_H, scale, kernel, kernel_width, antialiasing) weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices( in_W, out_W, scale, kernel, kernel_width, antialiasing) # process H dimension # symmetric copying img_aug = torch.FloatTensor(in_H + sym_len_Hs + sym_len_He, in_W, in_C) img_aug.narrow(0, sym_len_Hs, in_H).copy_(img) sym_patch = img[:sym_len_Hs, :, :] inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(0, inv_idx) img_aug.narrow(0, 0, sym_len_Hs).copy_(sym_patch_inv) sym_patch = img[-sym_len_He:, :, :] inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(0, inv_idx) img_aug.narrow(0, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv) out_1 = torch.FloatTensor(out_H, in_W, in_C) kernel_width = weights_H.size(1) for i in range(out_H): idx = int(indices_H[i][0]) for j in range(out_C): out_1[i, :, j] = img_aug[idx:idx + kernel_width, :, j].transpose(0, 1).mv(weights_H[i]) # process W dimension # symmetric copying out_1_aug = torch.FloatTensor(out_H, in_W + sym_len_Ws + sym_len_We, in_C) out_1_aug.narrow(1, sym_len_Ws, in_W).copy_(out_1) sym_patch = out_1[:, :sym_len_Ws, :] inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(1, inv_idx) out_1_aug.narrow(1, 0, sym_len_Ws).copy_(sym_patch_inv) sym_patch = out_1[:, -sym_len_We:, :] inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(1, inv_idx) out_1_aug.narrow(1, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv) out_2 = torch.FloatTensor(out_H, out_W, in_C) kernel_width = weights_W.size(1) for i in range(out_W): idx = int(indices_W[i][0]) for j in range(out_C): out_2[:, i, j] = out_1_aug[:, idx:idx + kernel_width, j].mv(weights_W[i]) if need_squeeze: out_2.squeeze_() return out_2.numpy()
10,256
37.852273
99
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Neighbor2Neighbor/utils/train_utils.py
import argparse import os import logging import numpy as np import random import sys import torch from datetime import datetime from torch.serialization import default_restore_location def add_logging_arguments(parser): parser.add_argument("--seed", default=0, type=int, help="random number generator seed") parser.add_argument("--output-dir", default="experiments", help="path to experiment directories") parser.add_argument("--experiment", default=None, help="experiment name to be used with Tensorboard") parser.add_argument("--resume-training", action="store_true", help="whether to resume training") parser.add_argument("--restore-mode", default=None, help="Either 'best' 'last' or '\path\to\checkpoint\dir'") parser.add_argument("--restore-file", default=None, help="filename to load checkpoint") parser.add_argument("--test-mode", default=None, help="Evaluate on which test set.") parser.add_argument("--no-save", action="store_true", help="don't save models or checkpoints") parser.add_argument("--step-checkpoints", action="store_true", help="store all step checkpoints") parser.add_argument("--no-log", action="store_true", help="don't save logs to file or Tensorboard directory") parser.add_argument("--log-interval", type=int, default=100, help="log every N steps") parser.add_argument("--no-visual", action="store_true", help="don't use Tensorboard") parser.add_argument("--visual-interval", type=int, default=100, help="log every N steps") parser.add_argument("--no-progress", action="store_true", help="don't use progress bar") parser.add_argument("--draft", action="store_true", help="save experiment results to draft directory") parser.add_argument("--dry-run", action="store_true", help="no log, no save, no visualization") return parser def init_logging(args): for handler in logging.root.handlers[:]: logging.root.removeHandler(handler) handlers = [logging.StreamHandler()] if not args.no_log and args.log_file is not None: mode = "a" if args.resume_training else "w" handlers.append(logging.FileHandler(args.log_file, mode=mode)) logging.basicConfig(handlers=handlers, format="[%(asctime)s] %(message)s", datefmt="%Y-%m-%d %H:%M:%S", level=logging.INFO) logging.info("Arguments: {}".format(vars(args))) def setup_experiment(args): torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = True torch.manual_seed(args.seed) np.random.seed(args.seed) random.seed(args.seed) if args.dry_run: args.no_save = args.no_log = args.no_visual = True return args.experiment = args.experiment or f"{args.model.replace('_', '-')}" #unet args.experiment = "-".join([args.experiment, 'std'+str(args.noise_std)]) if not args.resume_training: args.experiment = "-".join([args.experiment, datetime.now().strftime("%b-%d-%H:%M:%S")]) args.experiment = "-".join([args.experiment, 'tr'+str(args.train_size)]) args.experiment_dir = os.path.join(args.output_dir, args.experiment) os.makedirs(args.experiment_dir, exist_ok=True) #dir is only created if it not already exists. If it already exists no error is raised if not args.no_save: args.checkpoint_dir = os.path.join(args.experiment_dir, "checkpoints") os.makedirs(args.checkpoint_dir, exist_ok=True) if not args.no_log: args.log_dir = os.path.join(args.experiment_dir, "logs") os.makedirs(args.log_dir, exist_ok=True) args.log_file = os.path.join(args.log_dir, "train.log") def save_checkpoint(args, step, epoch, model, optimizer=None, scheduler=None, score=None, mode="min"): assert mode == "min" or mode == "max" last_step = getattr(save_checkpoint, "last_step", -1) #-1 as default argument that is given if attribute does not exist save_checkpoint.last_step = max(last_step, step) default_score = float("inf") if mode == "min" else float("-inf") best_score = getattr(save_checkpoint, "best_score", default_score) if (score < best_score and mode == "min") or (score > best_score and mode == "max"): save_checkpoint.best_step = step save_checkpoint.best_epoch = epoch save_checkpoint.best_score = score if not args.no_save and step % args.save_interval == 0: os.makedirs(args.checkpoint_dir, exist_ok=True) model = [model] if model is not None and not isinstance(model, list) else model optimizer = [optimizer] if optimizer is not None and not isinstance(optimizer, list) else optimizer scheduler = [scheduler] if scheduler is not None and not isinstance(scheduler, list) else scheduler state_dict = { "step": step, "epoch": epoch, "score": score, "last_step": save_checkpoint.last_step, "best_step": save_checkpoint.best_step, "best_epoch": save_checkpoint.best_epoch, "best_score": getattr(save_checkpoint, "best_score", None), "model": [m.state_dict() for m in model] if model is not None else None, "optimizer": [o.state_dict() for o in optimizer] if optimizer is not None else None, "scheduler": [s.state_dict() for s in scheduler] if scheduler is not None else None, "args": argparse.Namespace(**{k: v for k, v in vars(args).items() if not callable(v)}), } if args.step_checkpoints: torch.save(state_dict, os.path.join(args.checkpoint_dir, "checkpoint{}.pt".format(step))) if (score < best_score and mode == "min") or (score > best_score and mode == "max"): torch.save(state_dict, os.path.join(args.checkpoint_dir, "checkpoint_best.pt")) if step > last_step: torch.save(state_dict, os.path.join(args.checkpoint_dir, "checkpoint_last.pt")) def load_checkpoint(args, model=None, optimizer=None, scheduler=None): if args.restore_file is not None and os.path.isfile(args.restore_file): print('restoring model..') state_dict = torch.load(args.restore_file, map_location=lambda s, l: default_restore_location(s, "cpu")) model = [model] if model is not None and not isinstance(model, list) else model optimizer = [optimizer] if optimizer is not None and not isinstance(optimizer, list) else optimizer scheduler = [scheduler] if scheduler is not None and not isinstance(scheduler, list) else scheduler if "best_score" in state_dict: save_checkpoint.best_score = state_dict["best_score"] save_checkpoint.best_step = state_dict["best_step"] if "last_step" in state_dict: save_checkpoint.last_step = state_dict["last_step"] if model is not None and state_dict.get("model", None) is not None: for m, state in zip(model, state_dict["model"]): m.load_state_dict(state) if optimizer is not None and state_dict.get("optimizer", None) is not None: for o, state in zip(optimizer, state_dict["optimizer"]): o.load_state_dict(state) if scheduler is not None and state_dict.get("scheduler", None) is not None: for s, state in zip(scheduler, state_dict["scheduler"]): milestones = s.milestones state['milestones'] = milestones s.load_state_dict(state) s.milestones = milestones logging.info("Loaded checkpoint {}".format(args.restore_file)) return state_dict
7,573
52.716312
138
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Neighbor2Neighbor/utils/main_function_helpers.py
import torch import argparse import os import yaml import pathlib import pickle import logging import sys import time from torch.utils.tensorboard import SummaryWriter import torch.nn.functional as F import torchvision import glob from torch.serialization import default_restore_location from torch.utils.data import DataLoader import matplotlib.pyplot as plt import numpy as np from tensorboard.backend.event_processing import event_accumulator import utils import models from utils.data_helpers.load_datasets_helpers import * from utils.meters import * from utils.progress_bar import * from utils.noise_model import get_noise from utils.metrics import ssim,psnr from utils.util_calculate_psnr_ssim import calculate_psnr,calculate_ssim,calculate_psnr_neighbor, calculate_upsnr from utils.test_metrics import * import torchvision.transforms.functional as FF operation_seed_counter = 0 def space_to_depth(x, block_size): n, c, h, w = x.size() unfolded_x = torch.nn.functional.unfold(x, block_size, stride=block_size) return unfolded_x.view(n, c * block_size**2, h // block_size, w // block_size) def generate_subimages(img, mask): n, c, h, w = img.shape subimage = torch.zeros(n, c, h // 2, w // 2, dtype=img.dtype, layout=img.layout, device=img.device) # per channel for i in range(c): img_per_channel = space_to_depth(img[:, i:i + 1, :, :], block_size=2) img_per_channel = img_per_channel.permute(0, 2, 3, 1).reshape(-1) subimage[:, i:i + 1, :, :] = img_per_channel[mask].reshape( n, h // 2, w // 2, 1).permute(0, 3, 1, 2) return subimage def get_generator(seed = None): global operation_seed_counter operation_seed_counter += 1 g_cuda_generator = torch.Generator(device="cuda") if seed == None: g_cuda_generator.manual_seed(operation_seed_counter) else: g_cuda_generator.manual_seed(seed) return g_cuda_generator def generate_mask_pair(img,seed=None): # prepare masks (N x C x H/2 x W/2) n, c, h, w = img.shape mask1 = torch.zeros(size=(n * h // 2 * w // 2 * 4, ), dtype=torch.bool, device=img.device) mask2 = torch.zeros(size=(n * h // 2 * w // 2 * 4, ), dtype=torch.bool, device=img.device) # prepare random mask pairs idx_pair = torch.tensor( [[0, 1], [0, 2], [1, 3], [2, 3], [1, 0], [2, 0], [3, 1], [3, 2]], dtype=torch.int64, device=img.device) rd_idx = torch.zeros(size=(n * h // 2 * w // 2, ), dtype=torch.int64, device=img.device) if seed == None: torch.randint(low=0, high=8, size=(n * h // 2 * w // 2, ), generator=get_generator(), out=rd_idx) else: torch.randint(low=0, high=8, size=(n * h // 2 * w // 2, ), generator=get_generator(seed), out=rd_idx) rd_pair_idx = idx_pair[rd_idx] rd_pair_idx += torch.arange(start=0, end=n * h // 2 * w // 2 * 4, step=4, dtype=torch.int64, device=img.device).reshape(-1, 1) # get masks mask1[rd_pair_idx[:, 0]] = 1 mask2[rd_pair_idx[:, 1]] = 1 return mask1, mask2 def generate_val_mask(img): n, c, h, w = img.shape mask1 = torch.zeros(size=(n * h // 2 * w // 2 * 4, ), dtype=torch.bool, device=img.device) mask2 = torch.zeros(size=(n * h // 2 * w // 2 * 4, ), dtype=torch.bool, device=img.device) mask3 = torch.zeros(size=(n * h // 2 * w // 2 * 4, ), dtype=torch.bool, device=img.device) mask4 = torch.zeros(size=(n * h // 2 * w // 2 * 4, ), dtype=torch.bool, device=img.device) mask1[torch.arange(start=0,end=n * h // 2 * w // 2 * 4, step=4, dtype=torch.int64,device=img.device)]=1 mask2[torch.arange(start=1,end=n * h // 2 * w // 2 * 4, step=4, dtype=torch.int64,device=img.device)]=1 mask3[torch.arange(start=2,end=n * h // 2 * w // 2 * 4, step=4, dtype=torch.int64,device=img.device)]=1 mask4[torch.arange(start=3,end=n * h // 2 * w // 2 * 4, step=4, dtype=torch.int64,device=img.device)]=1 return mask1, mask2, mask3, mask4 def load_model(args): USE_CUDA = True device = torch.device('cuda') if (torch.cuda.is_available() and USE_CUDA) else torch.device('cpu') checkpoint_path = glob.glob(args.output_dir +'/unet*') if len(checkpoint_path) != 1: raise ValueError("There is either no or more than one model to load") checkpoint_path = pathlib.Path(checkpoint_path[0] + f"/checkpoints/checkpoint_{args.restore_mode}.pt") state_dict = torch.load(checkpoint_path, map_location=lambda s, l: default_restore_location(s, "cpu")) args = argparse.Namespace(**{ **vars(state_dict["args"]), "no_log": True}) #model = models.build_model(args).to(device) model = models.unet_fastMRI( in_chans=args.in_chans, chans = args.chans, num_pool_layers = args.num_pool_layers, drop_prob = 0.0, residual_connection = args.residual, ).to(device) model.load_state_dict(state_dict["model"][0]) model.eval() return model def cli_main_test(args): USE_CUDA = True device = torch.device('cuda') if (torch.cuda.is_available() and USE_CUDA) else torch.device('cpu') model = load_model(args) # evaluate test performance over following noise range noise_std_range = np.linspace(args.test_noise_std_min, args.test_noise_std_max, ((args.test_noise_std_max-args.test_noise_std_min)//args.test_noise_stepsize)+1,dtype=int)/255. metrics_path = os.path.join(args.output_dir, args.test_mode + '_' + str(args.test_noise_std_min)+'-'+str(args.test_noise_std_max)+f'_metrics_{args.restore_mode}.p') metrics_dict = metrics_avg_on_noise_range(model, args, noise_std_range, device = device) pickle.dump( metrics_dict, open(metrics_path, "wb" ) ) def cli_main(args): available_models = glob.glob(f'{args.output_dir}/*') if not args.resume_training and available_models: raise ValueError('There exists already a trained model and resume_training is set False') if args.resume_training: f_restore_file(args) # reset the attributes of the function save_checkpoint mode = "max" default_score = float("inf") if mode == "min" else float("-inf") utils.save_checkpoint.best_score = default_score utils.save_checkpoint.best_step = -1 utils.save_checkpoint.best_epoch = -1 utils.save_checkpoint.last_step = -1 utils.save_checkpoint.current_lr = args.lr device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') # Set the name of the directory for saving results utils.setup_experiment(args) utils.init_logging(args) # Build data loaders, a model and an optimizer model = models.unet_fastMRI( in_chans=args.in_chans, chans = args.chans, num_pool_layers = args.num_pool_layers, drop_prob = 0.0, residual_connection = args.residual, ).to(device) optimizer = torch.optim.Adam(model.parameters(), lr=args.lr) num_epoch = args.n_epoch ratio = num_epoch / 100 scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[ int(20 * ratio) - 1, int(40 * ratio) - 1, int(60 * ratio) - 1, int(80 * ratio) - 1 ], gamma=args.lr_gamma) logging.info(f"Built a model consisting of {sum(p.numel() for p in model.parameters()):,} parameters") trainset = ImagenetSubdataset(args.train_size,args.path_to_ImageNet_train,mode='train',patch_size=args.patch_size,val_crop=args.val_crop) train_loader = DataLoader(trainset, batch_size=args.batch_size, shuffle=True, num_workers=8, pin_memory=True,generator=torch.Generator().manual_seed(args.seed)) valset = ImagenetSubdataset(args.val_size,args.path_to_ImageNet_train,mode='val',patch_size=args.patch_size,val_crop=args.val_crop) val_loader = DataLoader(valset, batch_size=1, shuffle=False, num_workers=4, pin_memory=True,generator=torch.Generator().manual_seed(args.seed)) print(optimizer.param_groups[0]["lr"]) if args.resume_training: state_dict = utils.load_checkpoint(args, model, optimizer, scheduler) global_step = state_dict['last_step'] start_epoch = int(state_dict['last_step']/(len(train_loader)))+1 start_decay = True elif args.no_annealing: global_step = -1 start_epoch = 0 start_decay = True else: global_step = -1 start_epoch = 0 start_decay = False print(optimizer.param_groups[0]["lr"]) args.log_interval = min(len(trainset), 100) # len(train_loader)=log once per epoch args.no_visual = False # True for not logging to tensorboard # Track moving average of loss values #train_meters = {name: RunningAverageMeter(0.98) for name in (["train_loss_running_avg"])} #train_meters = {name: AverageMeter() for name in (["train_loss"])} train_meters = {"train_loss": AverageMeter(), "train_loss_running_avg":RunningAverageMeter(0.98),"reconstruction_loss": AverageMeter(), "regularization_loss":AverageMeter()} valid_meters = {name: AverageMeter() for name in (["valid_psnr", "valid_ssim", "valid_psnr_self_supervised", "valid_ssim_self_supervised"])} # Create tensorflow event file writer = SummaryWriter(log_dir=args.experiment_dir) if not args.no_visual else None break_counter = 0 # store the best val performance from lr-interval before the last lr decay best_val_last = 0 # track the best val performance for the current lr-inerval best_val_current = 0 # count for how many lr intervals there was no improvement and break only if there was no improvement for 2 lr_interval_counter = 0 # if best_val_current at the end of the current lr interval is smaller than best_val_last we perform early stopping for epoch in range(start_epoch, args.num_epochs): start = time.process_time() train_bar = ProgressBar(train_loader, epoch) # At beginning of each epoch reset the train meters for meter in train_meters.values(): meter.reset() for inputs, noise_seed in train_bar: model.train() #Sets the module in training mode. global_step += 1 inputs = inputs.to(device) noise = get_noise(inputs,noise_seed, fix_noise = args.fix_noise, noise_std = args.noise_std/255.) noisy_inputs = noise + inputs mask1, mask2 = generate_mask_pair(noisy_inputs) noisy_sub1 = generate_subimages(noisy_inputs, mask1) noisy_sub2 = generate_subimages(noisy_inputs, mask2) with torch.no_grad(): noisy_denoised = model(noisy_inputs) noisy_sub1_denoised = generate_subimages(noisy_denoised, mask1) noisy_sub2_denoised = generate_subimages(noisy_denoised, mask2) noisy_output = model(noisy_sub1) noisy_target = noisy_sub2 Lambda = (epoch / args.n_epoch)* args.increase_ratio diff = noisy_output - noisy_target exp_diff = noisy_sub1_denoised - noisy_sub2_denoised loss1 = torch.mean(diff**2) reg = torch.mean((diff - exp_diff)**2) loss2 = Lambda * reg loss = args.Lambda1 * loss1 + args.Lambda2 * loss2 model.zero_grad() loss.backward() optimizer.step() train_meters["train_loss"].update(loss.item()) train_meters["reconstruction_loss"].update(loss1.item()) train_meters["regularization_loss"].update(reg.item()) train_meters["train_loss_running_avg"].update(loss.item()) train_bar.log(dict(**train_meters, lr=optimizer.param_groups[0]["lr"]), verbose=True) # Add to tensorflow event file: if writer is not None: writer.add_scalar("lr", optimizer.param_groups[0]["lr"], global_step) writer.add_scalar("loss/train", train_meters["train_loss"].avg, global_step) writer.add_scalar("loss/train_running_avg", train_meters["train_loss_running_avg"].avg, global_step) writer.add_scalar("loss/rec",train_meters["reconstruction_loss"].avg,global_step) writer.add_scalar("loss/reg",train_meters["regularization_loss"].avg,global_step) #gradients = torch.cat([p.grad.view(-1) for p in model.parameters() if p.grad is not None], dim=0) #writer.add_histogram("gradients", gradients, global_step) sys.stdout.flush() if epoch % args.valid_interval == 0: model.eval() gen_val = torch.Generator() gen_val = gen_val.manual_seed(10) for meter in valid_meters.values(): meter.reset() valid_bar = ProgressBar(val_loader) for sample, noise_seed in valid_bar: with torch.no_grad(): sample = sample.to(device) sample_np = sample.cpu().detach().numpy() shape = np.shape(sample_np) if shape[3] % 2 == 1: sample_np = sample_np[:,:,:,:-1] sample = torch.from_numpy(sample_np) sample = sample.to(device) #print(sample.size()) # Self-supervised validation with fixed noise noise_self_supervised = get_noise(sample,noise_seed, fix_noise = args.fix_noise, noise_std = args.noise_std/255.) noisy_input_fixed = sample + noise_self_supervised output_self_supervised = model(noisy_input_fixed) #noisy_target = sample + noise_target_self_supervised #if args.val_crop: mask1, mask2, mask3, mask4 = generate_val_mask(noisy_input_fixed) noisy_sub1 = generate_subimages(noisy_input_fixed, mask1) noisy_sub2 = generate_subimages(noisy_input_fixed, mask2) noisy_sub3 = generate_subimages(noisy_input_fixed, mask3) noisy_sub4 = generate_subimages(noisy_input_fixed, mask4) noisy_sub1_denoised = generate_subimages(output_self_supervised, mask1) noisy_sub2_denoised = generate_subimages(output_self_supervised, mask2) #noisy_sub1_denoised = generate_subimages(output_self_supervised, mask1) noisy_output = model(noisy_sub1) valid_psnr_self_supervised = calculate_psnr_neighbor(noisy_output,noisy_sub2,noisy_sub1_denoised,noisy_sub2_denoised,args.increase_ratio) valid_ssim_self_supervised = ssim(noisy_output, noisy_sub2) valid_meters["valid_psnr_self_supervised"].update(valid_psnr_self_supervised.item()) valid_meters["valid_ssim_self_supervised"].update(valid_ssim_self_supervised.item()) # Ground truth validation wit fixed noise # It uses the same input and output as in the self-supervised case since the noise seed is fixed valid_psnr = psnr(output_self_supervised, sample) valid_ssim = ssim(output_self_supervised, sample) valid_meters["valid_psnr"].update(valid_psnr.item()) valid_meters["valid_ssim"].update(valid_ssim.item()) if writer is not None: # Average is correct valid_meters['valid_psnr'].avg since .val would be just the psnr of last sample in val set. writer.add_scalar("psnr/valid", valid_meters['valid_psnr'].avg, global_step) writer.add_scalar("ssim/valid", valid_meters['valid_ssim'].avg, global_step) writer.add_scalar("psnr_selfsupervised/valid", valid_meters['valid_psnr_self_supervised'].avg, global_step) writer.add_scalar("ssim_selfsupervised/valid", valid_meters["valid_ssim_self_supervised"].avg, global_step) writer.add_scalar("lr", optimizer.param_groups[0]["lr"], global_step) sys.stdout.flush() if args.val_flag == 0: # if we do self-supervised validation val_loss = valid_meters["valid_psnr_self_supervised"].avg else: # if we do supervised validation val_loss = valid_meters["valid_psnr"].avg if utils.save_checkpoint.best_score < val_loss and not start_decay: utils.save_checkpoint(args, global_step, epoch, model, optimizer, score=val_loss, mode="max") current_lr = utils.save_checkpoint.current_lr optimizer.param_groups[0]["lr"] = current_lr*args.lr_beta utils.save_checkpoint.current_lr = current_lr*args.lr_beta annealing_counter = 0 elif not start_decay: annealing_counter += 1 current_lr = utils.save_checkpoint.current_lr if annealing_counter == args.lr_patience_annealing: available_models = glob.glob(f'{args.output_dir}/*') if not available_models: raise ValueError('No file to restore') elif len(available_models)>1: raise ValueError('Too many files to restore from') model_path = os.path.join(available_models[0], "checkpoints/checkpoint_best.pt") state_dict = torch.load(model_path, map_location=lambda s, l: default_restore_location(s, "cpu")) model = [model] if model is not None and not isinstance(model, list) else model for m, state in zip(model, state_dict["model"]): m.load_state_dict(state) model = model[0] optimizer.param_groups[0]["lr"] = current_lr/(args.lr_beta*args.inital_decay_factor) start_decay = True num_epoch = args.n_epoch ratio = num_epoch / 100 scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[ int(20 * ratio) - 1, int(40 * ratio) - 1, int(60 * ratio) - 1, int(80 * ratio) - 1 ], gamma=args.lr_gamma) else: utils.save_checkpoint(args, global_step, epoch, model, optimizer, score=val_loss, mode="max") current_lr = optimizer.param_groups[0]["lr"] if val_loss > best_val_current: best_val_current = val_loss if writer is not None: writer.add_scalar("epoch", epoch, global_step) sys.stdout.flush() if start_decay: current_lr = optimizer.param_groups[0]["lr"] scheduler.step() new_lr = optimizer.param_groups[0]["lr"] #At every lr decay check if the model did not improve during the current or the previous lr interval and break if it didn't. if new_lr < current_lr: if best_val_current < best_val_last and lr_interval_counter==1: logging.info('Break training due to convergence of val loss!') break elif best_val_current < best_val_last and lr_interval_counter==0: lr_interval_counter += 1 logging.info('Do not yet break due to convergence of val loss!') else: best_val_last = best_val_current best_val_current = 0 lr_interval_counter = 0 end = time.process_time() - start logging.info(train_bar.print(dict(**train_meters, **valid_meters, lr=current_lr, time=np.round(end/60,3)))) if optimizer.param_groups[0]["lr"] == args.lr_min and start_decay: break_counter += 1 if break_counter == args.break_counter: print('Break training due to minimal learning rate constraint!') break logging.info(f"Done training! Best PSNR {utils.save_checkpoint.best_score:.3f} obtained after step {utils.save_checkpoint.best_step} (epoch {utils.save_checkpoint.best_epoch}).") def get_args(hp,ee,rr): parser = argparse.ArgumentParser(allow_abbrev=False) # Add data arguments parser.add_argument("--train-size", default=None, help="number of examples in training set") parser.add_argument("--val-size", default=40, help="number of examples in validation set") parser.add_argument("--test-size", default=100, help="number of examples in test set") parser.add_argument("--val-crop", default=True, type=bool, help="Crop validation images to train size.") parser.add_argument("--patch-size", default=128, help="size of the center cropped HR image") parser.add_argument("--batch-size", default=128, type=int, help="train batch size") # Add model arguments parser.add_argument("--model", default="unet", help="model architecture") # Add noise arguments parser.add_argument('--noise_std', default = 15, type = float, help = 'noise level') parser.add_argument('--test_noise_std_min', default = 15, type = float, help = 'minimal noise level for testing') parser.add_argument('--test_noise_std_max', default = 15, type = float, help = 'maximal noise level for testing') parser.add_argument('--test_noise_stepsize', default = 5, type = float, help = 'Stepsize between test_noise_std_min and test_noise_std_max') # Add optimization arguments parser.add_argument("--lr", default=1e-3, type=float, help="learning rate") parser.add_argument("--lr-gamma", default=0.5, type=float, help="factor by which to reduce learning rate") parser.add_argument("--lr-beta", default=2, type=float, help="factor by which to increase learning rate") parser.add_argument("--lr-patience", default=5, type=int, help="epochs without improvement before lr decay") parser.add_argument("--no_annealing", default=True, type=bool, help="Use lr annealing or not.") parser.add_argument("--lr-patience-annealing", default=3, type=int, help="epochs without improvement before lr annealing stops") parser.add_argument("--lr-min", default=1e-5, type=float, help="Once we reach this learning rate continue for break_counter many epochs then stop.") parser.add_argument("--lr-threshold", default=0.003, type=float, help="Improvements by less than this threshold are not counted for decay patience.") parser.add_argument("--break-counter", default=9, type=int, help="Once smallest learning rate is reached, continue for so many epochs before stopping.") parser.add_argument("--inital-decay-factor", default=2, type=int, help="After annealing found a lr for which val loss does not improve, go back initial_decay_factor many lrs") parser.add_argument("--num-epochs", default=100, type=int, help="force stop training at specified epoch") parser.add_argument("--valid-interval", default=1, type=int, help="evaluate every N epochs") parser.add_argument("--save-interval", default=1, type=int, help="save a checkpoint every N steps") # Add model arguments parser = models.unet_fastMRI.add_args(parser) parser = utils.add_logging_arguments(parser) #args = parser.parse_args() args, _ = parser.parse_known_args() # Set arguments specific for this experiment dargs = vars(args) for key in hp.keys(): dargs[key] = hp[key][ee] args.seed = int(42 + 10*rr) return args def f_restore_file(args): #available_models = glob.glob(f'{args.output_dir}/{args.experiment}-*') available_models = glob.glob(f'{args.output_dir}/*') if not available_models: raise ValueError('No file to restore') if not args.restore_mode: raise ValueError("Pick restore mode either 'best' 'last' or '\path\to\checkpoint\dir'") if args.restore_mode=='best': mode = "max" best_score = float("inf") if mode == "min" else float("-inf") best_model = None for modelp in available_models: model_path = os.path.join(modelp, "checkpoints/checkpoint_best.pt") if os.path.isfile(model_path): state_dict = torch.load(model_path, map_location=lambda s, l: default_restore_location(s, "cpu")) score = state_dict["best_score"] if (score < best_score and mode == "min") or (score > best_score and mode == "max"): best_score = score best_model = model_path best_modelp = modelp best_step = state_dict["best_step"] best_epoch = state_dict["best_epoch"] args.restore_file = best_model args.experiment_dir = best_modelp #logging.info(f"Prepare to restore best model {best_model} with PSNR {best_score} at step {best_step}, epoch {best_epoch}") elif args.restore_mode=='last': last_step = -1 last_model = None for modelp in available_models: model_path = os.path.join(modelp, "checkpoints/checkpoint_last.pt") if os.path.isfile(model_path): state_dict = torch.load(model_path, map_location=lambda s, l: default_restore_location(s, "cpu")) step = state_dict["last_step"] if step > last_step: last_step = step last_model = model_path last_modelp = modelp score = state_dict["score"] last_epoch = state_dict["epoch"] args.restore_file = last_model args.experiment_dir = last_modelp #logging.info(f"Prepare to restore last model {last_model} with PSNR {score} at step {last_step}, epoch {last_epoch}") else: args.restore_file = args.restore_mode args.experiment_dir = args.restore_mode[:args.restore_mode.find('/checkpoints')] def infer_images(args): USE_CUDA = True device = torch.device('cuda') if (torch.cuda.is_available() and USE_CUDA) else torch.device('cpu') net = load_model(args) # the denoiser seed_dict = { "val":10, "test":20, } gen = torch.Generator() gen = gen.manual_seed(seed_dict[args.test_mode]) # Load the test images load_path = '../training_set_lists/' if args.test_mode == 'test': files_source = torch.load(load_path+f'ImageNetTest{args.test_size}_filepaths.pt') #files_source.sort() elif args.test_mode == 'val': files_source = torch.load(load_path+f'ImageNetVal{args.val_size}_filepaths.pt') #files_source.sort() if not os.path.isdir(args.output_dir+'/test_images'): os.mkdir(args.output_dir+'/test_images') counter = 0 transformT = transforms.ToTensor() transformIm = transforms.ToPILImage() for f in files_source: counter = counter + 1 if counter > 3: break # Create noise ISource = torch.unsqueeze(transformT(Image.open(f).convert("RGB")),0).to(device) noise = torch.randn(ISource.shape,generator = gen) * args.noise_std/255. INoisy = noise.to(device) + ISource out = torch.clamp(net(INoisy), 0., 1.).cpu() out = torch.squeeze(out,0) # Get rid of the 1 in dim 0. im = transformIm(out) INoisy = torch.clamp(torch.squeeze(INoisy,0), 0., 1.).cpu() INoisy = transformIm(INoisy) clean_image = Image.open(f).convert("RGB") im.save(args.output_dir+f'/test_images/im{counter}_denoised_notclamped.png') clean_image.save(args.output_dir+f'/test_images/im{counter}_ground_truth_notclamped.png') INoisy.save(args.output_dir+f'/test_images/im{counter}_noisy_notclamped.png') im.save(args.output_dir+f'/test_images/im{counter}_denoised_notclamped.pdf') clean_image.save(args.output_dir+f'/test_images/im{counter}_ground_truth_notclamped.pdf') INoisy.save(args.output_dir+f'/test_images/im{counter}_noisy_notclamped.pdf')
