text
stringlengths
1
93.6k
os.environ["CUDA_VISIBLE_DEVICES"] = FLAGS.gpu
dataset = FLAGS.dataset
# optimal paramters
if dataset == 'KGBD':
FLAGS.lr = '0.00035'
if FLAGS.min_samples == '':
FLAGS.min_samples = '4'
elif dataset == 'CASIA_B':
FLAGS.lr = '0.00035'
if FLAGS.min_samples == '':
FLAGS.min_samples = '2'
if FLAGS.eps == '':
FLAGS.eps = '0.75'
else:
FLAGS.lr = '0.00035'
if dataset == 'KS20' or dataset == 'IAS':
if FLAGS.eps == '':
FLAGS.eps = '0.8'
elif dataset == 'BIWI':
if FLAGS.probe == 'Walking':
if FLAGS.eps == '':
FLAGS.eps = '0.8'
T_eps = FLAGS.eps
T_min_a = FLAGS.min_samples
if FLAGS.eps == '':
FLAGS.eps = '0.6'
if FLAGS.min_samples == '':
FLAGS.min_samples = '2'
eps = float(FLAGS.eps)
min_samples = int(FLAGS.min_samples)
time_step = int(FLAGS.length)
probe = FLAGS.probe
patience = int(FLAGS.patience)
batch_size = int(FLAGS.batch_size)
# not used
global_att = False
nhood = 1
residual = False
nonlinearity = tf.nn.elu
pre_dir = 'ReID_Models/'
# Customize the [directory] to save models with different hyper-parameters
change = '_Hi-MPC_Formal'
# [directory] = [pre_dir] + [dataset] + '/' + [probe] + [change] + '/' + 'best.ckpt'
# e.g., ReID_Models/BIWI/Walking_Hi-MPC/best.ckpt
if FLAGS.probe_type != '':
change += '_CME'
try:
os.mkdir(pre_dir)
except:
pass
if dataset == 'KS20':
nb_nodes = 25
if dataset == 'CASIA_B':
nb_nodes = 14
if FLAGS.dataset == 'CASIA_B':
FLAGS.length = '40'
print('----- Model hyperparams -----')
print('batch_size: ' + str(batch_size))
print('M: ' + FLAGS.M)
print('H: ' + FLAGS.H)
print('eps: ' + FLAGS.eps)
print('min_samples: ' + FLAGS.min_samples)
print('seqence_length: ' + str(time_step))
print('patience: ' + FLAGS.patience)
print('Mode: ' + FLAGS.mode)
if FLAGS.mode == 'Train':
print('----- Dataset Information -----')
print('Dataset: ' + dataset)
if dataset == 'CASIA_B':
print('Probe.Gallery: ', FLAGS.probe_type.split('.')[0], FLAGS.probe_type.split('.')[1])
else:
print('Probe: ' + FLAGS.probe)
"""
Codes from our project of SPC-MGR
We use joint-level (J), component-level (P), and limb-level (B) skeleton data
"""
norm = True