Image Classification
Keras
LiteRT
TF-Keras
Safetensors
English
efficientnetv2-s
efficientnetv2
fgic
transfer-learning
gem-pooling
focal-loss
swa
grad-cam
calibration
temperature-scaling
computer-vision
tensorflow.js
Eval Results (legacy)
Instructions to use 0xgr3y/Arch-Building-Image-Classification with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Keras
How to use 0xgr3y/Arch-Building-Image-Classification with Keras:
# Available backend options are: "jax", "torch", "tensorflow". import os os.environ["KERAS_BACKEND"] = "jax" import keras model = keras.saving.load_model("hf://0xgr3y/Arch-Building-Image-Classification") - Notebooks
- Google Colab
- Kaggle
Upload config.yaml with huggingface_hub
Browse files- config.yaml +123 -0
config.yaml
ADDED
|
@@ -0,0 +1,123 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Configuration Template — Arch-Building-Image-Classification
|
| 2 |
+
# File ini untuk referensi dokumentasi hyperparameter.
|
| 3 |
+
# File ini dapat diintegrasikan di masa depan untuk modular pipeline.
|
| 4 |
+
|
| 5 |
+
dataset:
|
| 6 |
+
source: huggingface
|
| 7 |
+
repo_id: 0xgr3y/arch-building-dataset
|
| 8 |
+
num_classes: 8
|
| 9 |
+
labels: [barn, bridge, castle, mosque, skyscraper, stadium, temple, windmill]
|
| 10 |
+
total_images: 13440
|
| 11 |
+
images_per_class: 1680
|
| 12 |
+
split_ratio: [0.8, 0.1, 0.1]
|
| 13 |
+
split_seed: 42
|
| 14 |
+
input_shape: [320, 320, 3]
|
| 15 |
+
batch_size: 32
|
| 16 |
+
|
| 17 |
+
augmentation:
|
| 18 |
+
rotation_range: 15
|
| 19 |
+
width_shift_range: 0.1
|
| 20 |
+
height_shift_range: 0.1
|
| 21 |
+
shear_range: 0.1
|
| 22 |
+
zoom_range: 0.20
|
| 23 |
+
brightness_range: [0.75, 1.15]
|
| 24 |
+
channel_shift_range: 10.0
|
| 25 |
+
horizontal_flip: true
|
| 26 |
+
fill_mode: nearest
|
| 27 |
+
mixup_alpha: 0.2
|
| 28 |
+
cutmix_alpha: 1.0
|
| 29 |
+
mixup_cutmix_prob: 0.5
|
| 30 |
+
random_erasing:
|
| 31 |
+
p: 0.5
|
| 32 |
+
area_range: [0.02, 0.15]
|
| 33 |
+
aspect_ratio: [0.3, 3.3]
|
| 34 |
+
|
| 35 |
+
architecture:
|
| 36 |
+
backbone: efficientnetv2-s
|
| 37 |
+
weights: imagenet
|
| 38 |
+
include_top: false
|
| 39 |
+
custom_head:
|
| 40 |
+
conv2d_filters: 256
|
| 41 |
+
conv2d_kernel: [3, 3]
|
| 42 |
+
conv2d_activation: relu
|
| 43 |
+
maxpool_size: [2, 2]
|
| 44 |
+
gem_pooling:
|
| 45 |
+
p_init: 3.0
|
| 46 |
+
eps: 1.0e-6
|
| 47 |
+
dense_units: 256
|
| 48 |
+
dense_activation: relu
|
| 49 |
+
dropout_rate: 0.4
|
| 50 |
+
output_activation: softmax
|
| 51 |
+
output_dtype: float32
|
| 52 |
+
|
| 53 |
+
training:
|
| 54 |
+
phase1:
|
| 55 |
+
name: head_training
|
| 56 |
+
epochs_max: 25
|
| 57 |
+
learning_rate: 0.001
|
| 58 |
+
warmup_epochs: 3
|
| 59 |
+
early_stopping_patience: 5
|
| 60 |
+
optimizer: adamw
|
| 61 |
+
weight_decay: 2.0e-5
|
| 62 |
+
loss: focal
|
| 63 |
+
focal_gamma: 2.0
|
| 64 |
+
label_smoothing: 0.1
|
| 65 |
+
use_mixup_cutmix: true
|
| 66 |
+
use_ema: true
|
| 67 |
+
ema_decay: 0.999
|
| 68 |
+
phase2:
|
| 69 |
+
name: selective_fine_tuning
|
| 70 |
+
epochs_max: 50
|
| 71 |
+
learning_rate: 0.0003
|
| 72 |
+
warmup_epochs: 5
|
| 73 |
+
early_stopping_patience: 3
|
| 74 |
+
optimizer: discriminative_adamw
|
| 75 |
+
weight_decay: 2.0e-5
|
| 76 |
+
loss: focal
|
| 77 |
+
focal_gamma: 2.0
|
| 78 |
+
label_smoothing: 0.05
|
| 79 |
+
use_mixup_cutmix: false
|
| 80 |
+
use_ema: true
|
| 81 |
+
ema_decay: 0.999
|
| 82 |
+
unfreeze: [block6, top_conv]
|
| 83 |
+
freeze_bn: true
|
| 84 |
+
lr_multipliers:
|
| 85 |
+
block6: 0.1
|
| 86 |
+
swa:
|
| 87 |
+
epochs: 10
|
| 88 |
+
learning_rate: 0.0001
|
| 89 |
+
bn_update_steps: 100
|
| 90 |
+
optimizer: adamw
|
| 91 |
+
|
| 92 |
+
reproducibility:
|
| 93 |
+
seed: 42
|
| 94 |
+
enable_op_determinism: false # disabled — slows training 5-10x
|
| 95 |
+
mixed_precision: disabled # float32 chosen for reproducibility
|
| 96 |
+
python_version: "3.12.13"
|
| 97 |
+
tensorflow_version: "2.19.0"
|
| 98 |
+
keras_version: "3.13.2"
|
| 99 |
+
cuda_version: "12.5.1"
|
| 100 |
+
cudnn_version: "9.x"
|
| 101 |
+
tf_use_legacy_keras: "0"
|
| 102 |
+
|
| 103 |
+
export:
|
| 104 |
+
formats:
|
| 105 |
+
- keras_compile_false
|
| 106 |
+
- weights_h5
|
| 107 |
+
- safetensors
|
| 108 |
+
- saved_model
|
| 109 |
+
- tflite
|
| 110 |
+
- tfjs
|
| 111 |
+
build_script: build_model.py
|
| 112 |
+
|
| 113 |
+
deployment:
|
| 114 |
+
hf_model_repo: 0xgr3y/Arch-Building-Image-Classification
|
| 115 |
+
hf_space_repo: 0xgr3y/arch-building-classifier
|
| 116 |
+
hf_dataset_repo: 0xgr3y/arch-building-dataset
|
| 117 |
+
space_inference_format: tflite
|
| 118 |
+
space_inference_ms: 197.8
|
| 119 |
+
github: https://github.com/arcxteam/building-architectural-image-classifier
|
| 120 |
+
license_code: MIT
|
| 121 |
+
license_model: apache-2.0
|
| 122 |
+
license_dataset: cc-by-4.0
|
| 123 |
+
public_version: v6
|