29,827
45.244961
182
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Neighbor2Neighbor/utils/util_calculate_psnr_ssim.py
import cv2 import numpy as np import torch # from https://github.com/JingyunLiang/SwinIR/blob/328dda0f4768772e6d8c5aa3d5aa8e24f1ad903b/utils/util_calculate_psnr_ssim.py#L80 def calculate_psnr(img1, img2, crop_border, input_order='HWC', test_y_channel=False): """Calculate PSNR (Peak Signal-to-Noise Ratio). Ref: https://en.wikipedia.org/wiki/Peak_signal-to-noise_ratio Args: img1 (ndarray): Images with range [0, 255]. img2 (ndarray): Images with range [0, 255]. crop_border (int): Cropped pixels in each edge of an image. These pixels are not involved in the PSNR calculation. input_order (str): Whether the input order is 'HWC' or 'CHW'. Default: 'HWC'. test_y_channel (bool): Test on Y channel of YCbCr. Default: False. Returns: float: psnr result. """ assert img1.shape == img2.shape, (f'Image shapes are differnet: {img1.shape}, {img2.shape}.') if input_order not in ['HWC', 'CHW']: raise ValueError(f'Wrong input_order {input_order}. Supported input_orders are ' '"HWC" and "CHW"') img1 = reorder_image(img1, input_order=input_order) img2 = reorder_image(img2, input_order=input_order) img1 = img1.astype(np.float64) img2 = img2.astype(np.float64) if crop_border != 0: img1 = img1[crop_border:-crop_border, crop_border:-crop_border, ...] img2 = img2[crop_border:-crop_border, crop_border:-crop_border, ...] if test_y_channel: img1 = to_y_channel(img1) img2 = to_y_channel(img2) mse = np.mean((img1 - img2) ** 2) if mse == 0: return float('inf') return 20. * np.log10(255. / np.sqrt(mse)) def calculate_psnr_neighbor(img1, img2,noisy_sub1_denoised,noisy_sub2_denoised,increase_ratio, crop_border=0, input_order='HWC', test_y_channel=False): """Calculate PSNR (Peak Signal-to-Noise Ratio). Ref: https://en.wikipedia.org/wiki/Peak_signal-to-noise_ratio Args: img1 (ndarray): Images with range [0, 255]. img2 (ndarray): Images with range [0, 255]. crop_border (int): Cropped pixels in each edge of an image. These pixels are not involved in the PSNR calculation. input_order (str): Whether the input order is 'HWC' or 'CHW'. Default: 'HWC'. test_y_channel (bool): Test on Y channel of YCbCr. Default: False. Returns: float: psnr result. """ assert img1.shape == img2.shape, (f'Image shapes are differnet: {img1.shape}, {img2.shape}.') if input_order not in ['HWC', 'CHW']: raise ValueError(f'Wrong input_order {input_order}. Supported input_orders are ' '"HWC" and "CHW"') img1 = reorder_image(img1, input_order=input_order) noisy_sub1_denoised = reorder_image(noisy_sub1_denoised,input_order=input_order) noisy_sub2_denoised = reorder_image(noisy_sub2_denoised,input_order=input_order) img2 = reorder_image(img2, input_order=input_order) #img1 = img1.astype(np.float64) #img2 = img2.astype(np.float64) #noisy_sub1_denoised = noisy_sub1_denoised.astype(np.float64) #noisy_sub2_denoised = noisy_sub2_denoised.astype(np.float64) img1 = img1.cpu().detach().numpy() img2 = img2.cpu().detach().numpy() noisy_sub1_denoised = noisy_sub1_denoised.cpu().detach().numpy() noisy_sub2_denoised = noisy_sub2_denoised.cpu().detach().numpy() if crop_border != 0: img1 = img1[crop_border:-crop_border, crop_border:-crop_border, ...] img2 = img2[crop_border:-crop_border, crop_border:-crop_border, ...] if test_y_channel: img1 = to_y_channel(img1) img2 = to_y_channel(img2) mse = np.mean((img1 - img2) ** 2) reg = increase_ratio * np.mean((img1 - img2 - (noisy_sub1_denoised - noisy_sub2_denoised)) ** 2) if mse == 0: return float('inf') return 20. * np.log10(1. / np.sqrt(mse)) def calculate_upsnr(img1,img2,img3,img4,noisy_sub1_denoised,noisy_sub2_denoised,increase_ratio, crop_border=0,input_order='HWC', test_y_channel=False): """Calculate PSNR (Peak Signal-to-Noise Ratio). Ref: https://en.wikipedia.org/wiki/Peak_signal-to-noise_ratio Args: img1 (ndarray): Images with range [0, 255]. img2 (ndarray): Images with range [0, 255]. crop_border (int): Cropped pixels in each edge of an image. These pixels are not involved in the PSNR calculation. input_order (str): Whether the input order is 'HWC' or 'CHW'. Default: 'HWC'. test_y_channel (bool): Test on Y channel of YCbCr. Default: False. Returns: float: psnr result. """ assert img1.shape == img2.shape, (f'Image shapes are differnet: {img1.shape}, {img2.shape}.') if input_order not in ['HWC', 'CHW']: raise ValueError(f'Wrong input_order {input_order}. Supported input_orders are ' '"HWC" and "CHW"') img1 = reorder_image(img1, input_order=input_order) img3 = reorder_image(img3,input_order=input_order) img4 = reorder_image(img4,input_order=input_order) img2 = reorder_image(img2, input_order=input_order) noisy_sub1_denoised = reorder_image(noisy_sub1_denoised,input_order=input_order) noisy_sub2_denoised = reorder_image(noisy_sub2_denoised,input_order=input_order) #img1 = img1.astype(np.float64) #img2 = img2.astype(np.float64) img1 = img1.cpu().detach().numpy() img2 = img2.cpu().detach().numpy() img3 = img3.cpu().detach().numpy() img4 = img4.cpu().detach().numpy() noisy_sub1_denoised = noisy_sub1_denoised.cpu().detach().numpy() noisy_sub2_denoised = noisy_sub2_denoised.cpu().detach().numpy() if crop_border != 0: img1 = img1[crop_border:-crop_border, crop_border:-crop_border, ...] img2 = img2[crop_border:-crop_border, crop_border:-crop_border, ...] if test_y_channel: img1 = to_y_channel(img1) img2 = to_y_channel(img2) umse = np.mean((img1 - img2) ** 2) + 0.5 * np.mean((img3 - img4) ** 2) reg = increase_ratio * np.mean((img1 - img2 - (noisy_sub1_denoised - noisy_sub2_denoised)) ** 2) if umse == 0: return float('inf') #return 20. * np.log10(1. / np.sqrt(umse+reg)) return 20. * np.log10(1. / np.sqrt(umse)) # E720 - 725 def _ssim(img1, img2): """Calculate SSIM (structural similarity) for one channel images. It is called by func:`calculate_ssim`. Args: img1 (ndarray): Images with range [0, 255] with order 'HWC'. img2 (ndarray): Images with range [0, 255] with order 'HWC'. Returns: float: ssim result. """ C1 = (0.01 * 255) ** 2 C2 = (0.03 * 255) ** 2 img1 = img1.astype(np.float64) img2 = img2.astype(np.float64) kernel = cv2.getGaussianKernel(11, 1.5) window = np.outer(kernel, kernel.transpose()) mu1 = cv2.filter2D(img1, -1, window)[5:-5, 5:-5] mu2 = cv2.filter2D(img2, -1, window)[5:-5, 5:-5] mu1_sq = mu1 ** 2 mu2_sq = mu2 ** 2 mu1_mu2 = mu1 * mu2 sigma1_sq = cv2.filter2D(img1 ** 2, -1, window)[5:-5, 5:-5] - mu1_sq sigma2_sq = cv2.filter2D(img2 ** 2, -1, window)[5:-5, 5:-5] - mu2_sq sigma12 = cv2.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2 ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) * (sigma1_sq + sigma2_sq + C2)) return ssim_map.mean() def calculate_ssim(img1, img2, crop_border, input_order='HWC', test_y_channel=False): """Calculate SSIM (structural similarity). Ref: Image quality assessment: From error visibility to structural similarity The results are the same as that of the official released MATLAB code in https://ece.uwaterloo.ca/~z70wang/research/ssim/. For three-channel images, SSIM is calculated for each channel and then averaged. Args: img1 (ndarray): Images with range [0, 255]. img2 (ndarray): Images with range [0, 255]. crop_border (int): Cropped pixels in each edge of an image. These pixels are not involved in the SSIM calculation. input_order (str): Whether the input order is 'HWC' or 'CHW'. Default: 'HWC'. test_y_channel (bool): Test on Y channel of YCbCr. Default: False. Returns: float: ssim result. """ assert img1.shape == img2.shape, (f'Image shapes are differnet: {img1.shape}, {img2.shape}.') if input_order not in ['HWC', 'CHW']: raise ValueError(f'Wrong input_order {input_order}. Supported input_orders are ' '"HWC" and "CHW"') img1 = reorder_image(img1, input_order=input_order) img2 = reorder_image(img2, input_order=input_order) img1 = img1.astype(np.float64) img2 = img2.astype(np.float64) if crop_border != 0: img1 = img1[crop_border:-crop_border, crop_border:-crop_border, ...] img2 = img2[crop_border:-crop_border, crop_border:-crop_border, ...] if test_y_channel: img1 = to_y_channel(img1) img2 = to_y_channel(img2) ssims = [] for i in range(img1.shape[2]): ssims.append(_ssim(img1[..., i], img2[..., i])) return np.array(ssims).mean() def reorder_image(img, input_order='HWC'): """Reorder images to 'HWC' order. If the input_order is (h, w), return (h, w, 1); If the input_order is (c, h, w), return (h, w, c); If the input_order is (h, w, c), return as it is. Args: img (ndarray): Input image. input_order (str): Whether the input order is 'HWC' or 'CHW'. If the input image shape is (h, w), input_order will not have effects. Default: 'HWC'. Returns: ndarray: reordered image. """ if input_order not in ['HWC', 'CHW']: raise ValueError(f'Wrong input_order {input_order}. Supported input_orders are ' "'HWC' and 'CHW'") if len(img.shape) == 2: img = img[..., None] if input_order == 'CHW': img = img.transpose(1, 2, 0) return img def to_y_channel(img): """Change to Y channel of YCbCr. Args: img (ndarray): Images with range [0, 255]. Returns: (ndarray): Images with range [0, 255] (float type) without round. """ img = img.astype(np.float32) / 255. if img.ndim == 3 and img.shape[2] == 3: img = bgr2ycbcr(img, y_only=True) img = img[..., None] return img * 255. def _convert_input_type_range(img): """Convert the type and range of the input image. It converts the input image to np.float32 type and range of [0, 1]. It is mainly used for pre-processing the input image in colorspace convertion functions such as rgb2ycbcr and ycbcr2rgb. Args: img (ndarray): The input image. It accepts: 1. np.uint8 type with range [0, 255]; 2. np.float32 type with range [0, 1]. Returns: (ndarray): The converted image with type of np.float32 and range of [0, 1]. """ img_type = img.dtype img = img.astype(np.float32) if img_type == np.float32: pass elif img_type == np.uint8: img /= 255. else: raise TypeError('The img type should be np.float32 or np.uint8, ' f'but got {img_type}') return img def _convert_output_type_range(img, dst_type): """Convert the type and range of the image according to dst_type. It converts the image to desired type and range. If `dst_type` is np.uint8, images will be converted to np.uint8 type with range [0, 255]. If `dst_type` is np.float32, it converts the image to np.float32 type with range [0, 1]. It is mainly used for post-processing images in colorspace convertion functions such as rgb2ycbcr and ycbcr2rgb. Args: img (ndarray): The image to be converted with np.float32 type and range [0, 255]. dst_type (np.uint8 | np.float32): If dst_type is np.uint8, it converts the image to np.uint8 type with range [0, 255]. If dst_type is np.float32, it converts the image to np.float32 type with range [0, 1]. Returns: (ndarray): The converted image with desired type and range. """ if dst_type not in (np.uint8, np.float32): raise TypeError('The dst_type should be np.float32 or np.uint8, ' f'but got {dst_type}') if dst_type == np.uint8: img = img.round() else: img /= 255. return img.astype(dst_type) def bgr2ycbcr(img, y_only=False): """Convert a BGR image to YCbCr image. The bgr version of rgb2ycbcr. It implements the ITU-R BT.601 conversion for standard-definition television. See more details in https://en.wikipedia.org/wiki/YCbCr#ITU-R_BT.601_conversion. It differs from a similar function in cv2.cvtColor: `BGR <-> YCrCb`. In OpenCV, it implements a JPEG conversion. See more details in https://en.wikipedia.org/wiki/YCbCr#JPEG_conversion. Args: img (ndarray): The input image. It accepts: 1. np.uint8 type with range [0, 255]; 2. np.float32 type with range [0, 1]. y_only (bool): Whether to only return Y channel. Default: False. Returns: ndarray: The converted YCbCr image. The output image has the same type and range as input image. """ img_type = img.dtype img = _convert_input_type_range(img) if y_only: out_img = np.dot(img, [24.966, 128.553, 65.481]) + 16.0 else: out_img = np.matmul( img, [[24.966, 112.0, -18.214], [128.553, -74.203, -93.786], [65.481, -37.797, 112.0]]) + [16, 128, 128] out_img = _convert_output_type_range(out_img, img_type) return out_img
13,571
40.631902
151
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Neighbor2Neighbor/utils/metrics.py
import numpy as np from skimage.metrics import peak_signal_noise_ratio, structural_similarity def ssim(clean, noisy, normalized=True): """Use skimage.meamsure.compare_ssim to calculate SSIM Args: clean (Tensor): (B, C, H, W) noisy (Tensor): (B, C, H, W) normalized (bool): If True, the range of tensors are [0., 1.] else [0, 255] Returns: SSIM per image: (B, ) """ if len(clean.shape)!=4 or len(noisy.shape)!=4: raise ValueError("ssim expects clean (Tensor): (B, C, H, W) noisy (Tensor): (B, C, H, W)") if normalized: clean = clean.mul(255).clamp(0, 255) noisy = noisy.mul(255).clamp(0, 255) clean = clean.cpu().detach().numpy().astype(np.float32) noisy = noisy.cpu().detach().numpy().astype(np.float32) clean = np.moveaxis(clean,1,-1) noisy = np.moveaxis(noisy,1,-1) return np.array([structural_similarity(c, n, data_range=255, multichannel=True) for c, n in zip(clean, noisy)]).mean() def psnr(clean, noisy, normalized=True): """Use skimage.meamsure.compare_ssim to calculate SSIM Args: clean (Tensor): (B, C, H, W) noisy (Tensor): (B, C, H, W) normalized (bool): If True, the range of tensors are [0., 1.] else [0, 255] Returns: SSIM per image: (B, ) """ if len(clean.shape)!=4 or len(noisy.shape)!=4: raise ValueError("psnr expects clean (Tensor): (B, C, H, W) noisy (Tensor): (B, C, H, W)") if normalized: clean = clean.mul(255).clamp(0, 255) noisy = noisy.mul(255).clamp(0, 255) clean = clean.cpu().detach().numpy().astype(np.float32) noisy = noisy.cpu().detach().numpy().astype(np.float32) return np.array([peak_signal_noise_ratio(c, n, data_range=255) for c, n in zip(clean, noisy)]).mean()
1,811
36.75
122
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Neighbor2Neighbor/utils/__init__.py
from .train_utils import * from .main_function_helpers import *
63
31
36
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Neighbor2Neighbor/utils/test_metrics.py
import torch import numpy as np import matplotlib.pyplot as plt import glob import os #import cv2 from utils.noise_model import get_noise from utils.metrics import ssim,psnr from utils.util_calculate_psnr_ssim import calculate_psnr,calculate_ssim from skimage import color import PIL.Image as Image import torchvision.transforms as transforms from utils.utils_image import * metrics_key = ['psnr_m', 'psnr_s', 'psnr_delta_m', 'psnr_delta_s', 'ssim_m', 'ssim_s', 'ssim_delta_m', 'ssim_delta_s']; def tensor_to_image(torch_image, low=0.0, high = 1.0, clamp = True): if clamp: torch_image = torch.clamp(torch_image, low, high); return torch_image[0,0].cpu().data.numpy() def normalize(data): return data/255. def convert_dict_to_string(metrics): return_string = ''; for x in metrics.keys(): return_string += x+': '+str(round(metrics[x], 3))+' '; return return_string def get_all_comparison_metrics(denoised, source, noisy = None, scale=None, return_title_string = False, clamp = True): metrics = {}; metrics['psnr'] = np.zeros(len(denoised)) metrics['ssim'] = np.zeros(len(denoised)) if noisy is not None: metrics['psnr_delta'] = np.zeros(len(denoised)) metrics['ssim_delta'] = np.zeros(len(denoised)) if clamp: denoised = torch.clamp(denoised, 0.0, 1.0) metrics['psnr'] = psnr(source, denoised); metrics['ssim'] = ssim(source, denoised); if noisy is not None: metrics['psnr_delta'] = metrics['psnr'] - psnr(source, noisy); metrics['ssim_delta'] = metrics['ssim'] - ssim(source, noisy); if return_title_string: return convert_dict_to_string(metrics) else: return metrics def average_on_folder(args, net, noise_std, verbose=True, device = torch.device('cuda')): #if verbose: #print('Loading data info ...\n') print(f'\n Dataset: {args.test_mode}, Restore mode: {args.restore_mode}') load_path = '../training_set_lists/' seed_dict = { "val":10, "test":20, } gen = torch.Generator() gen = gen.manual_seed(seed_dict[args.test_mode]) if args.test_mode == 'test': files_source = torch.load(load_path+f'ImageNetTest{args.test_size}_filepaths.pt') #files_source.sort() elif args.test_mode == 'val': files_source = torch.load(load_path+f'ImageNetVal{args.val_size}_filepaths.pt') avreage_metrics_key = ['psnr', 'psnr_delta', 'ssim', 'ssim_delta'] avg_metrics = {}; for x in avreage_metrics_key: avg_metrics[x] = []; psnr_list = [] ssim_list = [] for f in files_source: transformT = transforms.ToTensor() ISource = torch.unsqueeze(transformT(Image.open(args.path_to_ImageNet_train + f).convert("RGB")),0).to(device) if args.test_mode == 'val': noise_seed = int(f[f.find('train/')+17:-5].replace('_','')) gen = gen.manual_seed(noise_seed) noise = torch.randn(ISource.shape,generator = gen) * args.noise_std/255. INoisy = noise.to(device) + ISource out = torch.clamp(net(INoisy), 0., 1.) ind_metrics = get_all_comparison_metrics(out, ISource, INoisy, return_title_string = False) for x in avreage_metrics_key: avg_metrics[x].append(ind_metrics[x]) if(verbose): print("%s %s" % (f, convert_dict_to_string(ind_metrics))) metrics = {} for x in avreage_metrics_key: metrics[x+'_m'] = np.mean(avg_metrics[x]) metrics[x+'_s'] = np.std(avg_metrics[x]) if verbose: print("\n Average %s" % (convert_dict_to_string(metrics))) #if(not verbose): return metrics def metrics_avg_on_noise_range(net, args, noise_std_array, device = torch.device('cuda')): array_metrics = {} for x in metrics_key: array_metrics[x] = np.zeros(len(noise_std_array)) for j, noise_std in enumerate(noise_std_array): metric_list = average_on_folder(args, net, noise_std = noise_std, verbose=False, device=device); for x in metrics_key: array_metrics[x][j] += metric_list[x] print('noise: ', int(noise_std*255), ' ', x, ': ', str(array_metrics[x][j])) return array_metrics
4,404
29.804196
119
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Neighbor2Neighbor/utils/noise_model.py
import torch def get_noise(data, noise_seed, fix_noise, noise_std = float(25)/255.0): if fix_noise: device = torch.device('cuda') gen = torch.Generator(device=device) batch_size = data.size(dim=0) tensor_dim = list(data.size())[1:] for i in range(0,batch_size): gen = gen.manual_seed(noise_seed[i].item()) noise = torch.randn(tensor_dim,generator = gen, device=device) * noise_std noise = torch.unsqueeze(noise,0) if i == 0: noise_tensor = noise else: noise_tensor = torch.cat((noise_tensor, noise),0) noise = noise_tensor #noise = torch.randn(data.shape,generator = gen, device=device) * noise_std else: noise = torch.randn_like(data) noise.data = noise.data * noise_std return noise
880
31.62963
87
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Neighbor2Neighbor/utils/meters.py
import time import torch class AverageMeter(object): def __init__(self): self.reset() def reset(self): self.val = 0 self.avg = 0 self.sum = 0 self.count = 0 def update(self, val, n=1): if isinstance(val, torch.Tensor): val = val.item() self.val = val / n self.sum += val self.count += n self.avg = self.sum / self.count class RunningAverageMeter(object): def __init__(self, momentum=0.98): self.momentum = momentum self.reset() def reset(self): self.val = None self.avg = 0 def update(self, val): if isinstance(val, torch.Tensor): val = val.item() if self.val is None: self.avg = val else: self.avg = self.avg * self.momentum + val * (1 - self.momentum) self.val = val class TimeMeter(object): def __init__(self, init=0): self.reset(init) def reset(self, init=0): self.init = init self.start = time.time() self.n = 0 def update(self, val=1): self.n += val @property def avg(self): return self.n / self.elapsed_time @property def elapsed_time(self): return self.init + (time.time() - self.start)
1,321
20.322581
75
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Neighbor2Neighbor/utils/data_helpers/load_datasets_helpers.py
import os import os.path import numpy as np import h5py import torch import torchvision.transforms as transforms import PIL.Image as Image from utils.utils_image import * class ImagenetSubdataset(torch.utils.data.Dataset): def __init__(self, size, path_to_ImageNet_train, mode='train', patch_size='128', val_crop=True): super().__init__() load_path = '../training_set_lists/' self.path_to_ImageNet_train = path_to_ImageNet_train if mode=='train': self.files = torch.load(load_path+f'trsize{size}_filepaths.pt') self.transform = transforms.Compose([ transforms.CenterCrop(patch_size), transforms.ToTensor(), ]) elif mode=='val': self.files = torch.load(load_path+f'ImageNetVal{size}_filepaths.pt') #print(self.files) if val_crop: self.transform = transforms.Compose([ transforms.CenterCrop(patch_size), transforms.ToTensor(), ]) else: self.transform = transforms.Compose([ transforms.ToTensor(), ]) self.noise_seeds = {} for i, file in enumerate(self.files): key = file[file.find('train/')+16:-5] number = int(file[file.find('train/')+17:-5].replace('_','')) self.noise_seeds[key] = number def __len__(self): return len(self.files) def __getitem__(self, index): file = self.files[index] key = file[file.find('train/')+16:-5] noise_seed = self.noise_seeds[key] image = Image.open(self.path_to_ImageNet_train + self.files[index]).convert("RGB") #ImageNet contains some grayscale images data = self.transform(image) return data, noise_seed
1,949
32.62069
131
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noise2Noise/models/unet_fastMRI.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ import torch import math from torch import nn from torch.nn import functional as F class unet_fastMRI(nn.Module): """ PyTorch implementation of a U-Net model. O. Ronneberger, P. Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015. """ def __init__( self, in_chans: int, chans: int = 32, num_pool_layers: int = 4, drop_prob: float = 0.0, residual_connection: bool = True, ): """ Args: in_chans: Number of channels in the input to the U-Net model. out_chans: Number of channels in the output to the U-Net model. chans: Number of output channels of the first convolution layer. num_pool_layers: Number of down-sampling and up-sampling layers. drop_prob: Dropout probability. residual_connection: Network outputs the residual between input and output. """ super().__init__() self.in_chans = in_chans self.out_chans = in_chans self.chans = chans self.num_pool_layers = num_pool_layers self.drop_prob = drop_prob self.residual_connection = residual_connection self.down_sample_layers = nn.ModuleList([ConvBlock(in_chans, chans, drop_prob)]) #first conv block followed by downsampling ch = chans for _ in range(num_pool_layers - 1): self.down_sample_layers.append(ConvBlock(ch, ch * 2, drop_prob)) #conv blocks that are followed by downsampling ch *= 2 self.conv = ConvBlock(ch, ch * 2, drop_prob) self.up_conv = nn.ModuleList() self.up_transpose_conv = nn.ModuleList() for _ in range(num_pool_layers - 1): self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch)) #transposed conv blocks self.up_conv.append(ConvBlock(ch * 2, ch, drop_prob)) ch //= 2 self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch)) #last transposed conv block self.up_conv.append( nn.Sequential(# ConvBlock(ch * 2, ch, drop_prob), nn.Conv2d(ch, self.out_chans, kernel_size=1, stride=1),# )# ) def add_args(parser): """Add model-specific arguments to the parser.""" parser.add_argument("--bias", default=True, help="use residual bias") parser.add_argument("--residual", default=True, help="use residual connection") parser.add_argument("--in-chans", default=3, help="Either color (3) or grey (1)") parser.add_argument("--chans", default=32, help="Number of channels in outer most layer") parser.add_argument("--num-pool-layers", default=3, help="Number of layers per down- and up-sampling path.") parser.add_argument("--no-pooling", default=False, help="No downsampling. Use the no_unet_fastMRI module.") return parser def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H, W)`. """ stack = [] output = image.detach().clone() # apply down-sampling layers for layer in self.down_sample_layers: output = layer(output) stack.append(output) output = F.avg_pool2d(output, kernel_size=2, stride=2, padding=0) output = self.conv(output) # apply up-sampling layers for transpose_conv, conv in zip(self.up_transpose_conv, self.up_conv): downsample_layer = stack.pop() output = transpose_conv(output) # reflect pad on the right/botton if needed to handle odd input dimensions padding = [0, 0, 0, 0] if output.shape[-1] != downsample_layer.shape[-1]: padding[1] = 1 # padding right if output.shape[-2] != downsample_layer.shape[-2]: padding[3] = 1 # padding bottom if torch.sum(torch.tensor(padding)) != 0: output = F.pad(output, padding, "reflect") output = torch.cat([output, downsample_layer], dim=1) output = conv(output) if self.residual_connection: output = image - output; return output class ConvBlock(nn.Module): """ A Convolutional Block that consists of two convolution layers each followed by instance normalization, LeakyReLU activation and dropout. """ def __init__(self, in_chans: int, out_chans: int, drop_prob: float): """ Args: in_chans: Number of channels in the input. out_chans: Number of channels in the output. drop_prob: Dropout probability. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.drop_prob = drop_prob self.layers = nn.Sequential( nn.Conv2d(in_chans, out_chans, kernel_size=3, padding=1, bias=False), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Dropout2d(drop_prob), nn.Conv2d(out_chans, out_chans, kernel_size=3, padding=1, bias=False), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Dropout2d(drop_prob), ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H, W)`. """ return self.layers(image) class TransposeConvBlock(nn.Module): """ A Transpose Convolutional Block that consists of one convolution transpose layers followed by instance normalization and LeakyReLU activation. """ def __init__(self, in_chans: int, out_chans: int): """ Args: in_chans: Number of channels in the input. out_chans: Number of channels in the output. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.layers = nn.Sequential( nn.ConvTranspose2d( in_chans, out_chans, kernel_size=2, stride=2, bias=False ), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H*2, W*2)`. """ return self.layers(image)
7,138
34.517413
131
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noise2Noise/models/__init__.py
from .unet_fastMRI import *
27
27
27
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noise2Noise/utils/progress_bar.py
from collections import OrderedDict from numbers import Number from tqdm import tqdm from .meters import AverageMeter, RunningAverageMeter, TimeMeter class ProgressBar: def __init__(self, iterable, epoch=None, prefix=None, quiet=False): self.epoch = epoch self.quiet = quiet self.prefix = prefix + ' | ' if prefix is not None else '' if epoch is not None: self.prefix += f"epoch {epoch:02d}" self.iterable = iterable if self.quiet else tqdm(iterable, self.prefix, leave=False) def __iter__(self): return iter(self.iterable) def log(self, stats, verbose=False): if not self.quiet: self.iterable.set_postfix(self.format_stats(stats, verbose), refresh=True) def format_stats(self, stats, verbose=False): postfix = OrderedDict(stats) for key, value in postfix.items(): if isinstance(value, Number): fmt = "{:.3f}" if value > 0.001 else "{:.1e}" postfix[key] = fmt.format(value) elif isinstance(value, AverageMeter) or isinstance(value, RunningAverageMeter): if verbose: postfix[key] = f"{value.avg:.5f} ({value.val:.5f})" else: postfix[key] = f"{value.avg:.5f}" elif isinstance(value, TimeMeter): postfix[key] = f"{value.elapsed_time:.1f}s" elif not isinstance(postfix[key], str): postfix[key] = str(value) return postfix def print(self, stats, verbose=False): postfix = " | ".join(key + " " + value.strip() for key, value in self.format_stats(stats, verbose).items()) return f"{self.prefix + ' | ' if self.epoch is not None else ''}{postfix}"
1,789
37.913043
115
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noise2Noise/utils/utils_image.py
import os import math import random import numpy as np import torch import cv2 from torchvision.utils import make_grid def modcrop(img, scale): # img_in: BCHW or CHW or HW #img = np.copy(img_in) if img.ndim == 2: H, W = img.shape H_r, W_r = H % scale, W % scale img = img[:H - H_r, :W - W_r] elif img.ndim == 3: C, H, W = img.shape H_r, W_r = H % scale, W % scale img = img[:, :H - H_r, :W - W_r] elif img.ndim == 4: B, C, H, W = img.shape H_r, W_r = H % scale, W % scale img = img[:, :, :H - H_r, :W - W_r] else: raise ValueError('Wrong img ndim: [{:d}].'.format(img.ndim)) return img def modcrop_pil(image, modulo): w = image.width - image.width % modulo h = image.height - image.height % modulo return image.crop((0, 0, w, h)) def crop_center(pil_img, crop_width, crop_height): # Perform center crop on a PIL image img_width, img_height = pil_img.size return pil_img.crop(((img_width - crop_width) // 2, (img_height - crop_height) // 2, (img_width + crop_width) // 2, (img_height + crop_height) // 2)) ''' # -------------------------------------------- # matlab's bicubic imresize (numpy and torch) [0, 1] # -------------------------------------------- # from https://github.com/cszn/KAIR/blob/master/utils/utils_image.py ''' # matlab 'imresize' function, now only support 'bicubic' def cubic(x): absx = torch.abs(x) absx2 = absx**2 absx3 = absx**3 return (1.5*absx3 - 2.5*absx2 + 1) * ((absx <= 1).type_as(absx)) + \ (-0.5*absx3 + 2.5*absx2 - 4*absx + 2) * (((absx > 1)*(absx <= 2)).type_as(absx)) def calculate_weights_indices(in_length, out_length, scale, kernel, kernel_width, antialiasing): if (scale < 1) and (antialiasing): # Use a modified kernel to simultaneously interpolate and antialias- larger kernel width kernel_width = kernel_width / scale # Output-space coordinates x = torch.linspace(1, out_length, out_length) u = x / scale + 0.5 * (1 - 1 / scale) left = torch.floor(u - kernel_width / 2) P = math.ceil(kernel_width) + 2 # The indices of the input pixels involved in computing the k-th output # pixel are in row k of the indices matrix. indices = left.view(out_length, 1).expand(out_length, P) + torch.linspace(0, P - 1, P).view( 1, P).expand(out_length, P) # The weights used to compute the k-th output pixel are in row k of the # weights matrix. distance_to_center = u.view(out_length, 1).expand(out_length, P) - indices # apply cubic kernel if (scale < 1) and (antialiasing): weights = scale * cubic(distance_to_center * scale) else: weights = cubic(distance_to_center) # Normalize the weights matrix so that each row sums to 1. weights_sum = torch.sum(weights, 1).view(out_length, 1) weights = weights / weights_sum.expand(out_length, P) # If a column in weights is all zero, get rid of it. only consider the first and last column. weights_zero_tmp = torch.sum((weights == 0), 0) if not math.isclose(weights_zero_tmp[0], 0, rel_tol=1e-6): indices = indices.narrow(1, 1, P - 2) weights = weights.narrow(1, 1, P - 2) if not math.isclose(weights_zero_tmp[-1], 0, rel_tol=1e-6): indices = indices.narrow(1, 0, P - 2) weights = weights.narrow(1, 0, P - 2) weights = weights.contiguous() indices = indices.contiguous() sym_len_s = -indices.min() + 1 sym_len_e = indices.max() - in_length indices = indices + sym_len_s - 1 return weights, indices, int(sym_len_s), int(sym_len_e) # -------------------------------------------- # imresize for tensor image [0, 1] # -------------------------------------------- def imresize(img, scale, antialiasing=True): need_squeeze = True if img.dim() == 2 else False if need_squeeze: img.unsqueeze_(0) in_C, in_H, in_W = img.size() out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale) kernel_width = 4 kernel = 'cubic' # get weights and indices weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices( in_H, out_H, scale, kernel, kernel_width, antialiasing) weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices( in_W, out_W, scale, kernel, kernel_width, antialiasing) # process H dimension # symmetric copying img_aug = torch.FloatTensor(in_C, in_H + sym_len_Hs + sym_len_He, in_W) img_aug.narrow(1, sym_len_Hs, in_H).copy_(img) sym_patch = img[:, :sym_len_Hs, :] inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(1, inv_idx) img_aug.narrow(1, 0, sym_len_Hs).copy_(sym_patch_inv) sym_patch = img[:, -sym_len_He:, :] inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(1, inv_idx) img_aug.narrow(1, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv) out_1 = torch.FloatTensor(in_C, out_H, in_W) kernel_width = weights_H.size(1) for i in range(out_H): idx = int(indices_H[i][0]) for j in range(out_C): out_1[j, i, :] = img_aug[j, idx:idx + kernel_width, :].transpose(0, 1).mv(weights_H[i]) # process W dimension # symmetric copying out_1_aug = torch.FloatTensor(in_C, out_H, in_W + sym_len_Ws + sym_len_We) out_1_aug.narrow(2, sym_len_Ws, in_W).copy_(out_1) sym_patch = out_1[:, :, :sym_len_Ws] inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(2, inv_idx) out_1_aug.narrow(2, 0, sym_len_Ws).copy_(sym_patch_inv) sym_patch = out_1[:, :, -sym_len_We:] inv_idx = torch.arange(sym_patch.size(2) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(2, inv_idx) out_1_aug.narrow(2, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv) out_2 = torch.FloatTensor(in_C, out_H, out_W) kernel_width = weights_W.size(1) for i in range(out_W): idx = int(indices_W[i][0]) for j in range(out_C): out_2[j, :, i] = out_1_aug[j, :, idx:idx + kernel_width].mv(weights_W[i]) if need_squeeze: out_2.squeeze_() return out_2 # -------------------------------------------- # imresize for numpy image [0, 1] # -------------------------------------------- def imresize_np(img, scale, antialiasing=True): # Now the scale should be the same for H and W # input: img: Numpy, HWC or HW [0,1] # output: HWC or HW [0,1] w/o round img = torch.from_numpy(img) need_squeeze = True if img.dim() == 2 else False if need_squeeze: img.unsqueeze_(2) in_H, in_W, in_C = img.size() out_C, out_H, out_W = in_C, math.ceil(in_H * scale), math.ceil(in_W * scale) kernel_width = 4 kernel = 'cubic' # Return the desired dimension order for performing the resize. The # strategy is to perform the resize first along the dimension with the # smallest scale factor. # Now we do not support this. # get weights and indices weights_H, indices_H, sym_len_Hs, sym_len_He = calculate_weights_indices( in_H, out_H, scale, kernel, kernel_width, antialiasing) weights_W, indices_W, sym_len_Ws, sym_len_We = calculate_weights_indices( in_W, out_W, scale, kernel, kernel_width, antialiasing) # process H dimension # symmetric copying img_aug = torch.FloatTensor(in_H + sym_len_Hs + sym_len_He, in_W, in_C) img_aug.narrow(0, sym_len_Hs, in_H).copy_(img) sym_patch = img[:sym_len_Hs, :, :] inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(0, inv_idx) img_aug.narrow(0, 0, sym_len_Hs).copy_(sym_patch_inv) sym_patch = img[-sym_len_He:, :, :] inv_idx = torch.arange(sym_patch.size(0) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(0, inv_idx) img_aug.narrow(0, sym_len_Hs + in_H, sym_len_He).copy_(sym_patch_inv) out_1 = torch.FloatTensor(out_H, in_W, in_C) kernel_width = weights_H.size(1) for i in range(out_H): idx = int(indices_H[i][0]) for j in range(out_C): out_1[i, :, j] = img_aug[idx:idx + kernel_width, :, j].transpose(0, 1).mv(weights_H[i]) # process W dimension # symmetric copying out_1_aug = torch.FloatTensor(out_H, in_W + sym_len_Ws + sym_len_We, in_C) out_1_aug.narrow(1, sym_len_Ws, in_W).copy_(out_1) sym_patch = out_1[:, :sym_len_Ws, :] inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(1, inv_idx) out_1_aug.narrow(1, 0, sym_len_Ws).copy_(sym_patch_inv) sym_patch = out_1[:, -sym_len_We:, :] inv_idx = torch.arange(sym_patch.size(1) - 1, -1, -1).long() sym_patch_inv = sym_patch.index_select(1, inv_idx) out_1_aug.narrow(1, sym_len_Ws + in_W, sym_len_We).copy_(sym_patch_inv) out_2 = torch.FloatTensor(out_H, out_W, in_C) kernel_width = weights_W.size(1) for i in range(out_W): idx = int(indices_W[i][0]) for j in range(out_C): out_2[:, i, j] = out_1_aug[:, idx:idx + kernel_width, j].mv(weights_W[i]) if need_squeeze: out_2.squeeze_() return out_2.numpy()
9,403
36.466135
99
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noise2Noise/utils/train_utils.py
import argparse import os import logging import numpy as np import random import sys import torch from datetime import datetime from torch.serialization import default_restore_location def add_logging_arguments(parser): parser.add_argument("--seed", default=0, type=int, help="random number generator seed") parser.add_argument("--output-dir", default="experiments", help="path to experiment directories") parser.add_argument("--experiment", default=None, help="experiment name to be used with Tensorboard") parser.add_argument("--resume-training", action="store_true", help="whether to resume training") parser.add_argument("--restore-mode", default=None, help="Either 'best' 'last' or '\path\to\checkpoint\dir'") parser.add_argument("--restore-file", default=None, help="filename to load checkpoint") parser.add_argument("--test-mode", default=None, help="Evaluate on which test set.") parser.add_argument("--no-save", action="store_true", help="don't save models or checkpoints") parser.add_argument("--step-checkpoints", action="store_true", help="store all step checkpoints") parser.add_argument("--no-log", action="store_true", help="don't save logs to file or Tensorboard directory") parser.add_argument("--log-interval", type=int, default=100, help="log every N steps") parser.add_argument("--no-visual", action="store_true", help="don't use Tensorboard") parser.add_argument("--visual-interval", type=int, default=100, help="log every N steps") parser.add_argument("--no-progress", action="store_true", help="don't use progress bar") parser.add_argument("--draft", action="store_true", help="save experiment results to draft directory") parser.add_argument("--dry-run", action="store_true", help="no log, no save, no visualization") return parser def init_logging(args): for handler in logging.root.handlers[:]: logging.root.removeHandler(handler) handlers = [logging.StreamHandler()] if not args.no_log and args.log_file is not None: mode = "a" if args.resume_training else "w" handlers.append(logging.FileHandler(args.log_file, mode=mode)) logging.basicConfig(handlers=handlers, format="[%(asctime)s] %(message)s", datefmt="%Y-%m-%d %H:%M:%S", level=logging.INFO) logging.info("Arguments: {}".format(vars(args))) def setup_experiment(args): torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = True torch.manual_seed(args.seed) np.random.seed(args.seed) random.seed(args.seed) if args.dry_run: args.no_save = args.no_log = args.no_visual = True return args.experiment = args.experiment or f"{args.model.replace('_', '-')}" #unet args.experiment = "-".join([args.experiment, 'std'+str(args.noise_std)]) if not args.resume_training: args.experiment = "-".join([args.experiment, datetime.now().strftime("%b-%d-%H:%M:%S")]) args.experiment = "-".join([args.experiment, 'tr'+str(args.train_size)]) args.experiment_dir = os.path.join(args.output_dir, args.experiment) os.makedirs(args.experiment_dir, exist_ok=True) #dir is only created if it not already exists. If it already exists no error is raised if not args.no_save: args.checkpoint_dir = os.path.join(args.experiment_dir, "checkpoints") os.makedirs(args.checkpoint_dir, exist_ok=True) if not args.no_log: args.log_dir = os.path.join(args.experiment_dir, "logs") os.makedirs(args.log_dir, exist_ok=True) args.log_file = os.path.join(args.log_dir, "train.log") def save_checkpoint(args, step, epoch, model, optimizer=None, scheduler=None, score=None, mode="min"): assert mode == "min" or mode == "max" last_step = getattr(save_checkpoint, "last_step", -1) #-1 as default argument that is given if attribute does not exist save_checkpoint.last_step = max(last_step, step) default_score = float("inf") if mode == "min" else float("-inf") best_score = getattr(save_checkpoint, "best_score", default_score) if (score < best_score and mode == "min") or (score > best_score and mode == "max"): save_checkpoint.best_step = step save_checkpoint.best_epoch = epoch save_checkpoint.best_score = score if not args.no_save and step % args.save_interval == 0: os.makedirs(args.checkpoint_dir, exist_ok=True) model = [model] if model is not None and not isinstance(model, list) else model optimizer = [optimizer] if optimizer is not None and not isinstance(optimizer, list) else optimizer scheduler = [scheduler] if scheduler is not None and not isinstance(scheduler, list) else scheduler state_dict = { "step": step, "epoch": epoch, "score": score, "last_step": save_checkpoint.last_step, "best_step": save_checkpoint.best_step, "best_epoch": save_checkpoint.best_epoch, "best_score": getattr(save_checkpoint, "best_score", None), "model": [m.state_dict() for m in model] if model is not None else None, "optimizer": [o.state_dict() for o in optimizer] if optimizer is not None else None, "scheduler": [s.state_dict() for s in scheduler] if scheduler is not None else None, "args": argparse.Namespace(**{k: v for k, v in vars(args).items() if not callable(v)}), } if args.step_checkpoints: torch.save(state_dict, os.path.join(args.checkpoint_dir, "checkpoint{}.pt".format(step))) if (score < best_score and mode == "min") or (score > best_score and mode == "max"): torch.save(state_dict, os.path.join(args.checkpoint_dir, "checkpoint_best.pt")) if step > last_step: torch.save(state_dict, os.path.join(args.checkpoint_dir, "checkpoint_last.pt")) def load_checkpoint(args, model=None, optimizer=None, scheduler=None): if args.restore_file is not None and os.path.isfile(args.restore_file): print('restoring model..') state_dict = torch.load(args.restore_file, map_location=lambda s, l: default_restore_location(s, "cpu")) model = [model] if model is not None and not isinstance(model, list) else model optimizer = [optimizer] if optimizer is not None and not isinstance(optimizer, list) else optimizer scheduler = [scheduler] if scheduler is not None and not isinstance(scheduler, list) else scheduler if "best_score" in state_dict: save_checkpoint.best_score = state_dict["best_score"] save_checkpoint.best_step = state_dict["best_step"] if "last_step" in state_dict: save_checkpoint.last_step = state_dict["last_step"] if model is not None and state_dict.get("model", None) is not None: for m, state in zip(model, state_dict["model"]): m.load_state_dict(state) if optimizer is not None and state_dict.get("optimizer", None) is not None: for o, state in zip(optimizer, state_dict["optimizer"]): o.load_state_dict(state) if scheduler is not None and state_dict.get("scheduler", None) is not None: for s, state in zip(scheduler, state_dict["scheduler"]): milestones = s.milestones state['milestones'] = milestones s.load_state_dict(state) s.milestones = milestones logging.info("Loaded checkpoint {}".format(args.restore_file)) return state_dict
7,573
52.716312
138
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noise2Noise/utils/main_function_helpers.py
import torch import argparse import os import yaml import pathlib import pickle import logging import sys import time from torch.utils.tensorboard import SummaryWriter import torch.nn.functional as F import torchvision import glob from torch.serialization import default_restore_location from torch.utils.data import DataLoader import matplotlib.pyplot as plt import numpy as np from tensorboard.backend.event_processing import event_accumulator import utils import models from utils.data_helpers.load_datasets_helpers import * from utils.meters import * from utils.progress_bar import * from utils.noise_model import get_noise from utils.metrics import ssim,psnr from utils.util_calculate_psnr_ssim import calculate_psnr,calculate_ssim from utils.test_metrics import * def load_model(args): USE_CUDA = True device = torch.device('cuda') if (torch.cuda.is_available() and USE_CUDA) else torch.device('cpu') checkpoint_path = glob.glob(args.output_dir +'/unet*') if len(checkpoint_path) != 1: raise ValueError("There is either no or more than one model to load") checkpoint_path = pathlib.Path(checkpoint_path[0] + f"/checkpoints/checkpoint_{args.restore_mode}.pt") state_dict = torch.load(checkpoint_path, map_location=lambda s, l: default_restore_location(s, "cpu")) args = argparse.Namespace(**{ **vars(state_dict["args"]), "no_log": True}) #model = models.build_model(args).to(device) model = models.unet_fastMRI( in_chans=args.in_chans, chans = args.chans, num_pool_layers = args.num_pool_layers, drop_prob = 0.0, residual_connection = args.residual, ).to(device) model.load_state_dict(state_dict["model"][0]) model.eval() return model def cli_main_test(args): USE_CUDA = True device = torch.device('cuda') if (torch.cuda.is_available() and USE_CUDA) else torch.device('cpu') model = load_model(args) # evaluate test performance over following noise range noise_std_range = np.linspace(args.test_noise_std_min, args.test_noise_std_max, ((args.test_noise_std_max-args.test_noise_std_min)//args.test_noise_stepsize)+1,dtype=int)/255. metrics_path = os.path.join(args.output_dir, args.test_mode + '_' + str(args.test_noise_std_min)+'-'+str(args.test_noise_std_max)+f'_metrics_{args.restore_mode}.p') metrics_dict = metrics_avg_on_noise_range(model, args, noise_std_range, device = device) pickle.dump( metrics_dict, open(metrics_path, "wb" ) ) def cli_main(args): available_models = glob.glob(f'{args.output_dir}/*') if not args.resume_training and available_models: raise ValueError('There exists already a trained model and resume_training is set False') if args.resume_training: f_restore_file(args) # reset the attributes of the function save_checkpoint mode = "max" default_score = float("inf") if mode == "min" else float("-inf") utils.save_checkpoint.best_score = default_score utils.save_checkpoint.best_step = -1 utils.save_checkpoint.best_epoch = -1 utils.save_checkpoint.last_step = -1 utils.save_checkpoint.current_lr = args.lr device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') # Set the name of the directory for saving results utils.setup_experiment(args) utils.init_logging(args) # Build data loaders, a model and an optimizer model = models.unet_fastMRI( in_chans=args.in_chans, chans = args.chans, num_pool_layers = args.num_pool_layers, drop_prob = 0.0, residual_connection = args.residual, ).to(device) optimizer = torch.optim.Adam(model.parameters(), lr=args.lr) #scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, milestones=[50, 60, 70, 80, 90, 100], gamma=0.5) scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau( optimizer, mode='max', factor=args.lr_gamma, patience=args.lr_patience, threshold=args.lr_threshold, threshold_mode='abs', cooldown=0, min_lr=args.lr_min, eps=1e-08, verbose=True ) logging.info(f"Built a model consisting of {sum(p.numel() for p in model.parameters()):,} parameters") trainset = ImagenetSubdataset(args.train_size,args.path_to_ImageNet_train,mode='train',patch_size=args.patch_size,val_crop=args.val_crop) train_loader = DataLoader(trainset, batch_size=args.batch_size, shuffle=True, num_workers=8, pin_memory=True,generator=torch.Generator().manual_seed(args.seed)) valset = ImagenetSubdataset(args.val_size,args.path_to_ImageNet_train,mode='val',patch_size=args.patch_size,val_crop=args.val_crop) val_loader = DataLoader(valset, batch_size=1, shuffle=False, num_workers=4, pin_memory=True,generator=torch.Generator().manual_seed(args.seed)) print(optimizer.param_groups[0]["lr"]) if args.resume_training: state_dict = utils.load_checkpoint(args, model, optimizer, scheduler) global_step = state_dict['last_step'] start_epoch = int(state_dict['last_step']/(len(train_loader)))+1 start_decay = True elif args.no_annealing: global_step = -1 start_epoch = 0 start_decay = True else: global_step = -1 start_epoch = 0 start_decay = False print(optimizer.param_groups[0]["lr"]) args.log_interval = min(len(trainset), 100) # len(train_loader)=log once per epoch args.no_visual = False # True for not logging to tensorboard # Track moving average of loss values train_meters = { "train_loss":RunningAverageMeter(0.98)} valid_meters = {name: AverageMeter() for name in (["valid_psnr", "valid_ssim", "valid_psnr_self_supervised", "valid_ssim_self_supervised"])} # Create tensorflow event file writer = SummaryWriter(log_dir=args.experiment_dir) if not args.no_visual else None break_counter = 0 # store the best val performance from lr-interval before the last lr decay best_val_last = 0 # track the best val performance for the current lr-inerval best_val_current = 0 # count for how many lr intervals there was no improvement and break only if there was no improvement for 2 lr_interval_counter = 0 # if best_val_current at the end of the current lr interval is smaller than best_val_last we perform early stopping for epoch in range(start_epoch, args.num_epochs): start = time.process_time() train_bar = ProgressBar(train_loader, epoch) # At beginning of each epoch reset the train meters for meter in train_meters.values(): meter.reset() for inputs, noise_seed in train_bar: model.train() #Sets the module in training mode. global_step += 1 inputs = inputs.to(device) noise = get_noise(inputs,noise_seed, fix_noise = args.fix_noise, noise_std = args.noise_std/255.) noise_target = get_noise(inputs,torch.mul(noise_seed,10),fix_noise = args.fix_noise, noise_std = args.noise_std_target/255.) noisy_targets = noise_target + inputs noisy_inputs = noise + inputs outputs = model(noisy_inputs) # In loss function, I changed outputs to noisy_targets for self-supervision loss = F.mse_loss(outputs, noisy_targets, reduction="sum") / torch.prod(torch.tensor(inputs.size())) #(inputs.size(0) * 2) model.zero_grad() loss.backward() optimizer.step() train_meters["train_loss"].update(loss.item()) train_bar.log(dict(**train_meters, lr=optimizer.param_groups[0]["lr"]), verbose=True) # Add to tensorflow event file: if writer is not None: writer.add_scalar("lr", optimizer.param_groups[0]["lr"], global_step) writer.add_scalar("loss/train", train_meters["train_loss"].avg, global_step) sys.stdout.flush() if epoch % args.valid_interval == 0: model.eval() gen_val = torch.Generator() gen_val = gen_val.manual_seed(10) for meter in valid_meters.values(): meter.reset() valid_bar = ProgressBar(val_loader) for sample, noise_seed in valid_bar: with torch.no_grad(): sample = sample.to(device) # Self-supervised validation with fixed noise noise_self_supervised = get_noise(sample,noise_seed, fix_noise = args.fix_noise, noise_std = args.noise_std/255.) noise_target_self_supervised = get_noise(sample,torch.mul(noise_seed,10),fix_noise = args.fix_noise, noise_std = args.noise_std_target/255.) noisy_input_fixed = sample + noise_self_supervised noisy_target = sample + noise_target_self_supervised output_self_supervised = model(noisy_input_fixed) valid_psnr_self_supervised = psnr(output_self_supervised, noisy_target) valid_ssim_self_supervised = ssim(output_self_supervised, noisy_target) valid_meters["valid_psnr_self_supervised"].update(valid_psnr_self_supervised.item()) valid_meters["valid_ssim_self_supervised"].update(valid_ssim_self_supervised.item()) # Ground truth validation wit fixed noise # It uses the same input and output as in the self-supervised case since the noise seed is fixed valid_psnr = psnr(output_self_supervised, sample) valid_ssim = ssim(output_self_supervised, sample) valid_meters["valid_psnr"].update(valid_psnr.item()) valid_meters["valid_ssim"].update(valid_ssim.item()) if writer is not None: # Average is correct valid_meters['valid_psnr'].avg since .val would be just the psnr of last sample in val set. writer.add_scalar("psnr/valid", valid_meters['valid_psnr'].avg, global_step) writer.add_scalar("ssim/valid", valid_meters['valid_ssim'].avg, global_step) writer.add_scalar("psnr_selfsupervised/valid", valid_meters['valid_psnr_self_supervised'].avg, global_step) writer.add_scalar("ssim_selfsupervised/valid", valid_meters["valid_ssim_self_supervised"].avg, global_step) writer.add_scalar("lr", optimizer.param_groups[0]["lr"], global_step) sys.stdout.flush() if args.val_flag == 0: # if we do self-supervised validation val_loss = valid_meters["valid_psnr_self_supervised"].avg else: # if we do supervised validation val_loss = valid_meters["valid_psnr"].avg if utils.save_checkpoint.best_score < val_loss and not start_decay: utils.save_checkpoint(args, global_step, epoch, model, optimizer, score=val_loss, mode="max") current_lr = utils.save_checkpoint.current_lr optimizer.param_groups[0]["lr"] = current_lr*args.lr_beta utils.save_checkpoint.current_lr = current_lr*args.lr_beta annealing_counter = 0 elif not start_decay: annealing_counter += 1 current_lr = utils.save_checkpoint.current_lr if annealing_counter == args.lr_patience_annealing: available_models = glob.glob(f'{args.output_dir}/*') if not available_models: raise ValueError('No file to restore') elif len(available_models)>1: raise ValueError('Too many files to restore from') model_path = os.path.join(available_models[0], "checkpoints/checkpoint_best.pt") state_dict = torch.load(model_path, map_location=lambda s, l: default_restore_location(s, "cpu")) model = [model] if model is not None and not isinstance(model, list) else model for m, state in zip(model, state_dict["model"]): m.load_state_dict(state) model = model[0] optimizer.param_groups[0]["lr"] = current_lr/(args.lr_beta*args.inital_decay_factor) start_decay = True scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau( optimizer, mode='max', factor=args.lr_gamma, patience=args.lr_patience, threshold=args.lr_threshold, threshold_mode='abs', cooldown=0, min_lr=args.lr_min, eps=1e-08, verbose=True ) else: utils.save_checkpoint(args, global_step, epoch, model, optimizer, score=val_loss, mode="max") current_lr = optimizer.param_groups[0]["lr"] if val_loss > best_val_current: best_val_current = val_loss if writer is not None: writer.add_scalar("epoch", epoch, global_step) sys.stdout.flush() if start_decay: current_lr = optimizer.param_groups[0]["lr"] scheduler.step(val_loss) new_lr = optimizer.param_groups[0]["lr"] #At every lr decay check if the model did not improve during the current or the previous lr interval and break if it didn't. if new_lr < current_lr: if best_val_current < best_val_last and lr_interval_counter==1: logging.info('Break training due to convergence of val loss!') break elif best_val_current < best_val_last and lr_interval_counter==0: lr_interval_counter += 1 logging.info('Do not yet break due to convergence of val loss!') else: best_val_last = best_val_current best_val_current = 0 lr_interval_counter = 0 end = time.process_time() - start logging.info(train_bar.print(dict(**train_meters, **valid_meters, lr=current_lr, time=np.round(end/60,3)))) if optimizer.param_groups[0]["lr"] == args.lr_min and start_decay: break_counter += 1 if break_counter == args.break_counter: print('Break training due to minimal learning rate constraint!') break logging.info(f"Done training! Best PSNR {utils.save_checkpoint.best_score:.3f} obtained after step {utils.save_checkpoint.best_step} (epoch {utils.save_checkpoint.best_epoch}).") def get_args(hp,ee,rr): parser = argparse.ArgumentParser(allow_abbrev=False) # Add data arguments parser.add_argument("--train-size", default=None, help="number of examples in training set") parser.add_argument("--val-size", default=40, help="number of examples in validation set") parser.add_argument("--test-size", default=100, help="number of examples in test set") parser.add_argument("--val-crop", default=True, type=bool, help="Crop validation images to train size.") parser.add_argument("--patch-size", default=128, help="size of the center cropped HR image") parser.add_argument("--batch-size", default=128, type=int, help="train batch size") # Add model arguments parser.add_argument("--model", default="unet", help="model architecture") # Add noise arguments parser.add_argument('--noise_std', default = 15, type = float, help = 'noise level') parser.add_argument('--test_noise_std_min', default = 15, type = float, help = 'minimal noise level for testing') parser.add_argument('--test_noise_std_max', default = 15, type = float, help = 'maximal noise level for testing') parser.add_argument('--test_noise_stepsize', default = 5, type = float, help = 'Stepsize between test_noise_std_min and test_noise_std_max') # Add optimization arguments parser.add_argument("--lr", default=1e-3, type=float, help="learning rate") parser.add_argument("--lr-gamma", default=0.5, type=float, help="factor by which to reduce learning rate") parser.add_argument("--lr-beta", default=2, type=float, help="factor by which to increase learning rate") parser.add_argument("--lr-patience", default=5, type=int, help="epochs without improvement before lr decay") parser.add_argument("--no_annealing", default=True, type=bool, help="Use lr annealing or not.") parser.add_argument("--lr-patience-annealing", default=3, type=int, help="epochs without improvement before lr annealing stops") parser.add_argument("--lr-min", default=1e-5, type=float, help="Once we reach this learning rate continue for break_counter many epochs then stop.") parser.add_argument("--lr-threshold", default=0.003, type=float, help="Improvements by less than this threshold are not counted for decay patience.") parser.add_argument("--break-counter", default=9, type=int, help="Once smallest learning rate is reached, continue for so many epochs before stopping.") parser.add_argument("--inital-decay-factor", default=2, type=int, help="After annealing found a lr for which val loss does not improve, go back initial_decay_factor many lrs") parser.add_argument("--num-epochs", default=100, type=int, help="force stop training at specified epoch") parser.add_argument("--valid-interval", default=1, type=int, help="evaluate every N epochs") parser.add_argument("--save-interval", default=1, type=int, help="save a checkpoint every N steps") # Add model arguments parser = models.unet_fastMRI.add_args(parser) parser = utils.add_logging_arguments(parser) #args = parser.parse_args() args, _ = parser.parse_known_args() # Set arguments specific for this experiment dargs = vars(args) for key in hp.keys(): dargs[key] = hp[key][ee] args.seed = int(42 + 10*rr) return args def f_restore_file(args): #available_models = glob.glob(f'{args.output_dir}/{args.experiment}-*') available_models = glob.glob(f'{args.output_dir}/*') if not available_models: raise ValueError('No file to restore') if not args.restore_mode: raise ValueError("Pick restore mode either 'best' 'last' or '\path\to\checkpoint\dir'") if args.restore_mode=='best': mode = "max" best_score = float("inf") if mode == "min" else float("-inf") best_model = None for modelp in available_models: model_path = os.path.join(modelp, "checkpoints/checkpoint_best.pt") if os.path.isfile(model_path): state_dict = torch.load(model_path, map_location=lambda s, l: default_restore_location(s, "cpu")) score = state_dict["best_score"] if (score < best_score and mode == "min") or (score > best_score and mode == "max"): best_score = score best_model = model_path best_modelp = modelp best_step = state_dict["best_step"] best_epoch = state_dict["best_epoch"] args.restore_file = best_model args.experiment_dir = best_modelp #logging.info(f"Prepare to restore best model {best_model} with PSNR {best_score} at step {best_step}, epoch {best_epoch}") elif args.restore_mode=='last': last_step = -1 last_model = None for modelp in available_models: model_path = os.path.join(modelp, "checkpoints/checkpoint_last.pt") if os.path.isfile(model_path): state_dict = torch.load(model_path, map_location=lambda s, l: default_restore_location(s, "cpu")) step = state_dict["last_step"] if step > last_step: last_step = step last_model = model_path last_modelp = modelp score = state_dict["score"] last_epoch = state_dict["epoch"] args.restore_file = last_model args.experiment_dir = last_modelp #logging.info(f"Prepare to restore last model {last_model} with PSNR {score} at step {last_step}, epoch {last_epoch}") else: args.restore_file = args.restore_mode args.experiment_dir = args.restore_mode[:args.restore_mode.find('/checkpoints')] def infer_images(args): USE_CUDA = True device = torch.device('cuda') if (torch.cuda.is_available() and USE_CUDA) else torch.device('cpu') net = load_model(args) # the denoiser seed_dict = { "val":10, "test":20, "cbsd68":30, "urban100":40, "mcmaster18":50, "kodak24":60, "CBSD68":70, } gen = torch.Generator() gen = gen.manual_seed(seed_dict[args.test_mode]) # Load the test images load_path = '../training_set_lists/' if args.test_mode == 'test': files_source = torch.load(load_path+f'ImageNetTest{args.test_size}_filepaths.pt') #files_source.sort() elif args.test_mode == 'val': files_source = torch.load(load_path+f'ImageNetVal{args.val_size}_filepaths.pt') #files_source.sort() else: files_source = torch.load(load_path+f'{args.test_mode}_filepaths.pt') if not os.path.isdir(args.output_dir+'/test_images'): os.mkdir(args.output_dir+'/test_images') counter = 0 transformT = transforms.ToTensor() transformIm = transforms.ToPILImage() for f in files_source: counter = counter + 1 if counter > 3: break # Create noise ISource = torch.unsqueeze(transformT(Image.open(f).convert("RGB")),0).to(device) noise = torch.randn(ISource.shape,generator = gen) * args.noise_std/255. INoisy = noise.to(device) + ISource out = torch.clamp(net(INoisy), 0., 1.).cpu() out = torch.squeeze(out,0) # Get rid of the 1 in dim 0. im = transformIm(out) INoisy = torch.clamp(torch.squeeze(INoisy,0), 0., 1.).cpu() #INoisy = torch.squeeze(INoisy,0).cpu() INoisy = transformIm(INoisy) clean_image = Image.open(f).convert("RGB") im.save(args.output_dir+f'/test_images/im{counter}_denoised_notclamped.png') clean_image.save(args.output_dir+f'/test_images/im{counter}_ground_truth_notclamped.png') INoisy.save(args.output_dir+f'/test_images/im{counter}_noisy_notclamped.png') im.save(args.output_dir+f'/test_images/im{counter}_denoised_notclamped.pdf') clean_image.save(args.output_dir+f'/test_images/im{counter}_ground_truth_notclamped.pdf') INoisy.save(args.output_dir+f'/test_images/im{counter}_noisy_notclamped.pdf')
23,245
45.772636
182
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noise2Noise/utils/util_calculate_psnr_ssim.py
import cv2 import numpy as np import torch # from https://github.com/JingyunLiang/SwinIR/blob/328dda0f4768772e6d8c5aa3d5aa8e24f1ad903b/utils/util_calculate_psnr_ssim.py#L80 def calculate_psnr(img1, img2, crop_border, input_order='HWC', test_y_channel=False): """Calculate PSNR (Peak Signal-to-Noise Ratio). Ref: https://en.wikipedia.org/wiki/Peak_signal-to-noise_ratio Args: img1 (ndarray): Images with range [0, 255]. img2 (ndarray): Images with range [0, 255]. crop_border (int): Cropped pixels in each edge of an image. These pixels are not involved in the PSNR calculation. input_order (str): Whether the input order is 'HWC' or 'CHW'. Default: 'HWC'. test_y_channel (bool): Test on Y channel of YCbCr. Default: False. Returns: float: psnr result. """ assert img1.shape == img2.shape, (f'Image shapes are differnet: {img1.shape}, {img2.shape}.') if input_order not in ['HWC', 'CHW']: raise ValueError(f'Wrong input_order {input_order}. Supported input_orders are ' '"HWC" and "CHW"') img1 = reorder_image(img1, input_order=input_order) img2 = reorder_image(img2, input_order=input_order) img1 = img1.astype(np.float64) img2 = img2.astype(np.float64) if crop_border != 0: img1 = img1[crop_border:-crop_border, crop_border:-crop_border, ...] img2 = img2[crop_border:-crop_border, crop_border:-crop_border, ...] if test_y_channel: img1 = to_y_channel(img1) img2 = to_y_channel(img2) mse = np.mean((img1 - img2) ** 2) if mse == 0: return float('inf') return 20. * np.log10(255. / np.sqrt(mse)) def _ssim(img1, img2): """Calculate SSIM (structural similarity) for one channel images. It is called by func:`calculate_ssim`. Args: img1 (ndarray): Images with range [0, 255] with order 'HWC'. img2 (ndarray): Images with range [0, 255] with order 'HWC'. Returns: float: ssim result. """ C1 = (0.01 * 255) ** 2 C2 = (0.03 * 255) ** 2 img1 = img1.astype(np.float64) img2 = img2.astype(np.float64) kernel = cv2.getGaussianKernel(11, 1.5) window = np.outer(kernel, kernel.transpose()) mu1 = cv2.filter2D(img1, -1, window)[5:-5, 5:-5] mu2 = cv2.filter2D(img2, -1, window)[5:-5, 5:-5] mu1_sq = mu1 ** 2 mu2_sq = mu2 ** 2 mu1_mu2 = mu1 * mu2 sigma1_sq = cv2.filter2D(img1 ** 2, -1, window)[5:-5, 5:-5] - mu1_sq sigma2_sq = cv2.filter2D(img2 ** 2, -1, window)[5:-5, 5:-5] - mu2_sq sigma12 = cv2.filter2D(img1 * img2, -1, window)[5:-5, 5:-5] - mu1_mu2 ssim_map = ((2 * mu1_mu2 + C1) * (2 * sigma12 + C2)) / ((mu1_sq + mu2_sq + C1) * (sigma1_sq + sigma2_sq + C2)) return ssim_map.mean() def calculate_ssim(img1, img2, crop_border, input_order='HWC', test_y_channel=False): """Calculate SSIM (structural similarity). Ref: Image quality assessment: From error visibility to structural similarity The results are the same as that of the official released MATLAB code in https://ece.uwaterloo.ca/~z70wang/research/ssim/. For three-channel images, SSIM is calculated for each channel and then averaged. Args: img1 (ndarray): Images with range [0, 255]. img2 (ndarray): Images with range [0, 255]. crop_border (int): Cropped pixels in each edge of an image. These pixels are not involved in the SSIM calculation. input_order (str): Whether the input order is 'HWC' or 'CHW'. Default: 'HWC'. test_y_channel (bool): Test on Y channel of YCbCr. Default: False. Returns: float: ssim result. """ assert img1.shape == img2.shape, (f'Image shapes are differnet: {img1.shape}, {img2.shape}.') if input_order not in ['HWC', 'CHW']: raise ValueError(f'Wrong input_order {input_order}. Supported input_orders are ' '"HWC" and "CHW"') img1 = reorder_image(img1, input_order=input_order) img2 = reorder_image(img2, input_order=input_order) img1 = img1.astype(np.float64) img2 = img2.astype(np.float64) if crop_border != 0: img1 = img1[crop_border:-crop_border, crop_border:-crop_border, ...] img2 = img2[crop_border:-crop_border, crop_border:-crop_border, ...] if test_y_channel: img1 = to_y_channel(img1) img2 = to_y_channel(img2) ssims = [] for i in range(img1.shape[2]): ssims.append(_ssim(img1[..., i], img2[..., i])) return np.array(ssims).mean() def reorder_image(img, input_order='HWC'): """Reorder images to 'HWC' order. If the input_order is (h, w), return (h, w, 1); If the input_order is (c, h, w), return (h, w, c); If the input_order is (h, w, c), return as it is. Args: img (ndarray): Input image. input_order (str): Whether the input order is 'HWC' or 'CHW'. If the input image shape is (h, w), input_order will not have effects. Default: 'HWC'. Returns: ndarray: reordered image. """ if input_order not in ['HWC', 'CHW']: raise ValueError(f'Wrong input_order {input_order}. Supported input_orders are ' "'HWC' and 'CHW'") if len(img.shape) == 2: img = img[..., None] if input_order == 'CHW': img = img.transpose(1, 2, 0) return img def to_y_channel(img): """Change to Y channel of YCbCr. Args: img (ndarray): Images with range [0, 255]. Returns: (ndarray): Images with range [0, 255] (float type) without round. """ img = img.astype(np.float32) / 255. if img.ndim == 3 and img.shape[2] == 3: img = bgr2ycbcr(img, y_only=True) img = img[..., None] return img * 255. def _convert_input_type_range(img): """Convert the type and range of the input image. It converts the input image to np.float32 type and range of [0, 1]. It is mainly used for pre-processing the input image in colorspace convertion functions such as rgb2ycbcr and ycbcr2rgb. Args: img (ndarray): The input image. It accepts: 1. np.uint8 type with range [0, 255]; 2. np.float32 type with range [0, 1]. Returns: (ndarray): The converted image with type of np.float32 and range of [0, 1]. """ img_type = img.dtype img = img.astype(np.float32) if img_type == np.float32: pass elif img_type == np.uint8: img /= 255. else: raise TypeError('The img type should be np.float32 or np.uint8, ' f'but got {img_type}') return img def _convert_output_type_range(img, dst_type): """Convert the type and range of the image according to dst_type. It converts the image to desired type and range. If `dst_type` is np.uint8, images will be converted to np.uint8 type with range [0, 255]. If `dst_type` is np.float32, it converts the image to np.float32 type with range [0, 1]. It is mainly used for post-processing images in colorspace convertion functions such as rgb2ycbcr and ycbcr2rgb. Args: img (ndarray): The image to be converted with np.float32 type and range [0, 255]. dst_type (np.uint8 | np.float32): If dst_type is np.uint8, it converts the image to np.uint8 type with range [0, 255]. If dst_type is np.float32, it converts the image to np.float32 type with range [0, 1]. Returns: (ndarray): The converted image with desired type and range. """ if dst_type not in (np.uint8, np.float32): raise TypeError('The dst_type should be np.float32 or np.uint8, ' f'but got {dst_type}') if dst_type == np.uint8: img = img.round() else: img /= 255. return img.astype(dst_type) def bgr2ycbcr(img, y_only=False): """Convert a BGR image to YCbCr image. The bgr version of rgb2ycbcr. It implements the ITU-R BT.601 conversion for standard-definition television. See more details in https://en.wikipedia.org/wiki/YCbCr#ITU-R_BT.601_conversion. It differs from a similar function in cv2.cvtColor: `BGR <-> YCrCb`. In OpenCV, it implements a JPEG conversion. See more details in https://en.wikipedia.org/wiki/YCbCr#JPEG_conversion. Args: img (ndarray): The input image. It accepts: 1. np.uint8 type with range [0, 255]; 2. np.float32 type with range [0, 1]. y_only (bool): Whether to only return Y channel. Default: False. Returns: ndarray: The converted YCbCr image. The output image has the same type and range as input image. """ img_type = img.dtype img = _convert_input_type_range(img) if y_only: out_img = np.dot(img, [24.966, 128.553, 65.481]) + 16.0 else: out_img = np.matmul( img, [[24.966, 112.0, -18.214], [128.553, -74.203, -93.786], [65.481, -37.797, 112.0]]) + [16, 128, 128] out_img = _convert_output_type_range(out_img, img_type) return out_img
9,023
37.564103
129
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noise2Noise/utils/metrics.py
import numpy as np from skimage.metrics import peak_signal_noise_ratio, structural_similarity def ssim(clean, noisy, normalized=True): """Use skimage.meamsure.compare_ssim to calculate SSIM Args: clean (Tensor): (B, C, H, W) noisy (Tensor): (B, C, H, W) normalized (bool): If True, the range of tensors are [0., 1.] else [0, 255] Returns: SSIM per image: (B, ) """ if len(clean.shape)!=4 or len(noisy.shape)!=4: raise ValueError("ssim expects clean (Tensor): (B, C, H, W) noisy (Tensor): (B, C, H, W)") if normalized: clean = clean.mul(255).clamp(0, 255) noisy = noisy.mul(255).clamp(0, 255) clean = clean.cpu().detach().numpy().astype(np.float32) noisy = noisy.cpu().detach().numpy().astype(np.float32) clean = np.moveaxis(clean,1,-1) noisy = np.moveaxis(noisy,1,-1) return np.array([structural_similarity(c, n, data_range=255, multichannel=True) for c, n in zip(clean, noisy)]).mean() def psnr(clean, noisy, normalized=True): """Use skimage.meamsure.compare_ssim to calculate SSIM Args: clean (Tensor): (B, C, H, W) noisy (Tensor): (B, C, H, W) normalized (bool): If True, the range of tensors are [0., 1.] else [0, 255] Returns: SSIM per image: (B, ) """ if len(clean.shape)!=4 or len(noisy.shape)!=4: raise ValueError("psnr expects clean (Tensor): (B, C, H, W) noisy (Tensor): (B, C, H, W)") if normalized: clean = clean.mul(255).clamp(0, 255) noisy = noisy.mul(255).clamp(0, 255) clean = clean.cpu().detach().numpy().astype(np.float32) noisy = noisy.cpu().detach().numpy().astype(np.float32) return np.array([peak_signal_noise_ratio(c, n, data_range=255) for c, n in zip(clean, noisy)]).mean()
1,811
36.75
122
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noise2Noise/utils/__init__.py
from .train_utils import * from .main_function_helpers import *
63
31
36
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noise2Noise/utils/test_metrics.py
import torch import numpy as np import matplotlib.pyplot as plt import glob import os #import cv2 from utils.noise_model import get_noise from utils.metrics import ssim,psnr from utils.util_calculate_psnr_ssim import calculate_psnr,calculate_ssim from skimage import color import PIL.Image as Image import torchvision.transforms as transforms from utils.utils_image import * metrics_key = ['psnr_m', 'psnr_s', 'psnr_delta_m', 'psnr_delta_s', 'ssim_m', 'ssim_s', 'ssim_delta_m', 'ssim_delta_s']; def tensor_to_image(torch_image, low=0.0, high = 1.0, clamp = True): if clamp: torch_image = torch.clamp(torch_image, low, high); return torch_image[0,0].cpu().data.numpy() def normalize(data): return data/255. def convert_dict_to_string(metrics): return_string = ''; for x in metrics.keys(): return_string += x+': '+str(round(metrics[x], 3))+' '; return return_string def get_all_comparison_metrics(denoised, source, noisy = None, scale=None, return_title_string = False, clamp = True): metrics = {}; metrics['psnr'] = np.zeros(len(denoised)) metrics['ssim'] = np.zeros(len(denoised)) if noisy is not None: metrics['psnr_delta'] = np.zeros(len(denoised)) metrics['ssim_delta'] = np.zeros(len(denoised)) if clamp: denoised = torch.clamp(denoised, 0.0, 1.0) metrics['psnr'] = psnr(source, denoised); metrics['ssim'] = ssim(source, denoised); if noisy is not None: metrics['psnr_delta'] = metrics['psnr'] - psnr(source, noisy); metrics['ssim_delta'] = metrics['ssim'] - ssim(source, noisy); if return_title_string: return convert_dict_to_string(metrics) else: return metrics def average_on_folder(args, net, noise_std, verbose=True, device = torch.device('cuda')): #if verbose: #print('Loading data info ...\n') print(f'\n Dataset: {args.test_mode}, Restore mode: {args.restore_mode}') load_path = '../training_set_lists/' seed_dict = { "val":10, "test":20, } gen = torch.Generator() gen = gen.manual_seed(seed_dict[args.test_mode]) if args.test_mode == 'test': files_source = torch.load(load_path+f'ImageNetTest{args.test_size}_filepaths.pt') #files_source.sort() elif args.test_mode == 'val': files_source = torch.load(load_path+f'ImageNetVal{args.val_size}_filepaths.pt') #files_source.sort() avreage_metrics_key = ['psnr', 'psnr_delta', 'ssim', 'ssim_delta'] avg_metrics = {}; for x in avreage_metrics_key: avg_metrics[x] = []; psnr_list = [] ssim_list = [] #print(files_source) for f in files_source: transformT = transforms.ToTensor() ISource = torch.unsqueeze(transformT(Image.open(args.path_to_ImageNet_train + f).convert("RGB")),0).to(device) if args.test_mode == 'val': noise_seed = int(f[f.find('train/')+17:-5].replace('_','')) gen = gen.manual_seed(noise_seed) noise = torch.randn(ISource.shape,generator = gen) * args.noise_std/255. INoisy = noise.to(device) + ISource out = torch.clamp(net(INoisy), 0., 1.) ind_metrics = get_all_comparison_metrics(out, ISource, INoisy, return_title_string = False) for x in avreage_metrics_key: avg_metrics[x].append(ind_metrics[x]) if(verbose): print("%s %s" % (f, convert_dict_to_string(ind_metrics))) metrics = {} for x in avreage_metrics_key: metrics[x+'_m'] = np.mean(avg_metrics[x]) metrics[x+'_s'] = np.std(avg_metrics[x]) if verbose: print("\n Average %s" % (convert_dict_to_string(metrics))) #if(not verbose): return metrics def metrics_avg_on_noise_range(net, args, noise_std_array, device = torch.device('cuda')): array_metrics = {} for x in metrics_key: array_metrics[x] = np.zeros(len(noise_std_array)) for j, noise_std in enumerate(noise_std_array): metric_list = average_on_folder(args, net, noise_std = noise_std, verbose=False, device=device); for x in metrics_key: array_metrics[x][j] += metric_list[x] print('noise: ', int(noise_std*255), ' ', x, ': ', str(array_metrics[x][j])) return array_metrics
4,464
29.582192
119
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noise2Noise/utils/noise_model.py
import torch def get_noise(data, noise_seed, fix_noise, noise_std = float(25)/255.0): if fix_noise: device = torch.device('cuda') gen = torch.Generator(device=device) batch_size = data.size(dim=0) tensor_dim = list(data.size())[1:] for i in range(0,batch_size): gen = gen.manual_seed(noise_seed[i].item()) noise = torch.randn(tensor_dim,generator = gen, device=device) * noise_std noise = torch.unsqueeze(noise,0) if i == 0: noise_tensor = noise else: noise_tensor = torch.cat((noise_tensor, noise),0) noise = noise_tensor #noise = torch.randn(data.shape,generator = gen, device=device) * noise_std else: noise = torch.randn_like(data) noise.data = noise.data * noise_std return noise
880
31.62963
87
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noise2Noise/utils/meters.py
import time import torch class AverageMeter(object): def __init__(self): self.reset() def reset(self): self.val = 0 self.avg = 0 self.sum = 0 self.count = 0 def update(self, val, n=1): if isinstance(val, torch.Tensor): val = val.item() self.val = val / n self.sum += val self.count += n self.avg = self.sum / self.count class RunningAverageMeter(object): def __init__(self, momentum=0.98): self.momentum = momentum self.reset() def reset(self): self.val = None self.avg = 0 def update(self, val): if isinstance(val, torch.Tensor): val = val.item() if self.val is None: self.avg = val else: self.avg = self.avg * self.momentum + val * (1 - self.momentum) self.val = val class TimeMeter(object): def __init__(self, init=0): self.reset(init) def reset(self, init=0): self.init = init self.start = time.time() self.n = 0 def update(self, val=1): self.n += val @property def avg(self): return self.n / self.elapsed_time @property def elapsed_time(self): return self.init + (time.time() - self.start)
1,321
20.322581
75
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/Image_denoising_figure2/Noise2Noise/utils/data_helpers/load_datasets_helpers.py
import os import os.path import numpy as np import h5py import torch import torchvision.transforms as transforms import PIL.Image as Image from utils.utils_image import * class ImagenetSubdataset(torch.utils.data.Dataset): def __init__(self, size, path_to_ImageNet_train, mode='train', patch_size='128', val_crop=True): super().__init__() load_path = '../training_set_lists/' self.path_to_ImageNet_train = path_to_ImageNet_train if mode=='train': self.files = torch.load(load_path+f'trsize{size}_filepaths.pt') self.transform = transforms.Compose([ transforms.CenterCrop(patch_size), transforms.ToTensor(), ]) elif mode=='val': self.files = torch.load(load_path+f'ImageNetVal{size}_filepaths.pt') #print(self.files) if val_crop: self.transform = transforms.Compose([ transforms.CenterCrop(patch_size), transforms.ToTensor(), ]) else: self.transform = transforms.Compose([ transforms.ToTensor(), ]) self.noise_seeds = {} for i, file in enumerate(self.files): key = file[file.find('train/')+16:-5] number = int(file[file.find('train/')+17:-5].replace('_','')) self.noise_seeds[key] = number def __len__(self): return len(self.files) def __getitem__(self, index): file = self.files[index] key = file[file.find('train/')+16:-5] noise_seed = self.noise_seeds[key] image = Image.open(self.path_to_ImageNet_train + self.files[index]).convert("RGB") #ImageNet contains some grayscale images data = self.transform(image) return data, noise_seed
1,949
32.62069
131
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/CS_natural_images_figure4/train_network_for_histogram.py
# %% import torch import h5py import numpy as np import os import yaml import logging import glob import random import pickle from typing import Callable, Dict, List, Optional, Sequence, Tuple, Union import matplotlib.pyplot as plt from torch.nn import MSELoss import copy from argparse import ArgumentParser from torch.utils.tensorboard import SummaryWriter import sys from tqdm import tqdm import torchvision.transforms as transforms import PIL.Image as Image from skimage.transform import resize from CS_natural_images_functions.unet import Unet from CS_natural_images_functions.fftc import fft2c, ifft2c from CS_natural_images_functions.losses import SSIMLoss from CS_natural_images_functions.progress_bar import ProgressBar, init_logging, AverageMeter, TrackMeter, TrackMeter_testing from CS_natural_images_functions.log_progress_helpers import save_figure, add_img_to_tensorboard, save_test_image_with_dc from CS_natural_images_functions.load_save_model_helpers import setup_experiment_or_load_checkpoint, save_checkpoint from CS_natural_images_functions.data_transforms import UnetDataTransform from CS_natural_images_functions.data_transforms import compute_number_of_lines_in_input_target_kspace # %% class CropDataset(torch.utils.data.Dataset): """ A PyTorch Dataset that provides access to MR image slices. In this case we are only interested in downsampled magnitude images. """ def __init__( self, dataset: List, path_to_ImageNet_train: str, transforms_list: List, experiment_path: str, img_size: int, ): """ Args: dataset: A list containing one entry for every slice in the dataset. Each entry is a dictionary with keys 'path','slice','filename'ArithmeticError path_to_mridata: Path to fastMRI data on the server. transform: Function that transforms the ground truth image x into training input and target. """ self.transform_30 = transforms_list[0] self.transform_35 = transforms_list[1] self.experiment_path = experiment_path device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') # Store downsampled ground truth training images here self.examples = [] load_transform = transforms.Compose([ transforms.CenterCrop(img_size), transforms.ToTensor(), ]) # Load mri magnitude images, downsample and store for datapath in dataset: image = Image.open(path_to_ImageNet_train+datapath).convert("L") filename = datapath[16:-5] self.examples.append((load_transform(image)[0].to(device),filename)) def __len__(self): return len(self.examples) def __getitem__(self, i: int): # Determine input, target and ground truth x,filename = self.examples[i] y_input, x_input, y_target, x_target, x_gt, input_mask, target_mask, mean, std, fname = self.transform_30(x,filename,i) _, _, _, _, _, _, target_mask_35, _, _, _ = self.transform_35(x,filename,i) return y_input, x_input, y_target, x_target, x_gt, input_mask, target_mask, mean, std, fname,target_mask_35 # %% def read_args(): parser = ArgumentParser() parser.add_argument( '--path_to_ImageNet_train', type=str, help='Path to ImageNet train directory.', required=True ) parser.add_argument( '--training', default=True, action='store_false', help='Add this flag to disable training.' ) parser.add_argument( '--testing', default=False, action='store_false', help='Add this flag to disable testing.' ) parser.add_argument( '--experiment_number', default='300', type=str, help='Set consecutive numbering for the experiments.' ) parser.add_argument( '--gpu', choices=(0, 1, 2, 3), default=1, type=int, help='Pick one out of four gpus.' ) parser.add_argument( '--seed', default=0, type=int, help='Set seed for network initialization.' ) parser.add_argument( '--trainset_size', choices=(50, 100, 250, 500, 1000, 2500, 5000, 10000, 25000, 50000, 100000, 250000, 500000, 1000000), default=50, type=int, help='Set training set size.' ) parser.add_argument( '--img_size', default=100, type=int, help='Set img_size for downsampling.' ) parser.add_argument( '--num_epochs', default=1000, type=int, help='Set number of training epochs.' ) parser.add_argument( '--acceleration', default=4.0, type=float, help='Undersampling of training and test inputs.' ) parser.add_argument( '--center_fraction', default=0.08, type=float, help='Fraction of lines that are always sample from the center (input and target). Set to 0.0 for sampling all lines randomly.' ) parser.add_argument( '--fix_split', default=True, action='store_true', help='Add this flag to set fix_split=True for fixed input target split for self-supervised and fixed input for supervised training.' ) args = parser.parse_args() exp_nums = ['992'] # validation every second epoch # lr decay on plateau hyperparameters = { #'acceleration_total' : [ #3.0, #], 'trainset_size' : [ 10000, ], 'center_fraction' : [ 0.08, ], 'seed' : [ 1, ], 'fix_split' : [ True, ], 'num_epochs' : [ 2, ], 'patience' : [ # as we validate every second epoch a patience of 10 actually means 20 epochs 15, ] } # Sanity checks for key in hyperparameters.keys(): if len(hyperparameters[key]) != len(exp_nums): print(key) raise ValueError("Specify hyperparameters for every experiment!") for i in range(len(exp_nums)): args.trainset_size = hyperparameters['trainset_size'][i] args.center_fraction = hyperparameters['center_fraction'][i] args.seed = hyperparameters['seed'][i] args.fix_split = hyperparameters['fix_split'][i] args.num_epochs = hyperparameters['num_epochs'][i] args.patience = hyperparameters['patience'][i] args.experiment_number = exp_nums[i] experiment_name = f"N{args.experiment_number}_t{args.trainset_size}_" experiment_name+="sup_VS_ss3035_" if args.center_fraction==0.0: experiment_name+="RandCenter_" else: experiment_name+="FixCenter_" experiment_name+="grad_diff_" experiment_name += f"run{args.seed}" experiment_path = experiment_name+"/" #dataset_path = f"../datasets/train_{args.trainset_size}_selfsup_slice.yaml" os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu) if args.training: run_training(experiment_path=experiment_path, acceleration=args.acceleration, center_fraction=args.center_fraction, seed=args.seed, img_size=args.img_size, fix_split=args.fix_split, num_epochs=args.num_epochs, patience=args.patience, trainset_size=args.trainset_size, path_to_ImageNet_train=args.path_to_ImageNet_train) ################################################################################################ def run_training(experiment_path, acceleration, center_fraction, seed, img_size, fix_split, num_epochs, patience, trainset_size, path_to_ImageNet_train): device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = False # Set seed torch.manual_seed(seed) torch.cuda.manual_seed(seed) np.random.seed(seed) random.seed(seed) # Create directory that holds train files if not os.path.isdir(experiment_path): os.mkdir(experiment_path) else: print(experiment_path) #raise ValueError("Experiment already exists!!") print("Warning: Experiment already exists!!") # Init train.log file init_logging(experiment_path) logging.info("Training...") # train loss function loss_fct = MSELoss(reduction='sum') #val_ssim_fct = SSIMLoss() # Init model model = Unet( in_chans=2, out_chans=2, chans=24, num_pool_layers=3, drop_prob=0.0,).to(device) # Init optimizer and scheduler optimizer = torch.optim.Adam( params=model.parameters(), lr=0.001, betas=(0.9, 0.999), eps=1e-08, weight_decay=0.0, amsgrad=False) # Load train set train_pool = torch.load('CS_natural_images_functions/training_set_lists/trsize1000000_filepaths.pt') zero_norm_files = ['train/n03729826/n03729826_6483.JPEG', 'train/n04515003/n04515003_24673.JPEG', 'train/n02111277/n02111277_12490.JPEG', 'train/n03888605/n03888605_9775.JPEG', 'train/n02992529/n02992529_3197.JPEG', 'train/n01930112/n01930112_18908.JPEG', 'train/n06874185/n06874185_3219.JPEG', 'train/n06785654/n06785654_17232.JPEG', 'train/n04033901/n04033901_29617.JPEG', 'train/n07920052/n07920052_14729.JPEG', 'train/n03729826/n03729826_40479.JPEG', 'train/n03729826/n03729826_10716.JPEG', 'train/n04286575/n04286575_74296.JPEG', 'train/n03937543/n03937543_10198.JPEG', 'train/n03063599/n03063599_3942.JPEG', 'train/n04152593/n04152593_13802.JPEG', 'train/n04522168/n04522168_24105.JPEG', 'train/n03532672/n03532672_78983.JPEG', 'train/n04404412/n04404412_12316.JPEG', 'train/n04330267/n04330267_18003.JPEG', 'train/n04118776/n04118776_37671.JPEG', 'train/n04591713/n04591713_3568.JPEG', 'train/n02437616/n02437616_12697.JPEG', 'train/n02799071/n02799071_54867.JPEG', 'train/n02883205/n02883205_26196.JPEG', 'train/n02667093/n02667093_2919.JPEG', 'train/n03196217/n03196217_1135.JPEG', 'train/n03196217/n03196217_3568.JPEG', 'train/n15075141/n15075141_19601.JPEG', 'train/n01943899/n01943899_24166.JPEG'] for zero_norm_file in zero_norm_files: train_pool.remove(zero_norm_file) rng_dataset = np.random.default_rng(seed) train_set = rng_dataset.choice(train_pool, size=trainset_size, replace=False, p=None) torch.save(train_set,experiment_path+'train_set.pt') # Train loader data_transform_train_35 = UnetDataTransform(acceleration=acceleration,acceleration_total=3.5, fix_split=fix_split, experiment_path=experiment_path,center_fraction=center_fraction) data_transform_train_30 = UnetDataTransform(acceleration=acceleration,acceleration_total=3.0, fix_split=fix_split, experiment_path=experiment_path,center_fraction=center_fraction) trainset = CropDataset(dataset=train_set, path_to_ImageNet_train=path_to_ImageNet_train, transforms_list=[data_transform_train_30,data_transform_train_35], experiment_path=experiment_path, img_size=img_size) train_loader = torch.utils.data.DataLoader(dataset=trainset, batch_size=1, num_workers=0, shuffle=True, generator=torch.Generator().manual_seed(0)) # store training loss metrics train_meters = {'train_L2': AverageMeter()} sup_diff_tracks = {'divide_by_norm_of_risk_grad': TrackMeter_testing(), 'take_mse': TrackMeter_testing()} ss_diff_30_tracks = {'divide_by_norm_of_risk_grad': TrackMeter_testing(), 'take_mse': TrackMeter_testing()} ss_diff_35_tracks = {'divide_by_norm_of_risk_grad': TrackMeter_testing(), 'take_mse': TrackMeter_testing()} # Init tensorboard #writer = SummaryWriter(log_dir=experiment_path) #log_image_interval_tb = 10 # when to compute gradient histograms compute_gradients_interval = 1 # Start training break_counter=0 for epoch in range(num_epochs): # compute gradient histogrms if epoch % compute_gradients_interval == 0: model_copy = copy.deepcopy(model) train_bar_hist = ProgressBar(train_loader, epoch) model_copy.train() for meter in sup_diff_tracks.values(): meter.reset() for meter in ss_diff_30_tracks.values(): meter.reset() for meter in ss_diff_35_tracks.values(): meter.reset() # estimate ground truth gradient based on whole dataset for id,sample in enumerate(train_bar_hist): y_input, x_input, y_target, x_target, x, input_mask, target_mask_30, mean, std, fname, target_mask_35 = sample # prediction x_output = torch.moveaxis(model_copy(torch.moveaxis( x_input , -1, 1 )), 1, -1) # unnormalize x_output = x_output * std + mean # move to kspace y_output_sup = fft2c(x_output) # apply target mask (all ones for supervised training) y_output_ss_30 = y_output_sup * target_mask_30 + 0.0 y_output_ss_35 = y_output_sup * target_mask_35 + 0.0 y_target_sup = fft2c(x) y_target_ss_30 = y_target_sup * target_mask_30 + 0.0 y_target_ss_35 = y_target_sup * target_mask_35 + 0.0 save_figure(torch.log(torch.abs(y_input[0,:,:,0].detach().cpu())+ 1e-9),"y_input_real",experiment_path) if id==0 else None save_figure(torch.log(torch.abs(y_output_sup[0,:,:,0].detach().cpu())+ 1e-9),"y_output_sup_real",experiment_path) if id==0 else None save_figure(torch.log(torch.abs(y_output_ss_30[0,:,:,0].detach().cpu())+ 1e-9),"y_output_ss_30_real",experiment_path) if id==0 else None save_figure(torch.log(torch.abs(y_output_ss_35[0,:,:,0].detach().cpu())+ 1e-9),"y_output_ss_35_real",experiment_path) if id==0 else None save_figure(torch.log(torch.abs(y_target_sup[0,:,:,0].detach().cpu())+ 1e-9),"y_target_sup_real",experiment_path) if id==0 else None save_figure(torch.log(torch.abs(y_target_ss_30[0,:,:,0].detach().cpu())+ 1e-9),"y_target_ss_30_real",experiment_path) if id==0 else None save_figure(torch.log(torch.abs(y_target_ss_35[0,:,:,0].detach().cpu())+ 1e-9),"y_target_ss_35_real",experiment_path) if id==0 else None # compute loss train_loss_sup = loss_fct(y_output_sup,y_target_sup) / torch.sum(torch.abs(y_target_sup)**2) #train_loss_ss_30 = loss_fct(y_output_ss_30,y_target_ss_30) / torch.sum(torch.abs(y_target_ss_30)**2) #train_loss_ss_35 = loss_fct(y_output_ss_35,y_target_ss_35) / torch.sum(torch.abs(y_target_ss_35)**2) param = list(model_copy.parameters()) model_copy.zero_grad() train_loss_sup.backward(retain_graph=True) if id == 0: for p in param: p.grad_true_risk = p.grad p.grad = None else: for p in param: p.grad_true_risk += p.grad p.grad = None for p in param: p.grad_true_risk = p.grad_true_risk/len(train_loader) # compute stochastic supervised and self-supervised gradients based on the same dataset train_bar_hist = ProgressBar(train_loader, epoch) for id,sample in enumerate(train_bar_hist): y_input, x_input, y_target, x_target, x, input_mask, target_mask_30, mean, std, fname, target_mask_35 = sample # prediction x_output = torch.moveaxis(model_copy(torch.moveaxis( x_input , -1, 1 )), 1, -1) # unnormalize x_output = x_output * std + mean # move to kspace y_output_sup = fft2c(x_output) # apply target mask (all ones for supervised training) y_output_ss_30 = y_output_sup * target_mask_30 + 0.0 y_output_ss_35 = y_output_sup * target_mask_35 + 0.0 y_target_sup = fft2c(x) y_target_ss_30 = y_target_sup * target_mask_30 + 0.0 y_target_ss_35 = y_target_sup * target_mask_35 + 0.0 # compute loss train_loss_sup = loss_fct(y_output_sup,y_target_sup) / torch.sum(torch.abs(y_target_sup)**2) train_loss_ss_30 = loss_fct(y_output_ss_30,y_target_ss_30) / torch.sum(torch.abs(y_target_ss_30)**2) train_loss_ss_35 = loss_fct(y_output_ss_35,y_target_ss_35) / torch.sum(torch.abs(y_target_ss_35)**2) param = list(model_copy.parameters()) model_copy.zero_grad() train_loss_sup.backward(retain_graph=True) for p in param: p.grad_sup = p.grad p.grad = None train_loss_ss_30.backward(retain_graph=True) for p in param: p.grad_ss_30 = p.grad p.grad = None train_loss_ss_35.backward(retain_graph=True) for p in param: p.grad_ss_35 = p.grad p.grad = None diff_sup = torch.zeros(1).to(device) diff_ss_30 = torch.zeros(1).to(device) diff_ss_35 = torch.zeros(1).to(device) norm_grad_of_risk = torch.zeros(1).to(device) for p in param: diff_sup += torch.sum(torch.square(torch.sub(p.grad_sup,p.grad_true_risk))) diff_ss_30 += torch.sum(torch.square(torch.sub(p.grad_ss_30,p.grad_true_risk))) diff_ss_35 += torch.sum(torch.square(torch.sub(p.grad_ss_35,p.grad_true_risk))) norm_grad_of_risk += torch.sum(torch.square(p.grad_true_risk)) sup_diff_tracks['divide_by_norm_of_risk_grad'].update(torch.div(diff_sup,norm_grad_of_risk).item()) sup_diff_tracks['take_mse'].update(torch.mean(diff_sup).item()) ss_diff_30_tracks['divide_by_norm_of_risk_grad'].update(torch.div(diff_ss_30,norm_grad_of_risk).item()) ss_diff_30_tracks['take_mse'].update(torch.mean(diff_ss_30).item()) ss_diff_35_tracks['divide_by_norm_of_risk_grad'].update(torch.div(diff_ss_35,norm_grad_of_risk).item()) ss_diff_35_tracks['take_mse'].update(torch.mean(diff_ss_35).item()) pickle.dump( sup_diff_tracks, open(experiment_path + f"sup_diff_tracks_ep{epoch}.pkl", "wb" ) , pickle.HIGHEST_PROTOCOL ) pickle.dump( ss_diff_30_tracks, open(experiment_path + f"ss_diff_30_tracks_ep{epoch}.pkl", "wb" ) , pickle.HIGHEST_PROTOCOL ) pickle.dump( ss_diff_35_tracks, open(experiment_path + f"ss_diff_35_tracks_ep{epoch}.pkl", "wb" ) , pickle.HIGHEST_PROTOCOL ) # perform one training epoch train_bar = ProgressBar(train_loader, epoch) for meter in train_meters.values(): meter.reset() for id,sample in enumerate(train_bar): model.train() y_input, x_input, y_target, x_target, x, input_mask, target_mask_30, mean, std, fname, target_mask_35 = sample # prediction x_output = torch.moveaxis(model(torch.moveaxis( x_input , -1, 1 )), 1, -1) # unnormalize x_output = x_output * std + mean # move to kspace y_output = fft2c(x_output) # apply target mask (all ones for supervised training) # DO SUPERVISED TRAINING HERE #y_output = y_output #* target_mask + 0.0 y_target = fft2c(x) # compute loss train_loss = loss_fct(y_output,y_target) / torch.sum(torch.abs(y_target)**2) model.zero_grad() train_loss.backward() optimizer.step() # log train metrics train_meters['train_L2'].update(train_loss.item()) train_bar.log(dict(**train_meters), verbose=True) ################################################################################################ if __name__ == '__main__': read_args() # %%
21,063
36.681574
211
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/CS_natural_images_figure4/run_CS_natural_images.py
# %% import torch import h5py import numpy as np import os import yaml import logging import glob import json import random import pickle from typing import Callable, Dict, List, Optional, Sequence, Tuple, Union import matplotlib.pyplot as plt from torch.nn import MSELoss from argparse import ArgumentParser from torch.utils.tensorboard import SummaryWriter import sys from CS_natural_images_functions.unet import Unet from CS_natural_images_functions.fftc import fft2c, ifft2c from CS_natural_images_functions.losses import SSIMLoss from CS_natural_images_functions.progress_bar import ProgressBar, init_logging, AverageMeter, TrackMeter, TrackMeter_testing from CS_natural_images_functions.log_progress_helpers import save_figure, add_img_to_tensorboard, save_test_image_with_dc from CS_natural_images_functions.load_save_model_helpers import setup_experiment_or_load_checkpoint, save_checkpoint from CS_natural_images_functions.data_transforms import UnetDataTransform, CropDataset from CS_natural_images_functions.data_transforms import compute_number_of_lines_in_input_target_kspace # %% def read_args(): parser = ArgumentParser() # Required arguments parser.add_argument( '--config_file', type=str, help='Name of a config file in the experiment_configs folder.', required=True ) parser.add_argument( '--path_to_ImageNet_train', type=str, help='Path to ImageNet train directory.', required=True ) parser.add_argument( '--experiment_number', type=str, help="Set a unique identifier for the folder containing the experimental results. Start number with '001'. ", required=True ) parser.add_argument( '--run_which_seeds', type=str, choices=('run_best_seed','run_all_seeds'), help='Choose to run either only the best seed or all seeds shown in our results.', required=True ) # Optional arguments parser.add_argument( '--training', default=True, action='store_false', help='Add this flag to disable training.' ) parser.add_argument( '--testing', default=True, action='store_false', help='Add this flag to disable testing.' ) parser.add_argument( '--gpu', choices=(0, 1, 2, 3), default=3, type=int, help='Pick one out of four gpus.' ) parser.add_argument( '--trainset_size', choices=(50, 100, 250, 500, 1000, 2500, 5000, 10000, 25000, 50000, 100000, 250000, 500000, 1000000), default=50, type=int, help='Set training set size.' ) parser.add_argument( '--img_size', default=100, type=int, help='Set img_size for downsampling.' ) parser.add_argument( '--num_epochs', default=1000, type=int, help='Set number of training epochs.' ) parser.add_argument( '--val_epoch_interval', default=2, type=int, help='Set how often the validation loss is computed.' ) parser.add_argument( '--patience', default=10, type=int, help='Patience parameter for the learning rate scheduler.' ) parser.add_argument( '--acceleration', default=4.0, type=float, help='Undersampling of training and test inputs.' ) parser.add_argument( '--acceleration_total', default=1.0, type=float, help='Undersampling of data available for input target split. Set to 1 for supervised training.' ) parser.add_argument( '--center_fraction', default=0.08, type=float, help='Fraction of lines that are always sample from the center (input and target). Set to 0.0 for sampling all lines randomly.' ) parser.add_argument( '--fix_split', default=True, action='store_true', help='Add this flag to set use_seed=True for fixed input target split for self-supervised and fixed input for supervised training.' ) args = parser.parse_args() with open("experiment_configs/"+args.config_file) as handle: config_file = json.load(handle) args.acceleration_total = config_file['acceleration_total'] args.acceleration = config_file['acceleration'] args.trainset_size = config_file['trainset_size'] args.val_epoch_interval = config_file['val_epoch_interval'] args.patience = config_file['patience'] if args.run_which_seeds == 'run_best_seed': seeds = [config_file['best_seed']] elif args.run_which_seeds == 'run_all_seeds': seeds = config_file['all_seeds'] for seed in seeds: experiment_name = f"N{args.experiment_number}_t{args.trainset_size}_" if args.acceleration_total==1.0: experiment_name+="sup_" else: experiment_name+="selfsup_" if args.fix_split: experiment_name+="fixInput_" else: experiment_name+="RandInput_" experiment_name += f"run{seed}" experiment_path = experiment_name+"/" os.environ['CUDA_VISIBLE_DEVICES'] = str(args.gpu) if args.training: run_training(experiment_path=experiment_path, acceleration=args.acceleration, center_fraction=args.center_fraction, acceleration_total=args.acceleration_total, seed=seed, img_size=args.img_size, fix_split=args.fix_split, val_epoch_interval=args.val_epoch_interval, num_epochs=args.num_epochs, patience=args.patience, trainset_size=args.trainset_size, path_to_ImageNet_train=args.path_to_ImageNet_train) if args.testing: run_testing(experiment_path=experiment_path, acceleration=args.acceleration, center_fraction=args.center_fraction, acceleration_total=args.acceleration_total, img_size=args.img_size, path_to_ImageNet_train=args.path_to_ImageNet_train) ################################################################################################ def run_training(experiment_path, acceleration, center_fraction, acceleration_total, seed, img_size, fix_split, val_epoch_interval, num_epochs, patience, trainset_size, path_to_ImageNet_train): device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = False # Set seed torch.manual_seed(seed) torch.cuda.manual_seed(seed) np.random.seed(seed) random.seed(seed) # Create directory that holds train files if not os.path.isdir(experiment_path): os.mkdir(experiment_path) else: print(experiment_path) #raise ValueError("Experiment already exists!!") print("Warning: Experiment already exists!!") # Init train.log file init_logging(experiment_path) logging.info("Training...") # Log sanity checks on the number of lines in the input/target kspaces input_size, target_size, overlap_size_high, size_low, p, q, mu, nu, weight_on_random_lines = compute_number_of_lines_in_input_target_kspace(p=1/acceleration,mu=1/acceleration_total,nu=center_fraction, n=img_size) logging.info(f"mu: {mu}, p: {p}, q: {q}, nu: {nu}, weight_on_random_lines: {weight_on_random_lines}") logging.info(f"\n Lines in kspace: {img_size} \n Lines in input: {input_size} \n Lines in target: {target_size} \n Number of high freq overlapping lines: {overlap_size_high} \n Number of low freq lines: {size_low}") # train loss function loss_fct = MSELoss(reduction='sum') val_ssim_fct = SSIMLoss() # Init model model = Unet( in_chans=2, out_chans=2, chans=24, num_pool_layers=3, drop_prob=0.0,).to(device) logging.info(f"Built a model consisting of {sum(p.numel() for p in model.parameters()):,} parameters")# # Init optimizer and scheduler optimizer = torch.optim.Adam( params=model.parameters(), lr=0.001, betas=(0.9, 0.999), eps=1e-08, weight_decay=0.0, amsgrad=False) scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau( optimizer=optimizer, mode='max', factor=0.1, patience=patience, threshold=0.0001, threshold_mode='abs', cooldown=0, min_lr=1e-5, eps=1e-08, verbose=True ) # If a checkpoint exists, it is automtically loaded setup_experiment_or_load_checkpoint(experiment_path, resume_from='best', model=model, optimizer=optimizer, scheduler=scheduler) # Load train set train_pool = torch.load('CS_natural_images_functions/training_set_lists/trsize1000000_filepaths.pt') zero_norm_files = ['train/n03729826/n03729826_6483.JPEG', 'train/n04515003/n04515003_24673.JPEG', 'train/n02111277/n02111277_12490.JPEG', 'train/n03888605/n03888605_9775.JPEG', 'train/n02992529/n02992529_3197.JPEG', 'train/n01930112/n01930112_18908.JPEG', 'train/n06874185/n06874185_3219.JPEG', 'train/n06785654/n06785654_17232.JPEG', 'train/n04033901/n04033901_29617.JPEG', 'train/n07920052/n07920052_14729.JPEG', 'train/n03729826/n03729826_40479.JPEG', 'train/n03729826/n03729826_10716.JPEG', 'train/n04286575/n04286575_74296.JPEG', 'train/n03937543/n03937543_10198.JPEG', 'train/n03063599/n03063599_3942.JPEG', 'train/n04152593/n04152593_13802.JPEG', 'train/n04522168/n04522168_24105.JPEG', 'train/n03532672/n03532672_78983.JPEG', 'train/n04404412/n04404412_12316.JPEG', 'train/n04330267/n04330267_18003.JPEG', 'train/n04118776/n04118776_37671.JPEG', 'train/n04591713/n04591713_3568.JPEG', 'train/n02437616/n02437616_12697.JPEG', 'train/n02799071/n02799071_54867.JPEG', 'train/n02883205/n02883205_26196.JPEG', 'train/n02667093/n02667093_2919.JPEG', 'train/n03196217/n03196217_1135.JPEG', 'train/n03196217/n03196217_3568.JPEG', 'train/n15075141/n15075141_19601.JPEG', 'train/n01943899/n01943899_24166.JPEG'] for zero_norm_file in zero_norm_files: train_pool.remove(zero_norm_file) rng_dataset = np.random.default_rng(seed) if trainset_size == 1000000: train_set = train_pool else: train_set = rng_dataset.choice(train_pool, size=trainset_size, replace=False, p=None) torch.save(train_set,experiment_path+'train_set.pt') validation_set = torch.load('CS_natural_images_functions/training_set_lists/ImageNetVal80_filepaths.pt') # Train loader data_transform_train = UnetDataTransform(acceleration=acceleration,acceleration_total=acceleration_total, fix_split=fix_split, experiment_path=experiment_path,center_fraction=center_fraction) trainset = CropDataset(dataset=train_set, path_to_ImageNet_train=path_to_ImageNet_train, transform=data_transform_train, experiment_path=experiment_path, img_size=img_size) train_loader = torch.utils.data.DataLoader(dataset=trainset, batch_size=1, num_workers=0, shuffle=True, generator=torch.Generator().manual_seed(0)) # Val loader data_transform_val = UnetDataTransform(acceleration=acceleration,acceleration_total=acceleration_total, fix_split=True, experiment_path=experiment_path,center_fraction=center_fraction) valset = CropDataset(dataset=validation_set, path_to_ImageNet_train=path_to_ImageNet_train, transform=data_transform_val, experiment_path=experiment_path, img_size=img_size) val_loader = torch.utils.data.DataLoader( dataset=valset, batch_size=1, num_workers=0, shuffle=False, generator=torch.Generator().manual_seed(0)) # store training loss metrics train_meters = {'train_L2': AverageMeter()} train_tracks = {'train_L2': TrackMeter('decaying')} # store validation metrics valid_meters = {'val_SSIM' : AverageMeter(), 'val_PSNR' : AverageMeter(), 'val_L2' : AverageMeter(), 'val_L2_kspace': AverageMeter()} valid_tracks = {'val_SSIM' : TrackMeter('increasing'), 'val_PSNR' : TrackMeter('increasing'), 'val_L2' : TrackMeter('decaying'), 'val_L2_kspace': TrackMeter('decaying')} # Init tensorboard writer = SummaryWriter(log_dir=experiment_path) log_image_interval_tb = 30 break_counter=0 # Start training for epoch in range(save_checkpoint.start_epoch, num_epochs): train_bar = ProgressBar(train_loader, epoch) for meter in train_meters.values(): meter.reset() for id,sample in enumerate(train_bar): model.train() y_input, x_input, y_target, x_target, x, input_mask, target_mask, mean, std, fname = sample # sanity check on number of lines in input and target mask if epoch==0 and id==0: tm = target_mask.detach() target_mask_no_zeros = torch.where(tm != 0., tm , torch.tensor(1, dtype=tm.dtype).to(device)) target_mask_norm_to_one = tm / target_mask_no_zeros logging.info(f"\n Mask sanity check! Lines in kspace: {input_mask.shape[-2]} \n Lines in input: {torch.sum(input_mask)} \n Lines in target: {torch.sum(target_mask_norm_to_one)} \n Number of all overlapping lines: {torch.sum(input_mask*target_mask_norm_to_one)}") # prediction x_output = torch.moveaxis(model(torch.moveaxis( x_input , -1, 1 )), 1, -1) # unnormalize x_output = x_output * std + mean # move to kspace y_output = fft2c(x_output) # apply target mask (all ones for supervised training) y_output = y_output * target_mask + 0.0 # compute loss train_loss = loss_fct(y_output,y_target) / torch.sum(torch.abs(y_target)**2) model.zero_grad() train_loss.backward() optimizer.step() # log train metrics train_meters['train_L2'].update(train_loss.item()) train_bar.log(dict(**train_meters), verbose=True) if id ==0: # log a random train image to tensorboard name = f"train_0_img" add_img_to_tensorboard(writer, epoch, name, x_input.detach(),x_output.detach(),x_target.detach(),ksp=False) if epoch % log_image_interval_tb == 0 else None name = f"train_0_ksp" add_img_to_tensorboard(writer, epoch, name, y_input.detach(),y_output.detach(),y_target.detach(),ksp=True) if epoch % log_image_interval_tb == 0 else None if id ==1: # log a specific train image to tensorboard name = f"train_1_img" add_img_to_tensorboard(writer, epoch, name, x_input.detach(),x_output.detach(),x_target.detach(),ksp=False) if epoch % log_image_interval_tb == 0 else None name = f"train_1_ksp" add_img_to_tensorboard(writer, epoch, name, y_input.detach(),y_output.detach(),y_target.detach(),ksp=True) if epoch % log_image_interval_tb == 0 else None train_tracks['train_L2'].update(train_meters['train_L2'].avg,epoch) current_lr = optimizer.param_groups[0]["lr"] #scheduler.step() ############################################################################################################################ if epoch % val_epoch_interval == 0: # set this value such that it works with save_at_epochs and log_image_interval_tb model.eval() for meter in valid_meters.values(): meter.reset() valid_bar = ProgressBar(val_loader, epoch) rand_id = random.randint(0, len(val_loader)) # draw id to log a random slice to tensorboard for id, sample in enumerate(valid_bar): with torch.no_grad(): y_input, x_input, y_target, x_target, x, input_mask, target_mask, mean, std, fname = sample # prediction x_output = torch.moveaxis(model(torch.moveaxis( x_input , -1, 1 )), 1, -1) # unnormalize x_output = x_output * std + mean # move to kspace y_output = fft2c(x_output) if id ==0: # log one fixed and one random validation image to tensorboard name = f"val_0_img" add_img_to_tensorboard(writer, epoch, name, x_input.detach(),x_output.detach(),x.detach(),ksp=False) if epoch % log_image_interval_tb == 0 else None name = f"val_0_ksp" add_img_to_tensorboard(writer, epoch, name, y_input.detach(),y_output.detach(),y_target.detach(),ksp=True) if epoch % log_image_interval_tb == 0 else None elif id==rand_id: # log one fixed and one random validation image to tensorboard name = f"val_1_img" add_img_to_tensorboard(writer, epoch, name, x_input.detach(),x_output.detach(),x.detach(),ksp=False) if epoch % log_image_interval_tb == 0 else None name = f"val_1_ksp" add_img_to_tensorboard(writer, epoch, name, y_input.detach(),y_output.detach(),y_target.detach(),ksp=True) if epoch % log_image_interval_tb == 0 else None # apply target mask (all ones for supervised training) y_output = y_output * target_mask + 0.0 # val loss in kspace (L2) val_loss = loss_fct(y_output,y_target) / torch.sum(torch.abs(y_target)**2) valid_meters['val_L2_kspace'].update(val_loss) # L2 in image domain between complex output and target image val_loss = loss_fct(x_output,x) / torch.sum(torch.abs(x)**2) valid_meters['val_L2'].update(val_loss) output_magnitude = (x_output ** 2).sum(dim=-1).sqrt() x_magnitude = (x ** 2).sum(dim=-1).sqrt() # since x is real, this operation is identity x_magnitude = x_magnitude.unsqueeze(1) output_magnitude = output_magnitude.unsqueeze(1) # psnr max_value = x.max().unsqueeze(0) mse = torch.mean(torch.abs(output_magnitude-x_magnitude)**2) psnr = 20 * torch.log10(torch.tensor(max_value.item()))- 10 * torch.log10(mse) valid_meters["val_PSNR"].update(psnr.item()) # ssim ssim_loss = 1-val_ssim_fct(output_magnitude, x_magnitude, data_range=max_value) valid_meters["val_SSIM"].update(ssim_loss.item()) # log progrss valid_tracks['val_L2_kspace'].update(valid_meters['val_L2_kspace'].avg,epoch) valid_tracks['val_L2'].update(valid_meters['val_L2'].avg,epoch) valid_tracks['val_PSNR'].update(valid_meters['val_PSNR'].avg,epoch) valid_tracks['val_SSIM'].update(valid_meters['val_SSIM'].avg,epoch) valid_bar.log(dict(**valid_meters), verbose=True) scheduler.step(valid_meters['val_PSNR'].avg) if current_lr > optimizer.param_groups[0]["lr"]: scheduler.patience= scheduler.patience//2 if current_lr == scheduler.min_lrs[0]: break_counter+=1 if break_counter == 3: break if save_checkpoint.best_score < valid_meters['val_PSNR'].avg: if scheduler.num_bad_epochs == 0: logging.info(train_bar.print(dict(**train_meters, **valid_meters, lr=current_lr, New='Highscore', Scheduler_patience='reset'))) else: logging.info(train_bar.print(dict(**train_meters, **valid_meters, lr=current_lr, New='Highscore'))) else: if scheduler.num_bad_epochs == 0: logging.info(train_bar.print(dict(**train_meters, **valid_meters, lr=current_lr, Scheduler_patience='reset'))) else: logging.info(train_bar.print(dict(**train_meters, **valid_meters, lr=current_lr))) writer.add_scalar("lr", current_lr, epoch) writer.add_scalar("epoch", epoch, epoch) writer.add_scalar("train_L2", train_meters["train_L2"].avg, epoch) for val_loss_name in valid_meters.keys(): writer.add_scalar(val_loss_name, valid_meters[val_loss_name].avg, epoch) sys.stdout.flush() # Save checkpoint save_checkpoint(experiment_path, epoch, model, optimizer=optimizer, scheduler=scheduler, score=valid_meters['val_PSNR'].avg, save_at_epochs=[]) else: logging.info(train_bar.print(dict(**train_meters, lr=current_lr))) writer.add_scalar("lr", current_lr, epoch) writer.add_scalar("epoch", epoch, epoch) writer.add_scalar("train_L2", train_meters["train_L2"].avg, epoch) sys.stdout.flush() logging.info(f"Done training! Best Val score {valid_tracks['val_PSNR'].best_val:.5f} obtained after epoch {valid_tracks['val_PSNR'].best_count}.") pickle.dump( valid_tracks, open(experiment_path + 'valid_tracks_metrics.pkl', "wb" ) , pickle.HIGHEST_PROTOCOL ) pickle.dump( train_tracks, open(experiment_path + 'train_tracks_metrics.pkl', "wb" ) , pickle.HIGHEST_PROTOCOL ) ################################################################################################ def run_testing(experiment_path, acceleration, center_fraction, acceleration_total, img_size, path_to_ImageNet_train): device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = False # Init train.log file init_logging(experiment_path) # Log sanity checks on the number of lines in the input/target kspaces logging.info("Testing...") input_size, target_size, overlap_size_high, size_low, p, q, mu, nu, weight_on_random_lines = compute_number_of_lines_in_input_target_kspace(p=1/acceleration,mu=1/acceleration_total,nu=center_fraction, n=img_size) logging.info(f"mu: {mu}, p: {p}, q: {q}, nu: {nu}, weight_on_random_lines: {weight_on_random_lines}") logging.info(f"\n Lines in kspace: {img_size} \n Lines in input: {input_size} \n Lines in target: {target_size} \n Number of high freq overlapping lines: {overlap_size_high} \n Number of low freq lines: {size_low}") # train loss function loss_fct = MSELoss(reduction='sum') val_ssim_fct = SSIMLoss() # Init model model = Unet( in_chans=2, out_chans=2, chans=24, num_pool_layers=3, drop_prob=0.0,).to(device) validation_set = torch.load('CS_natural_images_functions/training_set_lists/ImageNetVal80_filepaths.pt') test_set = torch.load('CS_natural_images_functions/training_set_lists/ImageNetTest300_filepaths.pt') # test loader data_transform_test = UnetDataTransform(acceleration=acceleration,acceleration_total=acceleration_total, fix_split=True, experiment_path=experiment_path,center_fraction=center_fraction) testset = CropDataset(dataset=test_set, path_to_ImageNet_train=path_to_ImageNet_train, transform=data_transform_test, experiment_path=experiment_path, img_size=img_size) test_loader = torch.utils.data.DataLoader( dataset=testset, batch_size=1, num_workers=0, shuffle=False, generator=torch.Generator().manual_seed(0), ) # Val loader data_transform_val = UnetDataTransform(acceleration=acceleration,acceleration_total=acceleration_total, fix_split=True, experiment_path=experiment_path,center_fraction=center_fraction) valset = CropDataset(dataset=validation_set, path_to_ImageNet_train=path_to_ImageNet_train, transform=data_transform_val, experiment_path=experiment_path, img_size=img_size) val_loader = torch.utils.data.DataLoader( dataset=valset, batch_size=1, num_workers=0, shuffle=False, generator=torch.Generator().manual_seed(0), ) setup_experiment_or_load_checkpoint(experiment_path, resume_from='best', model=model, optimizer=None, scheduler=None) test_validationSet_tracks = {'SSIM' : TrackMeter_testing(), 'PSNR' : TrackMeter_testing(), 'L2' : TrackMeter_testing(),'SSIM_dc' : TrackMeter_testing(), 'PSNR_dc' : TrackMeter_testing(), 'L2_dc' : TrackMeter_testing()} test_testSet_tracks = {'SSIM' : TrackMeter_testing(), 'PSNR' : TrackMeter_testing(), 'L2' : TrackMeter_testing(),'SSIM_dc' : TrackMeter_testing(), 'PSNR_dc' : TrackMeter_testing(), 'L2_dc' : TrackMeter_testing()} model.eval() tmp=0 for data_loader, track_meter in zip([val_loader, test_loader],[test_validationSet_tracks, test_testSet_tracks]): tmp+=1 test_bar = ProgressBar(data_loader, epoch=0) for id, sample in enumerate(test_bar): y_input, x_input, y_target, x_target, x, input_mask, target_mask, mean, std, fname = sample # prediction x_output = torch.moveaxis(model(torch.moveaxis( x_input , -1, 1 )), 1, -1) # unnormalize x_output = x_output * std + mean # Apply data consistency y_output = fft2c(x_output) y_output_dc = y_output * (1-input_mask) + y_input x_output_dc = ifft2c(y_output_dc) # L2 in image domain between complex output and target image val_loss = loss_fct(x_output,x) / torch.sum(torch.abs(x)**2) track_meter['L2'].update(val_loss) val_loss_dc = loss_fct(x_output_dc,x) / torch.sum(torch.abs(x)**2) track_meter['L2_dc'].update(val_loss_dc) output_magnitude = (x_output ** 2).sum(dim=-1).sqrt() output_dc_magnitude = (x_output_dc ** 2).sum(dim=-1).sqrt() x_magnitude = (x ** 2).sum(dim=-1).sqrt() # since x is real, this operation is identity x_magnitude = x_magnitude.unsqueeze(1) output_magnitude = output_magnitude.unsqueeze(1) output_dc_magnitude = output_dc_magnitude.unsqueeze(1) # psnr max_value = x.max().unsqueeze(0) mse = torch.mean(torch.abs(output_magnitude-x_magnitude)**2) psnr = 20 * torch.log10(torch.tensor(max_value.item()))- 10 * torch.log10(mse) track_meter["PSNR"].update(psnr.item()) mse_dc = torch.mean(torch.abs(output_dc_magnitude-x_magnitude)**2) psnr_dc = 20 * torch.log10(torch.tensor(max_value.item()))- 10 * torch.log10(mse_dc) track_meter["PSNR_dc"].update(psnr_dc.item()) # ssim ssim_loss = 1-val_ssim_fct(output_magnitude, x_magnitude, data_range=max_value) track_meter["SSIM"].update(ssim_loss.item()) ssim_loss_dc = 1-val_ssim_fct(output_dc_magnitude, x_magnitude, data_range=max_value) track_meter["SSIM_dc"].update(ssim_loss_dc.item()) # Save the first image in test set if (tmp==1 and id==1) or (tmp==1 and id==22): x_input_abs = (x_input ** 2).sum(dim=-1).sqrt() x_input_abs = x_input_abs.unsqueeze(1) save_test_image_with_dc(experiment_path, ground_truth_image=x_magnitude, input_img=x_input_abs, output=output_magnitude, output_image_dc=output_dc_magnitude, fname=fname, track_meter=track_meter) pickle.dump( test_validationSet_tracks, open(experiment_path + 'test_validationSet_metrics.pkl', "wb" ) , pickle.HIGHEST_PROTOCOL ) pickle.dump( test_testSet_tracks, open(experiment_path + 'test_testSet_metrics.pkl', "wb" ) , pickle.HIGHEST_PROTOCOL ) logging.info(f"\nEvaluate validationset of length {len(val_loader)}:") for metric in test_validationSet_tracks.keys(): logging.info(f"{metric}: avg {test_validationSet_tracks[metric].avg:.6f}, std {test_validationSet_tracks[metric].std:.6f}") logging.info(f"\nEvaluate testset of length {len(test_loader)}:") for metric in test_testSet_tracks.keys(): logging.info(f"{metric}: avg {test_testSet_tracks[metric].avg:.6f}, std {test_testSet_tracks[metric].std:.6f}") if __name__ == '__main__': read_args() # %%
29,149
44.404984
278
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/CS_natural_images_figure4/CS_natural_images_functions/progress_bar.py
from collections import OrderedDict from numbers import Number from tqdm import tqdm import torch import logging import os import numpy as np def init_logging(experiment_path): for handler in logging.root.handlers[:]: logging.root.removeHandler(handler) handlers = [logging.StreamHandler()] mode = "a" if os.path.exists(experiment_path+"train.log") else "w" handlers.append(logging.FileHandler(experiment_path+"train.log", mode=mode)) logging.basicConfig(handlers=handlers, format="[%(asctime)s] %(message)s", datefmt="%Y-%m-%d %H:%M:%S", level=logging.INFO) class AverageMeter(object): def __init__(self): self.reset() def reset(self): self.val = 0 self.avg = 0 self.sum = 0 self.count = 0 def update(self, val, n=1): if isinstance(val, torch.Tensor): val = val.item() self.val = val / n self.sum += val self.count += n self.avg = self.sum / self.count class TrackMeter(object): def __init__(self, inc_or_dec='decaying'): self.inc_or_dec = inc_or_dec self.reset() def reset(self): self.val = [] self.epochs = [] self.count = 0 self.best_val = float("inf") if self.inc_or_dec=='decaying' else float("-inf") self.best_count = 0 self.best_epoch = 0 def update(self, val, epoch): if isinstance(val, torch.Tensor): val = val.item() self.val.append(val) self.epochs.append(epoch) if (self.inc_or_dec=='decaying' and val < self.best_val) or (self.inc_or_dec=='increasing' and val > self.best_val): self.best_val = val self.best_count = self.count self.best_count = epoch self.count += 1 class TrackMeter_testing(object): def __init__(self,): self.reset() def reset(self): self.val = [] self.avg = 0 self.std = 0 def update(self, val,): if isinstance(val, torch.Tensor): val = val.item() self.val.append(val) self.avg = np.mean(self.val) self.std = np.std(self.val) class ProgressBar: def __init__(self, iterable, epoch, quiet=False): self.epoch = epoch self.quiet = quiet self.prefix = f"epoch {epoch:02d}" self.iterable = iterable if self.quiet else tqdm(iterable, self.prefix, leave=False) def __iter__(self): return iter(self.iterable) def log(self, stats, verbose=False): if not self.quiet: self.iterable.set_postfix(self.format_stats(stats, verbose), refresh=True) def format_stats(self, stats, verbose=False): postfix = OrderedDict(stats) # method set_postfix requires ordered_dict for key, value in postfix.items(): if isinstance(value, Number): fmt = "{:.6f}" if value > 0.001 else "{:.3e}" postfix[key] = fmt.format(value) elif isinstance(value, AverageMeter): if verbose: postfix[key] = f"{value.avg:.6f} ({value.val:.6f})" else: postfix[key] = f"{value.avg:.6f}" elif not isinstance(postfix[key], str): postfix[key] = str(value) return postfix def print(self, stats, verbose=False): postfix = " | ".join(key + " " + value.strip() for key, value in self.format_stats(stats, verbose).items()) return f"{self.prefix + ' | ' if self.epoch is not None else ''}{postfix}"
3,573
30.910714
127
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/CS_natural_images_figure4/CS_natural_images_functions/losses.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ import torch import torch.nn as nn import torch.nn.functional as F class SSIMLoss(nn.Module): """ SSIM loss module. """ def __init__(self, win_size: int = 7, k1: float = 0.01, k2: float = 0.03): """ Args: win_size: Window size for SSIM calculation. k1: k1 parameter for SSIM calculation. k2: k2 parameter for SSIM calculation. """ super().__init__() device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') self.win_size = win_size self.k1, self.k2 = torch.tensor(k1).to(device), torch.tensor(k2).to(device) self.register_buffer("w", torch.ones(1, 1, win_size, win_size).to(device) / win_size ** 2) NP = win_size ** 2 self.cov_norm = torch.tensor(NP / (NP - 1)).to(device) def forward(self, X: torch.Tensor, Y: torch.Tensor, data_range: torch.Tensor): assert isinstance(self.w, torch.Tensor) data_range = data_range[:, None, None, None] C1 = (self.k1 * data_range) ** 2 C2 = (self.k2 * data_range) ** 2 ux = F.conv2d(X, self.w) # typing: ignore uy = F.conv2d(Y, self.w) # uxx = F.conv2d(X * X, self.w) uyy = F.conv2d(Y * Y, self.w) uxy = F.conv2d(X * Y, self.w) vx = self.cov_norm * (uxx - ux * ux) vy = self.cov_norm * (uyy - uy * uy) vxy = self.cov_norm * (uxy - ux * uy) A1, A2, B1, B2 = ( 2 * ux * uy + C1, 2 * vxy + C2, ux ** 2 + uy ** 2 + C1, vx + vy + C2, ) D = B1 * B2 S = (A1 * A2) / D return 1 - S.mean()
1,849
31.45614
98
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/CS_natural_images_figure4/CS_natural_images_functions/load_save_model_helpers.py
import glob import torch import os from torch.serialization import default_restore_location import logging def setup_experiment_or_load_checkpoint(experiment_path, resume_from='best', model=None, optimizer=None, scheduler=None): ''' Args: - resume_from: Either 'best' or 'some_number' where some_number could by any epoch at which a checkpoint was saved ''' # Look for checkpoints to load from. If avalable, always load. available_models = glob.glob(experiment_path + '*.pt') if available_models: restore_file = experiment_path + f"checkpoint_{resume_from}.pt" print('restoring model..') state_dict = torch.load(restore_file, map_location=lambda s, l: default_restore_location(s, "cpu")) save_checkpoint.last_epoch = state_dict["best_epoch"] if resume_from=='best' else state_dict["last_epoch"] save_checkpoint.start_epoch = state_dict["best_epoch"]+1 if resume_from=='best' else state_dict["last_epoch"]+1 save_checkpoint.best_score = state_dict["best_score"] save_checkpoint.best_epoch = state_dict["best_epoch"] model = [model] if model is not None and not isinstance(model, list) else model optimizer = [optimizer] if optimizer is not None and not isinstance(optimizer, list) else optimizer scheduler = [scheduler] if scheduler is not None and not isinstance(scheduler, list) else scheduler if model is not None and state_dict.get("model", None) is not None: for m, state in zip(model, state_dict["model"]): m.load_state_dict(state) if optimizer is not None and state_dict.get("optimizer", None) is not None: for o, state in zip(optimizer, state_dict["optimizer"]): o.load_state_dict(state) if scheduler is not None and state_dict.get("scheduler", None) is not None: for s, state in zip(scheduler, state_dict["scheduler"]): s.load_state_dict(state) logging.info("Loaded checkpoint {} with best_epoch {} last_epoch {}".format(restore_file, save_checkpoint.best_epoch, save_checkpoint.last_epoch)) else: print("No checkpoint to load. Start training from scratch.") save_checkpoint.best_epoch = -1 save_checkpoint.last_epoch = 0 save_checkpoint.start_epoch = 0 save_checkpoint.best_score = float("-inf") def save_checkpoint(experiment_path, epoch, model, optimizer=None, scheduler=None, score=None, save_at_epochs=None): '''' Args: - ''' save_checkpoint.last_epoch = epoch best_score = save_checkpoint.best_score if score > best_score: save_checkpoint.best_epoch = epoch save_checkpoint.best_score = score if score > best_score: model = [model] if model is not None and not isinstance(model, list) else model optimizer = [optimizer] if optimizer is not None and not isinstance(optimizer, list) else optimizer scheduler = [scheduler] if scheduler is not None and not isinstance(scheduler, list) else scheduler state_dict = { "last_epoch": save_checkpoint.last_epoch, "best_epoch": save_checkpoint.best_epoch, "best_score": save_checkpoint.best_score, "model": [m.state_dict() for m in model] if model is not None else None, "optimizer": [o.state_dict() for o in optimizer] if optimizer is not None else None, "scheduler": [s.state_dict() for s in scheduler] if scheduler is not None else None, } torch.save(state_dict, os.path.join(experiment_path + "checkpoint_best.pt")) if save_at_epochs: if epoch in save_at_epochs: model = [model] if model is not None and not isinstance(model, list) else model optimizer = [optimizer] if optimizer is not None and not isinstance(optimizer, list) else optimizer scheduler = [scheduler] if scheduler is not None and not isinstance(scheduler, list) else scheduler state_dict = { "last_epoch": save_checkpoint.last_epoch, #set "best_epoch": save_checkpoint.best_epoch, #set "best_score": getattr(save_checkpoint, "best_score", None), #set "model": [m.state_dict() for m in model] if model is not None else None, "optimizer": [o.state_dict() for o in optimizer] if optimizer is not None else None, "scheduler": [s.state_dict() for s in scheduler] if scheduler is not None else None, } torch.save(state_dict, os.path.join(experiment_path + f"checkpoint{epoch}.pt"))
4,690
45.445545
154
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/CS_natural_images_figure4/CS_natural_images_functions/unet.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ import torch from torch import nn from torch.nn import functional as F class Unet(nn.Module): """ PyTorch implementation of a U-Net model. O. Ronneberger, P. Fischer, and Thomas Brox. U-net: Convolutional networks for biomedical image segmentation. In International Conference on Medical image computing and computer-assisted intervention, pages 234–241. Springer, 2015. """ def __init__( self, in_chans: int, out_chans: int, chans: int = 32, num_pool_layers: int = 4, drop_prob: float = 0.0, ): """ Args: in_chans: Number of channels in the input to the U-Net model. out_chans: Number of channels in the output to the U-Net model. chans: Number of output channels of the first convolution layer. num_pool_layers: Number of down-sampling and up-sampling layers. drop_prob: Dropout probability. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.chans = chans self.num_pool_layers = num_pool_layers self.drop_prob = drop_prob self.down_sample_layers = nn.ModuleList([ConvBlock(in_chans, chans, drop_prob)]) ch = chans for _ in range(num_pool_layers - 1): self.down_sample_layers.append(ConvBlock(ch, ch * 2, drop_prob)) ch *= 2 self.conv = ConvBlock(ch, ch * 2, drop_prob) self.up_conv = nn.ModuleList() self.up_transpose_conv = nn.ModuleList() for _ in range(num_pool_layers - 1): self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch)) self.up_conv.append(ConvBlock(ch * 2, ch, drop_prob)) ch //= 2 self.up_transpose_conv.append(TransposeConvBlock(ch * 2, ch)) self.up_conv.append( nn.Sequential( ConvBlock(ch * 2, ch, drop_prob), nn.Conv2d(ch, self.out_chans, kernel_size=1, stride=1), # here is the only conv layer with a bias ) ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H, W)`. """ stack = [] output = image # apply down-sampling layers for layer in self.down_sample_layers: output = layer(output) stack.append(output) output = F.avg_pool2d(output, kernel_size=2, stride=2, padding=0) output = self.conv(output) # apply up-sampling layers for transpose_conv, conv in zip(self.up_transpose_conv, self.up_conv): downsample_layer = stack.pop() output = transpose_conv(output) # reflect pad on the right/botton if needed to handle odd input dimensions padding = [0, 0, 0, 0] if output.shape[-1] != downsample_layer.shape[-1]: padding[1] = 1 # padding right if output.shape[-2] != downsample_layer.shape[-2]: padding[3] = 1 # padding bottom if torch.sum(torch.tensor(padding)) != 0: output = F.pad(output, padding, "reflect") output = torch.cat([output, downsample_layer], dim=1) output = conv(output) return output class ConvBlock(nn.Module): """ A Convolutional Block that consists of two convolution layers each followed by instance normalization, LeakyReLU activation and dropout. """ def __init__(self, in_chans: int, out_chans: int, drop_prob: float): """ Args: in_chans: Number of channels in the input. out_chans: Number of channels in the output. drop_prob: Dropout probability. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.drop_prob = drop_prob self.layers = nn.Sequential( nn.Conv2d(in_chans, out_chans, kernel_size=3, padding=1, bias=False), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Dropout2d(drop_prob), nn.Conv2d(out_chans, out_chans, kernel_size=3, padding=1, bias=False), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), nn.Dropout2d(drop_prob), ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H, W)`. """ return self.layers(image) class TransposeConvBlock(nn.Module): """ A Transpose Convolutional Block that consists of one convolution transpose layers followed by instance normalization and LeakyReLU activation. """ def __init__(self, in_chans: int, out_chans: int): """ Args: in_chans: Number of channels in the input. out_chans: Number of channels in the output. """ super().__init__() self.in_chans = in_chans self.out_chans = out_chans self.layers = nn.Sequential( nn.ConvTranspose2d( in_chans, out_chans, kernel_size=2, stride=2, bias=False ), nn.InstanceNorm2d(out_chans), nn.LeakyReLU(negative_slope=0.2, inplace=True), ) def forward(self, image: torch.Tensor) -> torch.Tensor: """ Args: image: Input 4D tensor of shape `(N, in_chans, H, W)`. Returns: Output tensor of shape `(N, out_chans, H*2, W*2)`. """ return self.layers(image)
6,021
31.907104
113
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/CS_natural_images_figure4/CS_natural_images_functions/fftc.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ from typing import List, Optional import torch import torch.fft # type: ignore def fft2c(data: torch.Tensor) -> torch.Tensor: """ Apply centered 2 dimensional Fast Fourier Transform. Args: data: Complex valued input data containing at least 3 dimensions: dimensions -3 & -2 are spatial dimensions and dimension -1 has size 2. All other dimensions are assumed to be batch dimensions. Returns: The FFT of the input. """ if not data.shape[-1] == 2: raise ValueError("Tensor does not have separate complex dim.") data = ifftshift(data, dim=[-3, -2]) data = torch.view_as_real( torch.fft.fftn( # type: ignore torch.view_as_complex(data), dim=(-2, -1), norm="ortho" ) ) data = fftshift(data, dim=[-3, -2]) return data def ifft2c(data: torch.Tensor) -> torch.Tensor: """ Apply centered 2-dimensional Inverse Fast Fourier Transform. Args: data: Complex valued input data containing at least 3 dimensions: dimensions -3 & -2 are spatial dimensions and dimension -1 has size 2. All other dimensions are assumed to be batch dimensions. Returns: The IFFT of the input. """ if not data.shape[-1] == 2: raise ValueError("Tensor does not have separate complex dim.") data = ifftshift(data, dim=[-3, -2]) data = torch.view_as_real( torch.fft.ifftn( # type: ignore torch.view_as_complex(data), dim=(-2, -1), norm="ortho" ) ) data = fftshift(data, dim=[-3, -2]) return data # Helper functions def roll_one_dim(x: torch.Tensor, shift: int, dim: int) -> torch.Tensor: """ Similar to roll but for only one dim. Args: x: A PyTorch tensor. shift: Amount to roll. dim: Which dimension to roll. Returns: Rolled version of x. """ shift = shift % x.size(dim) if shift == 0: return x left = x.narrow(dim, 0, x.size(dim) - shift) right = x.narrow(dim, x.size(dim) - shift, shift) return torch.cat((right, left), dim=dim) def roll( x: torch.Tensor, shift: List[int], dim: List[int], ) -> torch.Tensor: """ Similar to np.roll but applies to PyTorch Tensors. Args: x: A PyTorch tensor. shift: Amount to roll. dim: Which dimension to roll. Returns: Rolled version of x. """ if len(shift) != len(dim): raise ValueError("len(shift) must match len(dim)") for (s, d) in zip(shift, dim): x = roll_one_dim(x, s, d) return x def fftshift(x: torch.Tensor, dim: Optional[List[int]] = None) -> torch.Tensor: """ Similar to np.fft.fftshift but applies to PyTorch Tensors Args: x: A PyTorch tensor. dim: Which dimension to fftshift. Returns: fftshifted version of x. """ if dim is None: # this weird code is necessary for toch.jit.script typing dim = [0] * (x.dim()) for i in range(1, x.dim()): dim[i] = i # also necessary for torch.jit.script shift = [0] * len(dim) for i, dim_num in enumerate(dim): shift[i] = x.shape[dim_num] // 2 return roll(x, shift, dim) def ifftshift(x: torch.Tensor, dim: Optional[List[int]] = None) -> torch.Tensor: """ Similar to np.fft.ifftshift but applies to PyTorch Tensors Args: x: A PyTorch tensor. dim: Which dimension to ifftshift. Returns: ifftshifted version of x. """ if dim is None: # this weird code is necessary for toch.jit.script typing dim = [0] * (x.dim()) for i in range(1, x.dim()): dim[i] = i # also necessary for torch.jit.script shift = [0] * len(dim) for i, dim_num in enumerate(dim): shift[i] = (x.shape[dim_num] + 1) // 2 return roll(x, shift, dim)
4,108
23.753012
80
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/CS_natural_images_figure4/CS_natural_images_functions/data_transforms.py
import numpy as np import torch from typing import Callable, Dict, List, Optional, Sequence, Tuple, Union import torchvision.transforms as transforms import PIL.Image as Image from CS_natural_images_functions.log_progress_helpers import save_figure from CS_natural_images_functions.fftc import fft2c, ifft2c class CropDataset(torch.utils.data.Dataset): """ A PyTorch Dataset that provides access to cropped images from ImageNet. """ def __init__( self, dataset: List, path_to_ImageNet_train: str, transform: Callable, experiment_path: str, img_size: int, ): self.transform = transform self.experiment_path = experiment_path device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') self.examples = [] load_transform = transforms.Compose([ transforms.CenterCrop(img_size), transforms.ToTensor(), ]) for datapath in dataset: image = Image.open(path_to_ImageNet_train+datapath).convert("L") filename = datapath[16:-5] self.examples.append((load_transform(image)[0].to(device),filename)) def __len__(self): return len(self.examples) def __getitem__(self, i: int): # Determine input, target and ground truth x,filename = self.examples[i] sample = self.transform(x,filename,i) return sample class UnetDataTransform: def __init__( self, acceleration, acceleration_total, fix_split, experiment_path, center_fraction, ): self.acceleration = acceleration self.acceleration_total = acceleration_total self.fix_split = fix_split self.experiment_path = experiment_path self.center_fraction = center_fraction def __call__( self, x: np.ndarray, fname: str, id: int, ) -> Tuple[torch.Tensor,torch.Tensor]: """ Args: Returns: tuple containing: x_input: zero-filled coarse reconstruction in image domain y_target: the fully sampled kspace x: the ground truth image input_mask: undersampled input mask target_mask: in the case of supervised training all ones """ device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') n = x.shape[-1] # transform x to a tensor with real channel and complex channel. Right now the complex channel is all zerors. #x = np.stack((x, np.zeros_like(x)), axis=-1) x = torch.stack((x, torch.zeros_like(x)), axis=-1) #x = torch.from_numpy(x) # obtain kspace y = fft2c(x) #save_figure(y[:,:,0],"y_real",self.experiment_path) if id==0 else None #save_figure(y[:,:,1],"y_imag",self.experiment_path) if id==0 else None ####################################### # sample input mask nu = self.center_fraction p = 1/self.acceleration mu = 1/self.acceleration_total q = (mu-p+nu-mu*nu)/(1-p) # 1. Determine the set S_low consisting of the indices of the nu*n many center frequencies which are always sampled size_low = int(round(n*nu)) pad = (n - size_low + 1) // 2 # set of indices of all lines in kspace S_all = np.arange(n) S_low = S_all[pad : pad + size_low] # 1.1 Determine S_mu_high, i.e, S_mu without S_low, so only the random high frequencies # set of indices of all high frequencies S_high = np.hstack((S_all[: pad],S_all[pad + size_low :])) S_mu_size_high = int(round((mu-nu)*n)) S_p_size_high = int(round((p-nu)*n)) #### Depending on whether the input/target split is fixed or re-sampled, the order of sampling needs to be adapted # This is so that validation during training samples the same input mask as during testing # Recall that during testing selfsup=False, hence S_mu_high is not sampled. seed = tuple(map(ord, fname)) rng = np.random.default_rng(seed) if self.fix_split: # If split is fixed, first sample S_p_high and then additional lines for S_mu_high # such that the set S_p_high is the same as if we would sample for selfsup=False S_p_high = rng.choice(S_high, size=S_p_size_high, replace=False, p=None) S_mu_size_high_remainding = S_mu_size_high - S_p_size_high S_high_remainding = np.array(list(set(S_high)-set(S_p_high))) S_q_high = rng.choice(S_high_remainding, size=S_mu_size_high_remainding, replace=False, p=None) else: # If split is random, first sample S_mu_high such that this set is always fixed. S_mu_high = rng.choice(S_high, size=S_mu_size_high, replace=False, p=None) # 2. From S_mu_high sample the set S_p_high of size (p-nu)n S_p_high = np.random.choice(S_mu_high, size=S_p_size_high, replace=False, p=None) # 3. All other indices in S_mu_high add to the set S_q_high S_q_high = np.array(list(set(S_mu_high)-set(S_p_high))) # 4. Determine the size of the overlap between S_p_high and S_q_high, sample this many indices from S_p_high and add them to S_q_high overlap_size_high = int(round(( (p-nu) / (1-nu) ) * ( (q-nu) / (1-nu) ) *(n-n*nu))) S_overlap = S_p_high[0:overlap_size_high] S_q_high = np.concatenate([S_q_high,S_overlap]) # 5. Define the final input and target masks by setting entries to zero or to one for S_p=S_low+S_p_high and S_q=S_low+S_q_high input_mask = np.zeros(n) input_mask[S_low] = 1.0 input_mask[S_p_high] = 1.0 input_mask = torch.from_numpy(input_mask.astype(np.float32)).unsqueeze(0).unsqueeze(-1).to(device) # 6. Create a target mask where the random entries are weighted weight_on_random_lines = np.sqrt((1-nu)/(q-nu)) target_mask = np.zeros(n) target_mask[S_low] = 1.0 target_mask[S_q_high] = weight_on_random_lines target_mask = torch.from_numpy(target_mask.astype(np.float32)).unsqueeze(0).unsqueeze(-1).to(device) ####################################### # apply mask to kspace y_input = y * input_mask + 0.0 #save_figure(y_input[:,:,0],"y_input_real",self.experiment_path) if id==0 else None #save_figure(y_input[:,:,1],"y_input_imag",self.experiment_path) if id==0 else None # compute zero-filed coarse reconstruction as input x_input = ifft2c(y_input) #save_figure(x_input[:,:,0],"x_input_real",self.experiment_path) if id==0 else None #save_figure(x_input[:,:,1],"x_input_imag",self.experiment_path) if id==0 else None mean = x_input.mean(dim=[0,1],keepdim=True) std = x_input.std(dim=[0,1],keepdim=True) x_input = (x_input - mean) / (std + 1e-11) # training target. target_mask is all ones if supervised training y_target = y * target_mask + 0.0 # training target in image domain x_target = ifft2c(y_target) return y_input, x_input, y_target, x_target, x, input_mask, target_mask, mean, std, fname def compute_number_of_lines_in_input_target_kspace(p,mu,nu,n=160): q = (mu-p+nu-mu*nu)/(1-p) size_low = int(round(n*nu)) S_p_size_high = int(round((p-nu)*n)) S_mu_size_high = int(round((mu-nu)*n)) S_mu_size_high_remainding = S_mu_size_high - S_p_size_high overlap_size_high = int(round(( (p-nu) / (1-nu) ) * ( (q-nu) / (1-nu) ) *(n-n*nu))) input_size = size_low + S_p_size_high target_size = size_low + S_mu_size_high_remainding + overlap_size_high weight_on_random_lines = np.sqrt((1-nu)/(q-nu)) return input_size, target_size, overlap_size_high, size_low, p, q, mu, nu, weight_on_random_lines
8,047
36.432558
141
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/CS_natural_images_figure4/CS_natural_images_functions/log_progress_helpers.py
import numpy as np import matplotlib.pyplot as plt from typing import Dict, Optional, Sequence, Tuple, Union, List import os import torchvision import io import torch from CS_natural_images_functions.losses import SSIMLoss def complex_abs(data: torch.Tensor) -> torch.Tensor: """ Compute the absolute value of a complex valued input tensor. Args: data: A complex valued tensor, where the size of the final dimension should be 2. Returns: Absolute value of data. """ if not data.shape[-1] == 2: raise ValueError("Tensor does not have separate complex dim.") return (data ** 2).sum(dim=-1).sqrt() def plot_to_image(figure): """Converts the matplotlib plot specified by 'figure' to a PNG image and returns it. The supplied figure is closed and inaccessible after this call.""" # Save the plot to a PNG in memory. buf = io.BytesIO() plt.savefig(buf, format='png') plt.close(figure) buf.seek(0) frameTensor = torch.tensor(np.frombuffer(buf.getvalue(), dtype=np.uint8), device='cpu') image = torchvision.io.decode_png(frameTensor) return image def get_figure(image,figsize,title): """Return a matplotlib figure of a given image.""" if len(image.shape) != 3: raise ValueError("Image dimensions not suitable for logging to tensorboard.") if image.shape[0] == 1 or image.shape[0] == 3: image = np.rollaxis(image,0,3) # Create a figure to contain the plot. if figsize: figure = plt.figure(figsize=figsize) else: figure = plt.figure() # Start next subplot. plt.subplot(1, 1, 1, title=title) plt.xticks([]) plt.yticks([]) plt.grid(False) plt.imshow(image, cmap='gray') figure.tight_layout() return figure def plot_figure( x: np.array,): """" x must have dimension height,width """ fig = plt.figure() ax = fig.add_subplot(111) ax.imshow(x,'gray') ax.axis('off') #ax.set_title(title,fontsize=10) fig.tight_layout() plt.show() def save_figure( x: np.array, figname: str, experiment_path: str, save: Optional[bool]=True,): """" x must have dimension height,width """ if save: save_path = experiment_path + 'train_figures/' if not os.path.isdir(save_path): os.mkdir(save_path) fig = plt.figure(figsize=(7,7)) ax = fig.add_subplot(111) ax.imshow(x,'gray') ax.axis('off') #ax.set_title(title,fontsize=10) fig.tight_layout() plt.savefig(save_path + figname + ".png") plt.close(fig) def save_test_image_with_dc(experiment_path, ground_truth_image, input_img, output, output_image_dc, fname, track_meter): save_path = experiment_path + 'test_figures/' if not os.path.isdir(save_path): os.mkdir(save_path) error = torch.abs(ground_truth_image - output) error_dc = torch.abs(ground_truth_image - output_image_dc) output = output - output.min() output = output / output.max() output_image_dc = output_image_dc - output_image_dc.min() output_image_dc = output_image_dc / output_image_dc.max() ground_truth_image = ground_truth_image - ground_truth_image.min() ground_truth_image = ground_truth_image / ground_truth_image.max() input_img = input_img - input_img.min() input_img = input_img / input_img.max() error = error - error.min() error_dc = error_dc - error_dc.min() max_norm = torch.stack([error,error_dc]).max() error = error / max_norm error_dc = error_dc / max_norm image = torch.cat([ground_truth_image, input_img, output, output_image_dc, error, error_dc], dim=0) image = torchvision.utils.make_grid(image, nrow=2, normalize=False, value_range=(0,1), pad_value=1) ssim_score = track_meter["SSIM"].val[-1] ssim_score_dc = track_meter["SSIM_dc"].val[-1] psnr_score = track_meter["PSNR"].val[-1] psnr_score_dc = track_meter["PSNR_dc"].val[-1] figure = get_figure(image.cpu().numpy(),figsize=(8,12),title=f"ssim={ssim_score:.4f}, dc={ssim_score_dc:.4f}, psnr={psnr_score:.3f}, dc={psnr_score_dc:.3f}") plt.savefig(experiment_path + 'test_figures/' + f"{fname[0]}.png", dpi='figure') plt.close() def add_img_to_tensorboard(writer, epoch, name, input_img_comp,output_comp,targetcomp,ksp): if ksp: input_img = torch.log(complex_abs(input_img_comp)[0]+ 1e-9) output = torch.log(complex_abs(output_comp)[0]+ 1e-9) target = torch.log(complex_abs(targetcomp)[0]+ 1e-9) else: input_img = complex_abs(input_img_comp)[0] output = complex_abs(output_comp)[0] target = complex_abs(targetcomp)[0] val_ssim_fct = SSIMLoss() max_value = target.max().unsqueeze(0) ssim_loss = 1-val_ssim_fct(output.unsqueeze(0).unsqueeze(0), target.unsqueeze(0).unsqueeze(0), data_range=max_value) error = torch.abs(target - output) input_img = input_img - input_img.min() input_img = input_img / input_img.max() output = output - output.min() output = output / output.max() target = target - target.min() target = target / target.max() error = error - error.min() error = error / error.max() image = torch.cat([input_img, target, output, error], dim=0) image = torchvision.utils.make_grid(image, nrow=1, normalize=False) figure = get_figure(image.cpu().numpy(),figsize=(3,12),title=f"ssim={ssim_loss.item():.6f}") writer.add_image(name+"_abs", plot_to_image(figure), epoch) if ksp: input_img = torch.log(torch.abs(input_img_comp[0,:,:,0])+ 1e-9) #input_img_max = torch.max(torch.stack((torch.log(torch.abs(input_img_comp[0,:,:,0])+ 1e-9),torch.log(torch.abs(input_img_comp[0,:,:,1])+ 1e-9)))) output = torch.log(torch.abs(output_comp[0,:,:,0])+ 1e-9) #output_max = torch.max(torch.stack((torch.log(torch.abs(output_comp[0,:,:,0])+ 1e-9),torch.log(torch.abs(output_comp[0,:,:,1])+ 1e-9)))) target = torch.log(torch.abs(targetcomp[0,:,:,0])+ 1e-9) #target_max = torch.max(torch.stack((torch.log(torch.abs(output_comp[0,:,:,0])+ 1e-9),torch.log(torch.abs(output_comp[0,:,:,1])+ 1e-9)))) else: input_img = input_img_comp[0,:,:,0] output = output_comp[0,:,:,0] target = targetcomp[0,:,:,0] val_ssim_fct = SSIMLoss() max_value = target.max().unsqueeze(0) ssim_loss = 1-val_ssim_fct(output.unsqueeze(0).unsqueeze(0), target.unsqueeze(0).unsqueeze(0), data_range=max_value) error = torch.abs(target - output) input_img = input_img - input_img.min() input_img = input_img / input_img.max() output = output - output.min() output = output / output.max() target = target - target.min() target = target / target.max() error = error - error.min() error = error / error.max() image = torch.cat([input_img, target, output, error], dim=0) image = torchvision.utils.make_grid(image, nrow=1, normalize=False) figure = get_figure(image.cpu().numpy(),figsize=(3,12),title=f"ssim={ssim_loss.item():.6f}") writer.add_image(name+"_re", plot_to_image(figure), epoch) if ksp: input_img = torch.log(torch.abs(input_img_comp[0,:,:,1])+ 1e-9) output = torch.log(torch.abs(output_comp[0,:,:,1])+ 1e-9) target = torch.log(torch.abs(targetcomp[0,:,:,1])+ 1e-9) else: input_img = input_img_comp[0,:,:,1] output = output_comp[0,:,:,1] target = targetcomp[0,:,:,1] val_ssim_fct = SSIMLoss() max_value = target.max().unsqueeze(0) ssim_loss = 1-val_ssim_fct(output.unsqueeze(0).unsqueeze(0), target.unsqueeze(0).unsqueeze(0), data_range=max_value) error = torch.abs(target - output) input_img = input_img - input_img.min() input_img = input_img / input_img.max() output = output - output.min() output = output / output.max() target = target - target.min() target = target / target.max() error = error - error.min() error = error / error.max() image = torch.cat([input_img, target, output, error], dim=0) image = torchvision.utils.make_grid(image, nrow=1, normalize=False) figure = get_figure(image.cpu().numpy(),figsize=(3,12),title=f"ssim={ssim_loss.item():.6f}") writer.add_image(name+"_im", plot_to_image(figure), epoch) plt.close()
8,395
35.663755
162
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/CS_accelerated_MRI_figure5/run_MRI.py
# %% ################################## # Import python packages import numpy as np import os import traceback # Import main() from functions.main import main_train, main_test #from functions.train_utils import get_event_details os.environ['CUDA_VISIBLE_DEVICES'] = '3' # %% ################################## # Customize hyperparemters # Specify an ID for each experiment, e.g. 001,002,... exp_nums = ['001'] ## Available datasets for MRI experiments: #train_49827_selfsup_slice #train_24900_selfsup_slice #train_9977_selfsup_slice #train_4943_selfsup_slice #train_2466_selfsup_slice #train_970_selfsup_slice #train_491_selfsup_slice #train_223_selfsup_slice #train_96_selfsup_slice #train_48_selfsup_slice #test_4713_selfsup_slice #val_313_selfsup_slice hp_all_exps = { ############################################################# ### SET FALGS AND PICK NUMBER OF RUNS ### # Specify the number of runs with different random initialization per experiment, e.g. [0,1,2] means three runs 'num_runs' : [[0]], # Start training or continue training if there exists already a checkpoint. 'training' : [True], # Evaluate best and last checkpoint on validation and test set. 'testing' : [True], # Load model from last or best checkpoint. "last" is for continuing an interrupted training, "best" is automatically set for test mode, Right now "best" is not supported during training. 'resume_from_which_checkpoint' : ["last"], ############################################################# ### CHOOSE DATASETS ### # Path to data directories: '../../../../srv/multicoil_brain_subset/' (zion) or '../../../../media/ssd1/fastMRIdata/brain/' (bigsur, yosemite) 'data_path' : ['../../../media/ssd1/fastMRIdata/brain/'], # Path to val data directories. Only specify if different from data_path. 'val_path' : [None], # Path to test data directories. Only specify if different from data_path. 'test_path' : [None], # Path to sensitivity maps directories. 'smaps_path' : ['../../../media/ssd1/fastMRIdata/brain_sensmaps_train_and_val/'], # If true sens maps are loaded in the train, validation and test loader. 'provide_senmaps' : [True], # Choose a training set, e.g. train_491_selfsup_slice 'train_set' : ['./datasets/train_48_selfsup_slice.yaml'], # Specify the size of the training set. 'train_size' : [50], # Choose a validation set. 'val_set' : ['./datasets/val_313_selfsup_slice.yaml'], # Choose test sets: 'test_sets' : [['./datasets/val_313_selfsup_slice.yaml','./datasets/test_4713_selfsup_slice.yaml']], # Which challenge to preprocess for. 'challenge' : ["multicoil"], ############################################################# ### DEFINE SETTINGS FOR MASK FUNCTIONS ### # Type of k-space mask. Use n2n. 'mask_type' : ["n2n"], # Number of center lines to use in mask. 0.08 for acceleration 4. 'center_fraction' : [0.08], # Acceleration rates to use for masks, e.g. 4 or 8. This acceleration factor is always applied to create network inputs 'acceleration' : [4.0], # If True: the same input and target masks for both supervised and self-supervised training are used over all epochs. # Note that masks still differ over slices in a volume. # If False: Supervised training draws a new input mask in every epoch for every training slice # Self-supervised still always draws the same set of frequecies defined by 'acceleration_total', but the split into input and target is re-sampled in every epoch. 'use_mask_seed_for_training' : [False], ############################################################# ### ADDITIONAL FLAGS FOR SELF-SUPERVISED TRAINING # If True the training loss is only computed on the masked kspace fraction selfsup_acceleration (plus accelerations if selfsup_compute_loss_on_input_freq is True) 'selfsup' : [True], # Determines the number of frequencies given during self-supervised training. Those are then split into input and target frequencies # such that the input has acceleration factor 'acceleration' and the target contains the remaining freqs plus some overlap. # Must be larger than 'accelerations'. If selfsup=False this value does not matter. 'acceleration_total' : [3.0], ############################################################# ### DEFINE THE MODEL ARCHITECTURE ### # Number of channels in the first layer. 'chans' : [64], # Number of downsampling/upsampling layers in the UNet (0 or 1 still has one downsampling/upsampling). 'num_pool_layers' : [4], # Use separate channels for real and imaginary part. Must be True. 'two_channel_imag_real' : [True], ############################################################# ### OPTIMIZER, LOSS FUNCTION, NUMBER OF EPOCHS, BATCH SIZE, INITIAL LEARNING RATE ### # Currently available RMSprop or Adam 'optimizer' : ['RMSprop'], # List of loss functions for training. Currently available L1, L2 or SSIM loss or L2_kspace. 'loss_functions' : [['L2_kspace']], # If True, then the network input is not cropped. 'compute_sup_loss_in_kspace' : [True], # Maximal number of epochs. 'num_epochs' : [1000], # Initial learning rate. 'lr' : [1e-3], # Batch size for mini batch training. 'batch_size' : [1], ############################################################# ### LEARNING RATE SCHEME FOR DECAY ### # Currently available MultiStepLR or ReduceLROnPlateau 'lr_scheduler' : ['ReduceLROnPlateau'], # If true, lr_convergence_break_counter and lr_min_break_counter are applied 'early_stop_lr_deacy' : [True], # Terminate training once this lr is reached. 'lr_min' : [1e-6], # Once lr_min is reached training continues for this many epochs before terminated. 'lr_min_break_counter' : [10], # Once for this many consecutive learning rate decays no improvement in val_loss is observed the training is early stopped. 'lr_convergence_break_counter' : [2], # Decay lr by this factor. 'lr_decay_factor' : [0.1], # If scheduler is MultiStepLR then decay lr after these number of epochs. 'lr_milestones' : [[50,60]], # If scheduler is ReduceLROnPlateau, decay lr after these many epochs without improving val loss. 'lr_patience' : [10], # If scheduler is ReduceLROnPlateau, improving val loss by less than this does not count as improvement. 'lr_threshold' : [0.0001], # Metric used to determine a Plateau 'decay_metric' : ['L2_kspace'], ############################################################# ### CHOOSE OPTIONS FOR ACCELERATING TRAINING AND CHECKPOINTING # Number of workers for dataloader. 'num_workers' : [8], # Enable logging to Tensorboard 'tb_logging' : [True], # List of validaiton sample indices to be logged to Tensorboard. Put empty dictionary to not log images 'log_val_images' : [{'file_brain_AXT2_210_2100189':3, 'file_brain_AXT2_200_2000396': 7, 'file_brain_AXT2_200_6002509': 12}], 'log_train_images' : [{'file_brain_AXT2_207_2070586':8, 'file_brain_AXT2_210_6001651': 6}], # Log images ever log_image_interval epochs 'log_image_interval' : [8], # Optional: A list of epochs at which the model is saved as "checkpoint{}.pt".format(epoch) 'epoch_checkpoints' : [None], # Interval for validation 'val_interval' : [1], # Save images from the testset to the log directory. Dictionary containing filenames and slice numbers. Put empty dictionary to not log images 'save_test_images': [{'file_brain_AXT2_207_2070041': 7, 'file_brain_AXT2_201_2010484': 12, 'file_brain_AXT2_203_2030348': 3,'file_brain_AXT2_209_6001477':9, 'file_brain_AXT2_210_6001600': 6, 'file_brain_AXT2_205_2050106': 7}], } # Sanity checks for key in hp_all_exps.keys(): # expand lists with constant settings to the length of the number of experiments if len(hp_all_exps[key]) == 1 and len(hp_all_exps[key]) != len(exp_nums): hp_all_exps[key] = [hp_all_exps[key][0] for _ in range(len(exp_nums))] #hp_all_exps[key] = hp_all_exps[key]*(len(exp_nums)) if len(hp_all_exps[key]) != len(exp_nums): print(key) raise ValueError("Specify hyperparameters for every experiment!") # %% # Print out the names of the new experiments printouts = [] print('') for ee in range(len(exp_nums)): for rr in hp_all_exps['num_runs'][ee]: exp_name = 'E' + str(exp_nums[ee]) + \ '_t' + str(hp_all_exps['train_size'][ee]) + \ '_l' + str(hp_all_exps['num_pool_layers'][ee]) + \ 'c' + str(hp_all_exps['chans'][ee]) + \ '_bs' + str(hp_all_exps['batch_size'][ee]) +\ '_lr' + str(np.round(hp_all_exps['lr'][ee],6))[2:] if rr>0: exp_name = exp_name + '_run{}'.format(rr+1) printouts.append(exp_name) print(printouts) # %% for ee in range(len(exp_nums)): hp_exp = {} for k in hp_all_exps.keys(): hp_exp[k] = hp_all_exps[k][ee]#.copy() for rr in hp_exp['num_runs']: exp_name = 'E' + str(exp_nums[ee]) + \ '_t' + str(hp_all_exps['train_size'][ee]) + \ '_l' + str(hp_all_exps['num_pool_layers'][ee]) + \ 'c' + str(hp_all_exps['chans'][ee]) + \ '_bs' + str(hp_all_exps['batch_size'][ee]) +\ '_lr' + str(np.round(hp_all_exps['lr'][ee],6))[2:] if rr>0: exp_name = exp_name + '_run{}'.format(rr+1) if not os.path.isdir('./'+exp_name): os.mkdir('./'+exp_name) hp_exp['seed'] = int(42 + 10*rr) hp_exp['exp_name'] = exp_name if not hp_exp['train_set']: hp_exp['train_set'] = f"./datasets/train_{hp_exp['train_size']}_filenames_slice.yaml" ######## # Training ######## try: if hp_exp['training']: print('\n{} - Training\n'.format(exp_name)) hp_exp['mode'] = 'train' # Perform training train_meters, val_metric_dict = main_train(hp_exp) print('\n{} - Training finished\n'.format(exp_name)) except: with open("./"+exp_name+"/errors_train.txt", "a+") as text_file: error_str = traceback.format_exc() print(error_str, file=text_file) print(error_str) ######## # Testing ######## try: if hp_exp['testing']: # for test_set in hp_exp['test_sets']: print('\n{} - Testing {}\n'.format(exp_name, test_set)) # Prepare args for testing hp_exp['test_set'] = test_set hp_exp['mode'] = 'test' hp_exp['resume_from_which_checkpoint'] = 'best' # Create a directory reconstructions with all reconstructions main_test(hp_exp) print('\n{} - Testing finished\n'.format(exp_name)) except: with open("./"+exp_name+"/errors_test.txt", "a+") as text_file: error_str = traceback.format_exc() print(error_str, file=text_file) print(error_str) # %%
11,447
44.070866
230
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/CS_accelerated_MRI_figure5/functions/main.py
################# # Import python packages import torch import logging import time from torch.utils.tensorboard import SummaryWriter import sys import os from torch.serialization import default_restore_location from collections import defaultdict import numpy as np import torchvision import pickle import matplotlib.pyplot as plt from packaging import version from torch.nn import L1Loss, MSELoss from functions.math import complex_abs, complex_mul, complex_conj from functions.data.transforms import center_crop_to_smallest, normalize_to_given_mean_std if version.parse(torch.__version__) >= version.parse("1.7.0"): from functions.fftc import fft2c_new as fft2c from functions.fftc import ifft2c_new as ifft2c else: from functions.fftc import fft2c_old as fft2c from functions.fftc import ifft2c_old as ifft2c # Implementation of SSIMLoss from functions.training.losses import SSIMLoss from functions.training.debug_helper import print_tensor_stats, save_figure # Set seeds, create directories, set path to checkpoints if available from functions.train_utils import setup_experiment from functions.train_utils import load_checkpoint,save_checkpoint, init_logging # Function that returns a MaskFunc object from functions.data.subsample import create_mask_for_mask_type from functions.data.transforms import UnetDataTransform from functions.data.mri_dataset import SliceDataset from functions.models.unet import Unet # Create scheduler and optimizer objects from functions.training.training_functions import configure_optimizers, Compute_batch_train_loss # Class that allows to track the average of some quantity over an epoch from functions.training.meters import AverageMeter # Gives a customized tqdm object that can be used as iterable instead of train_loader from functions.training.progress_bar import ProgressBar # Functions to log images with a header to tensorboard from functions.log_save_image_utils import plot_to_image, get_figure def add_img_to_tensorboard(writer, epoch, name, input_img, target, output, val_ssim_fct, max_value, crop): output, _ = center_crop_to_smallest(output, target) input_img, _ = center_crop_to_smallest(input_img, target) # Normalize output to mean and std of target #target, output = normalize_to_given_mean_std(target, output) ssim_loss = 1-val_ssim_fct(output, target, data_range=max_value) error = torch.abs(target - output) input_img = input_img - input_img.min() input_img = input_img / input_img.max() output = output - output.min() output = output / output.max() target = target - target.min() target = target / target.max() error = error - error.min() error = error / error.max() image = torch.cat([input_img, target, output, error], dim=0) image = torchvision.utils.make_grid(image, nrow=1, normalize=False) if crop: figure = get_figure(image.cpu().numpy(),figsize=(3,12),title=f"ssim={ssim_loss.item():.6f}") else: figure = get_figure(image.cpu().numpy(),figsize=(3,20),title=f"ssim={ssim_loss.item():.6f}") writer.add_image(name, plot_to_image(figure), epoch) def main_train(hp_exp): device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') # ------------ # setup: # Set seeds, create directories, set path to checkpoints if available # ------------ hp_exp = setup_experiment(hp_exp) init_logging(hp_exp) writer = SummaryWriter(log_dir=hp_exp['log_path']) if hp_exp['tb_logging'] else None # Get list of filenames logged to tensorboard during validation (from the validation set) val_log_filenames_list = [] for k in hp_exp['log_val_images'].keys(): val_log_filenames_list.append(k) train_log_filenames_list = [] for k in hp_exp['log_train_images'].keys(): train_log_filenames_list.append(k) mode_lookup = { 'SSIM' : 'max', 'PSNR' : 'max', 'L1' : 'min', 'L2' : 'min', 'MSE' : 'min', 'L2_kspace' : 'min', 'L1_kspace' : 'min', } # ------------ # data # ------------ mask_func = create_mask_for_mask_type( hp_exp['mask_type'], hp_exp['selfsup'], hp_exp['center_fraction'], hp_exp['acceleration'], hp_exp['acceleration_total'] ) data_transform_train = UnetDataTransform(hp_exp['challenge'],mask_func=mask_func, use_seed=hp_exp['use_mask_seed_for_training'], hp_exp=hp_exp,mode="train") data_transform_val = UnetDataTransform(hp_exp['challenge'],mask_func=mask_func, use_seed=True, hp_exp=hp_exp,mode="val") def _init_fn(worker_id): np.random.seed(12 + worker_id) trainset = SliceDataset( dataset=hp_exp['train_set'], path_to_dataset=hp_exp['data_path'], path_to_sensmaps=hp_exp['smaps_path'], provide_senmaps=hp_exp['provide_senmaps'], challenge=hp_exp['challenge'], transform=data_transform_train, use_dataset_cache=True, ) train_loader = torch.utils.data.DataLoader( dataset=trainset, batch_size=hp_exp['batch_size'], num_workers=hp_exp['num_workers'], shuffle=True, generator=torch.Generator().manual_seed(hp_exp['seed']), pin_memory =True, ) valset = SliceDataset( dataset=hp_exp['val_set'], path_to_dataset=hp_exp['data_path'], path_to_sensmaps=hp_exp['smaps_path'], provide_senmaps=hp_exp['provide_senmaps'], challenge=hp_exp['challenge'], transform=data_transform_val, use_dataset_cache=True, ) val_loader = torch.utils.data.DataLoader( dataset=valset, batch_size=1, num_workers=hp_exp['num_workers'], shuffle=False, generator=torch.Generator().manual_seed(hp_exp['seed']), ) # ------------ # model # ------------ if hp_exp['two_channel_imag_real']: in_chans = 2 else: in_chans = 1 model = Unet( in_chans=in_chans, out_chans=in_chans, chans=hp_exp['chans'], num_pool_layers=hp_exp['num_pool_layers'], drop_prob=0.0, ).to(device) logging.info(f"Built a model consisting of {sum(p.numel() for p in model.parameters()):,} parameters")# # ------------ # trainer # ------------ optimizer, scheduler = configure_optimizers(hp_exp, model.parameters()) compute_batch_train_loss = Compute_batch_train_loss() train_meters = {'train_' + name: AverageMeter() for name in (hp_exp['loss_functions'])} if len(hp_exp['loss_functions']) > 1: train_meters['cumulated_loss'] = AverageMeter() train_meters['train_L2_gt_abs'] = AverageMeter() valid_meters = {'val_SSIM' : AverageMeter(), 'val_PSNR' : AverageMeter(), 'val_L1' : AverageMeter(), 'val_L2' : AverageMeter(), 'val_L2_kspace': AverageMeter(), 'val_L2_gt_abs': AverageMeter()} if hp_exp['two_channel_imag_real']: train_meters['train_L2_gt_comp'] = AverageMeter() valid_meters['val_L2_gt_comp'] = AverageMeter() val_ssim_fct = SSIMLoss() val_l1_fct = L1Loss(reduction='sum') val_mse_fct = MSELoss() val_mse_reduceSum_fct = MSELoss(reduction='sum') # ------------ # load a stored model if available # ------------ if hp_exp['restore_file']: load_checkpoint(hp_exp, model, optimizer, scheduler) ######### # Training save_train_figures = True save_val_figures = False mask_dict = {} for epoch in range(save_checkpoint.start_epoch, hp_exp['num_epochs']): start = time.process_time() train_bar = ProgressBar(train_loader, epoch) for meter in train_meters.values(): meter.reset() for batch_id, batch in enumerate(train_bar): hp_exp['mode'] = 'train' model.train() save_checkpoint.global_step +=1 binary_background_mask, input_image, input_kspace, input_mask, target_image, target_kspace, target_mask, target_mask_weighted, ground_truth_image, sens_maps, mean, std, fname, slice_num = batch input_image=input_image.to(device) target_image=target_image.to(device) target_kspace=target_kspace.to(device) input_kspace=input_kspace.to(device) input_mask=input_mask.to(device) target_mask=target_mask.to(device) target_mask_weighted=target_mask_weighted.to(device) ground_truth_image=ground_truth_image.to(device) sens_maps=sens_maps.to(device) mean=mean.to(device) std=std.to(device) binary_background_mask=binary_background_mask.to(device) output = model(input_image) output = output * std + mean output_tensorboard = output.detach().clone() ################################ # Compute the training loss ################################ if hp_exp['selfsup'] or hp_exp['compute_sup_loss_in_kspace']: # move complex dim to end output_per_coil_imgs = torch.moveaxis(output , 1, -1 ) output_per_coil_imgs = complex_mul(output_per_coil_imgs, sens_maps) # Transform coil images to kspace output_kspace = fft2c(output_per_coil_imgs) output_kspace = output_kspace * target_mask_weighted + 0.0 target_kspace = target_kspace * target_mask_weighted + 0.0 output_train_loss = output_kspace target_train_loss = target_kspace else: output_train_loss = output target_train_loss = target_image # Use max value per ground truth slice instead of per volume to compute ssim and psnr in image domain max_value = ground_truth_image.max().unsqueeze(0) train_loss = compute_batch_train_loss.get_batch_train_loss(hp_exp, output_train_loss, target_train_loss, max_value, train_meters) model.zero_grad() train_loss.backward() optimizer.step() ################################ # Compute train metrics that can be compared over all different setupts ################################ # Apply center cropping to outputs if necessary output_train_metrics, _ = center_crop_to_smallest(output_tensorboard, ground_truth_image) target_image_train_metrics, _ = center_crop_to_smallest(target_image, ground_truth_image) # Apply binary masking to outputs if binary masks are given binary_background_mask, _ = center_crop_to_smallest(binary_background_mask, ground_truth_image) output_train_metrics = output_train_metrics * binary_background_mask if hp_exp['two_channel_imag_real']: # target_image is already masked (if possible) loss = val_mse_reduceSum_fct(output_train_metrics, target_image_train_metrics) / torch.sum(torch.abs(target_image_train_metrics)**2) train_meters["train_L2_gt_comp"].update(loss.item()) # prepare for train_L2_gt_abs output_train_metrics = complex_abs(torch.moveaxis(output_train_metrics , 1, -1 )).unsqueeze(1) loss = val_mse_reduceSum_fct(output_train_metrics, ground_truth_image) / torch.sum(torch.abs(ground_truth_image)**2) train_meters["train_L2_gt_abs"].update(loss.item()) train_bar.log(dict(**train_meters), verbose=True) ################################ # Log some training images to tensorboard ################################ if hp_exp['tb_logging'] and fname[0] in train_log_filenames_list and epoch % hp_exp['log_image_interval'] == 0: if slice_num.item() == hp_exp['log_train_images'][fname[0]]: with torch.no_grad(): if hp_exp['two_channel_imag_real']: crop = False name = f"train_{fname[0]}_s{slice_num.item()}_ch1/"+hp_exp['exp_name'] inp = input_image[:,0,:,:].unsqueeze(1) tar = target_image[:,0,:,:].unsqueeze(1) out = output_tensorboard[:,0,:,:].unsqueeze(1) # If we want to look at center crop then crop target if crop: tar, _ = center_crop_to_smallest(tar, target_image) add_img_to_tensorboard(writer, epoch, name, inp, tar, out, val_ssim_fct, max_value, crop) name = f"train_{fname[0]}_s{slice_num.item()}_ch2/"+hp_exp['exp_name'] inp = input_image[:,1,:,:].unsqueeze(1) tar = target_image[:,1,:,:].unsqueeze(1) out = output_tensorboard[:,1,:,:].unsqueeze(1) # If we want to look at center crop then crop target if crop: tar, _ = center_crop_to_smallest(tar, target_image) add_img_to_tensorboard(writer, epoch, name, inp, tar, out, val_ssim_fct, max_value, crop) name = f"train_{fname[0]}_s{slice_num.item()}_abs/"+hp_exp['exp_name'] inp = complex_abs(torch.moveaxis(input_image , 1, -1 )).unsqueeze(1) tar = complex_abs(torch.moveaxis(target_image , 1, -1 )).unsqueeze(1) out = complex_abs(torch.moveaxis(output_tensorboard , 1, -1 )).unsqueeze(1) # If we want to look at center crop then crop target if crop: tar, _ = center_crop_to_smallest(tar, target_image) add_img_to_tensorboard(writer, epoch, name, inp, tar, out, val_ssim_fct, max_value, crop) else: name = f"train_{fname[0]}_s{slice_num.item()}_abs/"+hp_exp['exp_name'] add_img_to_tensorboard(writer, epoch, name, input_image, target_image, output_tensorboard, val_ssim_fct, max_value, crop=True) if epoch % hp_exp['val_interval'] == 0: hp_exp['mode'] = 'val' model.eval() for meter in valid_meters.values(): meter.reset() valid_bar = ProgressBar(val_loader, epoch) for sample_id, sample in enumerate(valid_bar): with torch.no_grad(): binary_background_mask, input_image, input_kspace, input_mask, target_image, target_kspace, target_mask, target_mask_weighted, ground_truth_image, sens_maps, mean, std, fname, slice_num = sample input_image=input_image.to(device) input_kspace=input_kspace.to(device) input_mask=input_mask.to(device) target_image=target_image.to(device) target_kspace=target_kspace.to(device) target_mask=target_mask.to(device) target_mask_weighted=target_mask_weighted.to(device) ground_truth_image=ground_truth_image.to(device) sens_maps=sens_maps.to(device) mean=mean.to(device) std=std.to(device) binary_background_mask=binary_background_mask.to(device) output = model(input_image) output = output * std + mean ############################################################### # Validation L1 and L2 are computed as during training if hp_exp['selfsup'] or hp_exp['compute_sup_loss_in_kspace']: # move complex dim to end output_per_coil_imgs = torch.moveaxis(output , 1, -1 ) output_per_coil_imgs = complex_mul(output_per_coil_imgs, sens_maps) # Transform coil images to kspace output_kspace = fft2c(output_per_coil_imgs) output_kspace_fully_sampled = output_kspace.clone() output_kspace = output_kspace * target_mask_weighted + 0.0 target_kspace = target_kspace * target_mask_weighted + 0.0 L2kspace = val_mse_reduceSum_fct(output_kspace, target_kspace) / torch.sum(torch.abs(target_kspace)**2) valid_meters["val_L2_kspace"].update(L2kspace.item()) # L1 and L2 validation are computed on the full images without cropping and complex absolute value and without masking or dc # L1 validation loss loss = val_l1_fct(output, target_image) / torch.sum(torch.abs(target_image)) valid_meters["val_L1"].update(loss.item()) # L2 validation loss loss = val_mse_reduceSum_fct(output, target_image) / torch.sum(torch.abs(target_image)**2) valid_meters["val_L2"].update(loss.item()) else: if hp_exp['two_channel_imag_real']: # To enable data consistency later on # move complex dim to end output_per_coil_imgs = torch.moveaxis(output , 1, -1 ) output_per_coil_imgs = complex_mul(output_per_coil_imgs, sens_maps) # Transform coil images to kspace output_kspace = fft2c(output_per_coil_imgs) #save_figure(torch.log(complex_abs(output_kspace[0,0,:,:]) + 1e-9).detach().cpu(),'output_kspace_val',hp_exp,save=save_val_figures) output_kspace_fully_sampled = output_kspace.clone() L2kspace = val_mse_reduceSum_fct(output_kspace, target_kspace) / torch.sum(torch.abs(target_kspace)**2) valid_meters["val_L2_kspace"].update(L2kspace.item()) # L1 and L2 validation are computed on the full images without cropping and complex absolute value and without masking or dc # L1 validation loss loss = val_l1_fct(output, target_image) / torch.sum(torch.abs(target_image)) valid_meters["val_L1"].update(loss.item()) # L2 validation loss loss = val_mse_reduceSum_fct(output, target_image) / torch.sum(torch.abs(target_image)**2) valid_meters["val_L2"].update(loss.item()) ############################################################### ############################################################### # Validation PSNR and SSIM are computed on masked, cropped and real images # Apply masking before computing scores in the image domain in order to eliminate artifacts in the background output = output * binary_background_mask output_tensorboard = output.clone() # PSNR and SSIM are computed on the center cropped magnitude reconstruction output, _ = center_crop_to_smallest(output, ground_truth_image) if hp_exp['two_channel_imag_real']: target_image_train_metrics, _ = center_crop_to_smallest(target_image, ground_truth_image) loss = val_mse_reduceSum_fct(output, target_image_train_metrics) / torch.sum(torch.abs(target_image_train_metrics)**2) valid_meters["val_L2_gt_comp"].update(loss.item()) # Move complex dim to end, apply complex abs, insert channel dimension output = complex_abs(torch.moveaxis(output , 1, -1 )) output = output.unsqueeze(1) # Use max value per ground truth slice instead of per volume max_value = ground_truth_image.max().unsqueeze(0) # SSIM ssim_loss = 1-val_ssim_fct(output, ground_truth_image, data_range=max_value) valid_meters["val_SSIM"].update(ssim_loss.item()) loss = val_mse_reduceSum_fct(output, ground_truth_image) / torch.sum(torch.abs(ground_truth_image)**2) valid_meters["val_L2_gt_abs"].update(loss.item()) # MSE for PSNR loss = val_mse_fct(output, ground_truth_image) # reduce with mean # PSNR psnr = 20 * torch.log10(torch.tensor(max_value.item()))- 10 * torch.log10(loss) valid_meters["val_PSNR"].update(psnr.item()) valid_bar.log(dict(**valid_meters), verbose=True) if hp_exp['tb_logging'] and fname[0] in val_log_filenames_list and epoch % hp_exp['log_image_interval'] == 0: if slice_num.item() == hp_exp['log_val_images'][fname[0]]: if hp_exp['two_channel_imag_real']: crop = False name = f"val_{fname[0]}_s{slice_num.item()}_ch1/"+hp_exp['exp_name'] inp = input_image[:,0,:,:].unsqueeze(1) tar = target_image[:,0,:,:].unsqueeze(1) out = output_tensorboard[:,0,:,:].unsqueeze(1) # If we want to look at center crop then crop target if crop: tar, _ = center_crop_to_smallest(tar, ground_truth_image) add_img_to_tensorboard(writer, epoch, name, inp, tar, out, val_ssim_fct, max_value, crop) name = f"val_{fname[0]}_s{slice_num.item()}_ch2/"+hp_exp['exp_name'] inp = input_image[:,1,:,:].unsqueeze(1) tar = target_image[:,1,:,:].unsqueeze(1) out = output_tensorboard[:,1,:,:].unsqueeze(1) # If we want to look at center crop then crop target if crop: tar, _ = center_crop_to_smallest(tar, ground_truth_image) add_img_to_tensorboard(writer, epoch, name, inp, tar, out, val_ssim_fct, max_value, crop) name = f"val_{fname[0]}_s{slice_num.item()}_abs/"+hp_exp['exp_name'] inp = complex_abs(torch.moveaxis(input_image , 1, -1 )).unsqueeze(1) tar = complex_abs(torch.moveaxis(target_image , 1, -1 )).unsqueeze(1) out = complex_abs(torch.moveaxis(output_tensorboard , 1, -1 )).unsqueeze(1) # If we want to look at center crop then crop target if crop: tar, _ = center_crop_to_smallest(tar, ground_truth_image) add_img_to_tensorboard(writer, epoch, name, inp, tar, out, val_ssim_fct, max_value, crop) else: name = f"val_{fname[0]}_s{slice_num.item()}_abs/"+hp_exp['exp_name'] add_img_to_tensorboard(writer, epoch, name, input_image, target_image, output_tensorboard, val_ssim_fct, max_value, crop=True) if hp_exp['two_channel_imag_real']: val_metric_dict = { #keys should have the same name as the keys used to pick a training loss in hp_exp['loss_functions'] 'SSIM' : valid_meters['val_SSIM'].avg, 'L1' : valid_meters['val_L1'].avg, 'L2' : valid_meters['val_L2'].avg, 'PSNR' : valid_meters['val_PSNR'].avg, 'L2_kspace' : valid_meters['val_L2_kspace'].avg, 'L2_gt_abs' : valid_meters["val_L2_gt_abs"].avg, 'L2_gt_comp' : valid_meters["val_L2_gt_comp"].avg, } else: val_metric_dict = { #keys should have the same name as the keys used to pick a training loss in hp_exp['loss_functions'] 'SSIM' : valid_meters['val_SSIM'].avg, 'L1' : valid_meters['val_L1'].avg, 'L2' : valid_meters['val_L2'].avg, 'PSNR' : valid_meters['val_PSNR'].avg, 'L2_kspace' : valid_meters['val_L2_kspace'].avg, 'L2_gt_abs' : valid_meters["val_L2_gt_abs"].avg, } current_lr = save_checkpoint.current_lr current_best_score = save_checkpoint.best_score # Logging to tensorboard if hp_exp['tb_logging']: writer.add_scalar("lr", current_lr, epoch) writer.add_scalar("epoch", epoch, epoch) for tr_loss_name in train_meters.keys(): writer.add_scalar(tr_loss_name, train_meters[tr_loss_name].avg, epoch) for val_loss_name in val_metric_dict.keys(): writer.add_scalar('val_'+val_loss_name, val_metric_dict[val_loss_name], epoch) sys.stdout.flush() #### Learning rate decay score = val_metric_dict[hp_exp['decay_metric']] if hp_exp['lr_scheduler'] == 'MultiStepLR': scheduler.step() elif hp_exp['lr_scheduler'] == 'ReduceLROnPlateau': scheduler.step(score) else: raise ValueError('Scheduler is not defined') save_checkpoint(hp_exp, epoch, model, optimizer, scheduler, score=score) # This potentially updates save_checkpoint.best_score end = time.process_time() - start # Logging to train.log if (val_metric_dict[hp_exp['decay_metric']] < current_best_score and save_checkpoint.mode == "min") or (val_metric_dict[hp_exp['decay_metric']] > current_best_score and save_checkpoint.mode == "max"): logging.info(train_bar.print(dict(**train_meters, **valid_meters, lr=current_lr, time=np.round(end/60,3), New='Highscore'))) else: logging.info(train_bar.print(dict(**train_meters, **valid_meters, lr=current_lr, time=np.round(end/60,3)))) new_lr = optimizer.param_groups[0]["lr"] save_checkpoint.current_lr = new_lr # current lr during next epoch if hp_exp['early_stop_lr_deacy']: if (score < current_best_score and save_checkpoint.mode == "min") or (score > current_best_score and save_checkpoint.mode == "max"): save_checkpoint.best_val_current_lr_interval = score #At every lr decay check if the model did not improve during the lr_convergence_break_counter many lastt lr intervals and break if it didn't. if new_lr < current_lr: if save_checkpoint.best_val_current_lr_interval != save_checkpoint.best_score: save_checkpoint.lr_interval_counter += 1 if save_checkpoint.lr_interval_counter == hp_exp['lr_convergence_break_counter']: logging.info(f'lr decayed to {new_lr}. Break training due to convergence of val loss!') break else: logging.info(f'lr decayed to {new_lr}. lr_interval_counter increased but do not yet break due to convergence of val loss!') else: save_checkpoint.best_val_current_lr_interval = float("inf") if mode_lookup[hp_exp['decay_metric']] == "min" else float("-inf") save_checkpoint.lr_interval_counter = 0 logging.info(f'lr decayed to {new_lr}. No convergence detected. Reset lr_interval_counter.') if np.round(current_lr,10) <= hp_exp['lr_min']: if hp_exp['lr_min_break_counter'] == save_checkpoint.break_counter: logging.info('Break training due to minimal learning rate constraint!') break else: save_checkpoint.break_counter += 1 else: current_lr = save_checkpoint.current_lr if hp_exp['lr_scheduler'] == 'MultiStepLR': scheduler.step() else: raise ValueError('Scheduler is not defined') if hp_exp['tb_logging']: writer.add_scalar("lr", current_lr, epoch) writer.add_scalar("epoch", epoch, epoch) for tr_loss_name in train_meters.keys(): writer.add_scalar(tr_loss_name, train_meters[tr_loss_name].avg, epoch) new_lr = optimizer.param_groups[0]["lr"] save_checkpoint.current_lr = new_lr # current lr during next epoch end = time.process_time() - start logging.info(train_bar.print(dict(**train_meters, lr=current_lr, time=np.round(end/60,3)))) logging.info(f"Done training! Best PSNR {save_checkpoint.best_score:.5f} obtained after epoch {save_checkpoint.best_epoch}.") # return names of matrics logged to tensorboard return train_meters, val_metric_dict def main_test(hp_exp): device = torch.device('cuda') if torch.cuda.is_available() else torch.device('cpu') # ------------ # setup: # Set seeds, create directories, set path to checkpoints if available # ------------ hp_exp = setup_experiment(hp_exp) init_logging(hp_exp) # Get list of filenames saved during testing (from the test set) test_log_filenames_list = [] for k in hp_exp['save_test_images'].keys(): test_log_filenames_list.append(k) # ------------ # data # ------------ # For testing we want the target mask to be all ones. # This can be achieved either by setting self_sup=False or acceleration_total=1.0 mask_func = create_mask_for_mask_type( hp_exp['mask_type'], self_sup=False, center_fraction=hp_exp['center_fraction'], acceleration=hp_exp['acceleration'], acceleration_total=1.0 ) data_transform = UnetDataTransform(hp_exp['challenge'],mask_func=mask_func, use_seed=True, hp_exp=hp_exp, mode="test") testset = SliceDataset( dataset=hp_exp['test_set'], path_to_dataset=hp_exp['data_path'], path_to_sensmaps=hp_exp['smaps_path'], provide_senmaps=hp_exp['provide_senmaps'], challenge=hp_exp['challenge'], transform=data_transform, use_dataset_cache=True, ) test_loader = torch.utils.data.DataLoader( dataset=testset, batch_size=1, num_workers=hp_exp['num_workers'], shuffle=False, generator=torch.Generator().manual_seed(hp_exp['seed']), ) # ------------ # model # ------------ if hp_exp['two_channel_imag_real']: in_chans = 2 else: in_chans = 1 model = Unet( in_chans=in_chans, out_chans=in_chans, chans=hp_exp['chans'], num_pool_layers=hp_exp['num_pool_layers'], drop_prob=0.0, ).to(device) # ------------ # load a stored model if available # ------------ if hp_exp['restore_file']: load_checkpoint(hp_exp, model, None, None) test_ssim_fct = SSIMLoss() test_l1_fct_sum = L1Loss(reduction='sum') test_mse_fct_sum = MSELoss(reduction='sum') test_mse_fct_mean = MSELoss(reduction='mean') model.eval() test_bar = ProgressBar(test_loader, epoch=0) # Collect scores ssim_vals = [] L1_vals = [] psnr_vals = [] L2_vals = [] # Always compute both, scores after binary masking and after data consistency. ssim_vals_dc = [] L1_vals_dc = [] psnr_vals_dc = [] L2_vals_dc = [] for sample_id, sample in enumerate(test_bar): with torch.no_grad(): binary_background_mask, input_image, input_kspace, input_mask, target_image, target_kspace, target_mask, target_mask_weighted, ground_truth_image, sens_maps, mean, std, fname, slice_num = sample input_image=input_image.to(device) input_kspace=input_kspace.to(device) input_mask=input_mask.to(device) target_kspace=target_kspace.to(device) target_mask=target_mask.to(device) ground_truth_image=ground_truth_image.to(device) sens_maps=sens_maps.to(device) mean=mean.to(device) std=std.to(device) binary_background_mask=binary_background_mask.to(device) output = model(input_image) output = output * std + mean ##################### sens_maps_conj = complex_conj(sens_maps) output_per_coil_imgs = torch.moveaxis(output.clone() , 1, -1 ) output_per_coil_imgs = complex_mul(output_per_coil_imgs, sens_maps) # Transform coil images to kspace output_kspace = fft2c(output_per_coil_imgs) ################ ################ # Get scores in image domain after data consistency output_image_data_consistency = ifft2c(output_kspace* (1-input_mask) + input_kspace) output_image_data_consistency = complex_mul(output_image_data_consistency, sens_maps_conj) output_image_data_consistency = output_image_data_consistency.sum(dim=1, keepdim=False) output_image_data_consistency = torch.moveaxis(output_image_data_consistency , -1, 1) output_image_data_consistency, _ = center_crop_to_smallest(output_image_data_consistency, ground_truth_image) output_image_data_consistency = complex_abs(torch.moveaxis(output_image_data_consistency , 1, -1 )) output_image_dc = output_image_data_consistency.unsqueeze(1) # L1 loss = test_l1_fct_sum(output_image_dc, ground_truth_image) / torch.sum(torch.abs(ground_truth_image)) L1_vals_dc.append(loss.item()) # L2 loss = test_mse_fct_sum(output_image_dc, ground_truth_image) / torch.sum(torch.abs(ground_truth_image)**2) L2_vals_dc.append(loss.item()) max_value = ground_truth_image.max().unsqueeze(0) # MSE for PSNR mse = test_mse_fct_mean(output_image_dc, ground_truth_image) # PSNR psnr = 20 * torch.log10(torch.tensor(max_value.item()))- 10 * torch.log10(mse) psnr_vals_dc.append(psnr.item()) # SSIM ssim_loss = 1-test_ssim_fct(output_image_dc, ground_truth_image, data_range=max_value) ssim_vals_dc.append(ssim_loss.item()) output = output * binary_background_mask ###################### ###################### # Get scores after binary masking without data consistency # at test time L1, L2, PSNR and SSIM are all computed on center cropped magnitude values output, _ = center_crop_to_smallest(output, ground_truth_image) if hp_exp['two_channel_imag_real']: # Move complex dim to end, apply complex abs, insert channel dimension output = complex_abs(torch.moveaxis(output , 1, -1 )) output = output.unsqueeze(1) # L1 loss = test_l1_fct_sum(output, ground_truth_image) / torch.sum(torch.abs(ground_truth_image)) L1_vals.append(loss.item()) # L2 loss = test_mse_fct_sum(output, ground_truth_image) / torch.sum(torch.abs(ground_truth_image)**2) L2_vals.append(loss.item()) # Normalize output to mean and std of target #target, output = normalize_to_given_mean_std(target, output) # Use max value per ground truth slice instead of per volume max_value = ground_truth_image.max().unsqueeze(0) # MSE for PSNR mse = test_mse_fct_mean(output, ground_truth_image) # PSNR psnr = 20 * torch.log10(torch.tensor(max_value.item()))- 10 * torch.log10(mse) psnr_vals.append(psnr.item()) # SSIM ssim_loss = 1-test_ssim_fct(output, ground_truth_image, data_range=max_value) ssim_vals.append(ssim_loss.item()) ###################### # Save some test images if fname[0] in test_log_filenames_list: if slice_num.item() == hp_exp['save_test_images'][fname[0]]: error = torch.abs(ground_truth_image - output) error_dc = torch.abs(ground_truth_image - output_image_dc) output = output - output.min() output = output / output.max() output_image_dc = output_image_dc - output_image_dc.min() output_image_dc = output_image_dc / output_image_dc.max() ground_truth_image = ground_truth_image - ground_truth_image.min() ground_truth_image = ground_truth_image / ground_truth_image.max() error = error - error.min() error_dc = error_dc - error_dc.min() max_norm = torch.stack([error,error_dc]).max() error = error / max_norm error_dc = error_dc / max_norm image = torch.cat([ground_truth_image, ground_truth_image, output, output_image_dc, error, error_dc], dim=0) image = torchvision.utils.make_grid(image, nrow=2, normalize=False, value_range=(0,1), pad_value=1) figure = get_figure(image.cpu().numpy(),figsize=(8,12),title=f"ssim={ssim_loss.item():.6f}, ssim_dc={ssim_vals_dc[-1]:.6f}") if not os.path.isdir(hp_exp['log_path'] + 'test_imgs/'): os.mkdir(hp_exp['log_path'] + 'test_imgs/') plt.savefig(hp_exp['log_path'] + f"test_imgs/{fname[0]}_s{slice_num.item()}.png", dpi='figure') plt.close() test_metric_dict = { 'ssim_m' : np.mean(np.array(ssim_vals)), 'ssim_s' : np.std(np.array(ssim_vals)), 'L1_m' : np.mean(np.array(L1_vals)), 'L1_s' : np.std(np.array(L1_vals)), 'psnr_m' : np.mean(np.array(psnr_vals)), 'psnr_s' : np.std(np.array(psnr_vals)), 'L2_m' : np.mean(np.array(L2_vals)), 'L2_s' : np.std(np.array(L2_vals)), } print(test_metric_dict) test_metric_dict_dc = { 'ssim_m' : np.mean(np.array(ssim_vals_dc)), 'ssim_s' : np.std(np.array(ssim_vals_dc)), 'L1_m' : np.mean(np.array(L1_vals_dc)), 'L1_s' : np.std(np.array(L1_vals_dc)), 'psnr_m' : np.mean(np.array(psnr_vals_dc)), 'psnr_s' : np.std(np.array(psnr_vals_dc)), 'L2_m' : np.mean(np.array(L2_vals_dc)), 'L2_s' : np.std(np.array(L2_vals_dc)), } print(test_metric_dict_dc) testset_name = hp_exp['log_path'] + hp_exp['test_set'][hp_exp['test_set'].find('datasets/')+9 : hp_exp['test_set'].find('.yaml')] pickle.dump( test_metric_dict, open(testset_name + '_metrics_' + hp_exp['resume_from_which_checkpoint'] + '.p', "wb" ) ) pickle.dump( test_metric_dict_dc, open(testset_name + '_metrics_DC_' + hp_exp['resume_from_which_checkpoint'] + '.p', "wb" ) ) logging.info("Evaluate testset : {}".format(testset_name)) for test_metric in test_metric_dict.keys(): logging.info("{}: {}".format(test_metric, test_metric_dict[test_metric])) logging.info("Evaluate testset : {} with data consistency".format(testset_name)) for test_metric in test_metric_dict_dc.keys(): logging.info("{}: {}".format(test_metric, test_metric_dict_dc[test_metric]))
40,715
45.961938
214
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/CS_accelerated_MRI_figure5/functions/coil_combine.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ import torch from functions.math import complex_abs_sq def rss(data: torch.Tensor, dim: int = 0) -> torch.Tensor: """ Compute the Root Sum of Squares (RSS). RSS is computed assuming that dim is the coil dimension. Args: data: The input tensor dim: The dimensions along which to apply the RSS transform Returns: The RSS value. """ return torch.sqrt((data ** 2).sum(dim)) def rss_complex(data: torch.Tensor, dim: int = 0) -> torch.Tensor: """ Compute the Root Sum of Squares (RSS) for complex inputs. RSS is computed assuming that dim is the coil dimension. Args: data: The input tensor dim: The dimensions along which to apply the RSS transform Returns: The RSS value. """ return torch.sqrt(complex_abs_sq(data).sum(dim))
1,015
22.627907
66
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/CS_accelerated_MRI_figure5/functions/math.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ import numpy as np import torch def complex_mul(x: torch.Tensor, y: torch.Tensor) -> torch.Tensor: """ Complex multiplication. This multiplies two complex tensors assuming that they are both stored as real arrays with the last dimension being the complex dimension. Args: x: A PyTorch tensor with the last dimension of size 2. y: A PyTorch tensor with the last dimension of size 2. Returns: A PyTorch tensor with the last dimension of size 2. """ if not x.shape[-1] == y.shape[-1] == 2: raise ValueError("Tensors do not have separate complex dim.") re = x[..., 0] * y[..., 0] - x[..., 1] * y[..., 1] im = x[..., 0] * y[..., 1] + x[..., 1] * y[..., 0] return torch.stack((re, im), dim=-1) def complex_conj(x: torch.Tensor) -> torch.Tensor: """ Complex conjugate. This applies the complex conjugate assuming that the input array has the last dimension as the complex dimension. Args: x: A PyTorch tensor with the last dimension of size 2. y: A PyTorch tensor with the last dimension of size 2. Returns: A PyTorch tensor with the last dimension of size 2. """ if not x.shape[-1] == 2: raise ValueError("Tensor does not have separate complex dim.") return torch.stack((x[..., 0], -x[..., 1]), dim=-1) def complex_abs(data: torch.Tensor) -> torch.Tensor: """ Compute the absolute value of a complex valued input tensor. Args: data: A complex valued tensor, where the size of the final dimension should be 2. Returns: Absolute value of data. """ if not data.shape[-1] == 2: raise ValueError("Tensor does not have separate complex dim.") return (data ** 2).sum(dim=-1).sqrt() def complex_abs_sq(data: torch.Tensor) -> torch.Tensor: """ Compute the squared absolute value of a complex tensor. Args: data: A complex valued tensor, where the size of the final dimension should be 2. Returns: Squared absolute value of data. """ if not data.shape[-1] == 2: raise ValueError("Tensor does not have separate complex dim.") return (data ** 2).sum(dim=-1) def tensor_to_complex_np(data: torch.Tensor) -> np.ndarray: """ Converts a complex torch tensor to numpy array. Args: data: Input data to be converted to numpy. Returns: Complex numpy version of data. """ data = data.numpy() return data[..., 0] + 1j * data[..., 1]
2,728
25.754902
77
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/CS_accelerated_MRI_figure5/functions/log_save_image_utils.py
import matplotlib.pyplot as plt import torchvision import io import torch import numpy as np def plot_to_image(figure): """Converts the matplotlib plot specified by 'figure' to a PNG image and returns it. The supplied figure is closed and inaccessible after this call.""" # Save the plot to a PNG in memory. buf = io.BytesIO() plt.savefig(buf, format='png') # Closing the figure prevents it from being displayed directly inside # the notebook. plt.close(figure) buf.seek(0) # Convert PNG buffer to TF image #img = buf.getvalue()#.to(torch.uint8) #img = torch.import_ir_module_from_buffer(img,dtype=torch.uint8) frameTensor = torch.tensor(np.frombuffer(buf.getvalue(), dtype=np.uint8), device='cpu') image = torchvision.io.decode_png(frameTensor) # Add the batch dimension #image = torch.unsqueeze(image, 0) return image def get_figure(image,figsize,title): """Return a matplotlib figure of a given image.""" if len(image.shape) != 3: raise ValueError("Image dimensions not suitable for logging to tensorboard.") if image.shape[0] == 1 or image.shape[0] == 3: image = np.rollaxis(image,0,3) # Create a figure to contain the plot. if figsize: figure = plt.figure(figsize=figsize) else: figure = plt.figure() # Start next subplot. plt.subplot(1, 1, 1, title=title) plt.xticks([]) plt.yticks([]) plt.grid(False) plt.imshow(image, cmap='gray') figure.tight_layout() return figure
1,533
33.088889
91
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/CS_accelerated_MRI_figure5/functions/train_utils.py
import torch import numpy as np import random import os import glob import logging from torch.serialization import default_restore_location from tensorboard.backend.event_processing import event_accumulator import time import matplotlib.pyplot as plt def setup_experiment(hp_exp): ''' - Handle seeding - Create directories - Look for checkpoints to load from ''' torch.backends.cudnn.deterministic = True torch.backends.cudnn.benchmark = False torch.manual_seed(hp_exp['seed']) torch.cuda.manual_seed(hp_exp['seed']) np.random.seed(hp_exp['seed']) random.seed(hp_exp['seed']) hp_exp['log_path'] = './'+ hp_exp['exp_name'] + '/log_files/' os.makedirs(hp_exp['log_path'] + 'checkpoints/', exist_ok=True) hp_exp['log_file'] = os.path.join(hp_exp['log_path'], "train.log") # Look for checkpoints to load from available_models = glob.glob(hp_exp['log_path'] + 'checkpoints/*.pt') if available_models and hp_exp['resume_from_which_checkpoint']=='last': hp_exp['restore_file'] = hp_exp['log_path'] + 'checkpoints/checkpoint_last.pt' elif available_models and hp_exp['resume_from_which_checkpoint']=='best': hp_exp['restore_file'] = hp_exp['log_path'] + 'checkpoints/checkpoint_best.pt' else: hp_exp['restore_file'] = None # Set attributes of the function save_checkpoint. They will be used to track the validation score and trigger saving a checkpoint mode_lookup = { 'SSIM' : 'max', 'PSNR' : 'max', 'L1' : 'min', 'L2' : 'min', 'MSE' : 'min', 'L2_kspace' : 'min', 'L1_kspace' : 'min', } save_checkpoint.best_epoch = -1 save_checkpoint.last_epoch = 0 save_checkpoint.start_epoch = 0 save_checkpoint.global_step = 0 save_checkpoint.current_lr = hp_exp['lr'] save_checkpoint.break_counter = 0 save_checkpoint.best_val_current_lr_interval = float("inf") if mode_lookup[hp_exp['decay_metric']] == "min" else float("-inf") save_checkpoint.lr_interval_counter = 0 save_checkpoint.mode = mode_lookup[hp_exp['decay_metric']] save_checkpoint.best_score = float("inf") if save_checkpoint.mode == "min" else float("-inf") return hp_exp def init_logging(hp_exp): for handler in logging.root.handlers[:]: logging.root.removeHandler(handler) handlers = [logging.StreamHandler()] mode = "a" if hp_exp['restore_file'] else "w" handlers.append(logging.FileHandler(hp_exp['log_file'], mode=mode)) logging.basicConfig(handlers=handlers, format="[%(asctime)s] %(message)s", datefmt="%Y-%m-%d %H:%M:%S", level=logging.INFO) if hp_exp['mode'] == 'train': logging.info("Arguments: {}".format(hp_exp)) def save_checkpoint(hp_exp, epoch, model, optimizer=None, scheduler=None, score=None): '''' This function is used to save a range of parameters related to the training progress. Saving those parameters allows to interrupt and then pick up training later at any point. At the beginning of every experiment the parameters are initialized in setup_experiment() Parameters: - best_score: Holds the best validation score so far - best_epoch: Holds the epoch in which the best validation score was achieved - last_epoch: Holds the current epoch. - break_counter: Count the number of epochs with minimal lr - best_val_current_lr_interval: Holds the best val performance for the current lr-inerval - lr_interval_counter: Counts for how many lr intervals there was no improvement - ''' save_checkpoint.last_epoch = epoch best_score = save_checkpoint.best_score if (score < best_score and save_checkpoint.mode == "min") or (score > best_score and save_checkpoint.mode == "max"): save_checkpoint.best_epoch = epoch save_checkpoint.best_score = score model = [model] if model is not None and not isinstance(model, list) else model optimizer = [optimizer] if optimizer is not None and not isinstance(optimizer, list) else optimizer scheduler = [scheduler] if scheduler is not None and not isinstance(scheduler, list) else scheduler state_dict = { "last_step": save_checkpoint.global_step, #set "last_score": score, #set "break_counter": save_checkpoint.break_counter, "best_val_current_lr_interval": save_checkpoint.best_val_current_lr_interval, "lr_interval_counter": save_checkpoint.lr_interval_counter, "last_epoch": save_checkpoint.last_epoch, #set "best_epoch": save_checkpoint.best_epoch, #set "current_lr":save_checkpoint.current_lr, #set "mode": save_checkpoint.mode, "best_score": getattr(save_checkpoint, "best_score", None), #set "model": [m.state_dict() for m in model] if model is not None else None, "optimizer": [o.state_dict() for o in optimizer] if optimizer is not None else None, "scheduler": [s.state_dict() for s in scheduler] if scheduler is not None else None, "args": hp_exp, } torch.save(state_dict, os.path.join(hp_exp['log_path'] + 'checkpoints/', "checkpoint_last.pt")) if hp_exp['epoch_checkpoints']: if epoch in hp_exp['epoch_checkpoints']: torch.save(state_dict, os.path.join(hp_exp['log_path'] + 'checkpoints/', "checkpoint{}.pt".format(epoch))) if (score < best_score and save_checkpoint.mode == "min") or (score > best_score and save_checkpoint.mode == "max"): torch.save(state_dict, os.path.join(hp_exp['log_path'] + 'checkpoints/', "checkpoint_best.pt")) def load_checkpoint(hp_exp, model=None, optimizer=None, scheduler=None): print('restoring model..') state_dict = torch.load(hp_exp['restore_file'], map_location=lambda s, l: default_restore_location(s, "cpu")) save_checkpoint.last_epoch = state_dict["last_epoch"] save_checkpoint.start_epoch = state_dict["last_epoch"]+1 save_checkpoint.global_step = state_dict["last_step"] save_checkpoint.best_score = state_dict["best_score"] save_checkpoint.best_epoch = state_dict["best_epoch"] save_checkpoint.break_counter = state_dict["break_counter"] save_checkpoint.best_val_current_lr_interval = state_dict["best_val_current_lr_interval"] save_checkpoint.lr_interval_counter = state_dict["lr_interval_counter"] save_checkpoint.current_lr = state_dict["current_lr"] save_checkpoint.mode = state_dict["mode"] model = [model] if model is not None and not isinstance(model, list) else model optimizer = [optimizer] if optimizer is not None and not isinstance(optimizer, list) else optimizer scheduler = [scheduler] if scheduler is not None and not isinstance(scheduler, list) else scheduler if model is not None and state_dict.get("model", None) is not None: for m, state in zip(model, state_dict["model"]): m.load_state_dict(state) if optimizer is not None and state_dict.get("optimizer", None) is not None: for o, state in zip(optimizer, state_dict["optimizer"]): o.load_state_dict(state) if scheduler is not None and state_dict.get("scheduler", None) is not None: for s, state in zip(scheduler, state_dict["scheduler"]): #milestones = s.milestones #state['milestones'] = milestones s.load_state_dict(state) #s.milestones = milestones logging.info("Loaded checkpoint {} from best_epoch {} last_epoch {}".format(hp_exp['restore_file'], save_checkpoint.best_epoch, save_checkpoint.last_epoch))
7,597
44.771084
160
py
sample_complexity_ss_recon
sample_complexity_ss_recon-main/CS_accelerated_MRI_figure5/functions/fftc.py
""" Copyright (c) Facebook, Inc. and its affiliates. This source code is licensed under the MIT license found in the LICENSE file in the root directory of this source tree. """ from typing import List, Optional import torch from packaging import version if version.parse(torch.__version__) >= version.parse("1.7.0"): import torch.fft # type: ignore def fft2c_old(data: torch.Tensor) -> torch.Tensor: """ Apply centered 2 dimensional Fast Fourier Transform. Args: data: Complex valued input data containing at least 3 dimensions: dimensions -3 & -2 are spatial dimensions and dimension -1 has size 2. All other dimensions are assumed to be batch dimensions. Returns: The FFT of the input. """ if not data.shape[-1] == 2: raise ValueError("Tensor does not have separate complex dim.") data = ifftshift(data, dim=[-3, -2]) data = torch.fft(data, 2, normalized=True) data = fftshift(data, dim=[-3, -2]) return data def ifft2c_old(data: torch.Tensor) -> torch.Tensor: """ Apply centered 2-dimensional Inverse Fast Fourier Transform. Args: data: Complex valued input data containing at least 3 dimensions: dimensions -3 & -2 are spatial dimensions and dimension -1 has size 2. All other dimensions are assumed to be batch dimensions. Returns: The IFFT of the input. """ if not data.shape[-1] == 2: raise ValueError("Tensor does not have separate complex dim.") data = ifftshift(data, dim=[-3, -2]) data = torch.ifft(data, 2, normalized=True) data = fftshift(data, dim=[-3, -2]) return data def fft2c_new(data: torch.Tensor) -> torch.Tensor: """ Apply centered 2 dimensional Fast Fourier Transform. Args: data: Complex valued input data containing at least 3 dimensions: dimensions -3 & -2 are spatial dimensions and dimension -1 has size 2. All other dimensions are assumed to be batch dimensions. Returns: The FFT of the input. """ if not data.shape[-1] == 2: raise ValueError("Tensor does not have separate complex dim.") data = ifftshift(data, dim=[-3, -2]) data = torch.view_as_real( torch.fft.fftn( # type: ignore torch.view_as_complex(data), dim=(-2, -1), norm="ortho" ) ) data = fftshift(data, dim=[-3, -2]) return data def ifft2c_new(data: torch.Tensor) -> torch.Tensor: """ Apply centered 2-dimensional Inverse Fast Fourier Transform. Args: data: Complex valued input data containing at least 3 dimensions: dimensions -3 & -2 are spatial dimensions and dimension -1 has size 2. All other dimensions are assumed to be batch dimensions. Returns: The IFFT of the input. """ if not data.shape[-1] == 2: raise ValueError("Tensor does not have separate complex dim.") data = ifftshift(data, dim=[-3, -2]) data = torch.view_as_real( torch.fft.ifftn( # type: ignore torch.view_as_complex(data), dim=(-2, -1), norm="ortho" ) ) data = fftshift(data, dim=[-3, -2]) return data # Helper functions def roll_one_dim(x: torch.Tensor, shift: int, dim: int) -> torch.Tensor: """ Similar to roll but for only one dim. Args: x: A PyTorch tensor. shift: Amount to roll. dim: Which dimension to roll. Returns: Rolled version of x. """ shift = shift % x.size(dim) if shift == 0: return x left = x.narrow(dim, 0, x.size(dim) - shift) right = x.narrow(dim, x.size(dim) - shift, shift) return torch.cat((right, left), dim=dim) def roll( x: torch.Tensor, shift: List[int], dim: List[int], ) -> torch.Tensor: """ Similar to np.roll but applies to PyTorch Tensors. Args: x: A PyTorch tensor. shift: Amount to roll. dim: Which dimension to roll. Returns: Rolled version of x. """ if len(shift) != len(dim): raise ValueError("len(shift) must match len(dim)") for (s, d) in zip(shift, dim): x = roll_one_dim(x, s, d) return x def fftshift(x: torch.Tensor, dim: Optional[List[int]] = None) -> torch.Tensor: """ Similar to np.fft.fftshift but applies to PyTorch Tensors Args: x: A PyTorch tensor. dim: Which dimension to fftshift. Returns: fftshifted version of x. """ if dim is None: # this weird code is necessary for toch.jit.script typing dim = [0] * (x.dim()) for i in range(1, x.dim()): dim[i] = i # also necessary for torch.jit.script shift = [0] * len(dim) for i, dim_num in enumerate(dim): shift[i] = x.shape[dim_num] // 2 return roll(x, shift, dim) def ifftshift(x: torch.Tensor, dim: Optional[List[int]] = None) -> torch.Tensor: """ Similar to np.fft.ifftshift but applies to PyTorch Tensors Args: x: A PyTorch tensor. dim: Which dimension to ifftshift. Returns: ifftshifted version of x. """ if dim is None: # this weird code is necessary for toch.jit.script typing dim = [0] * (x.dim()) for i in range(1, x.dim()): dim[i] = i # also necessary for torch.jit.script shift = [0] * len(dim) for i, dim_num in enumerate(dim): shift[i] = (x.shape[dim_num] + 1) // 2 return roll(x, shift, dim)
5,535
25.236967
80
py