Classification CNNs in TensorFlow

Configuration

• Edit model image sizes, model type, classes, epochs, etc:
  configuration.py

Supported Networks

0: MobileNet-v1, 1: MobileNet-v2, 2: MobileNet-v3-Large, 3: MobileNet-v3-Small
4: EfficientNet-B0, 5: EfficientNet-B1, 6: EfficientNet-B2, 7: EfficientNet-B3
8: EfficientNet-B4, 9: EfficientNet-B5, 10: EfficientNet-B6, 11: EfficientNet-B7
12: ResNeXt50, 13: ResNeXt101
14: InceptionV4, 15: InceptionResNetV1, 16: InceptionResNetV2
17: SE_ResNet_50, 18: SE_ResNet_101, 19: SE_ResNet_152
20: SqueezeNet
21: DenseNet_121, 22: DenseNet_169, 23: DenseNet_201, 24: DenseNet_269
25: ShuffleNetV2-0.5x, 26: ShuffleNetV2-1.0x, 27: ShuffleNetV2-1.5x, 28: ShuffleNetV2-2.0x
29: ResNet_18, 30: ResNet_34, 31: ResNet_50, 32: ResNet_101, 33: ResNet_152
34: SEResNeXt_50, 35: SEResNeXt_101
36: RegNet

Training & Evaluation

1. Prepare Data

• For monolithic datasets:

  • Place input data into /dataset with subfolder name as class name
  • Automatically split dataset into
    train/train, train/valid, train/test

  python src/split.py

• For pre-split datasets:

  • Place input data into
    train/train, train/valid, train/test

• Tesize images and generate binary records
  train/train.tfrecord, train/valid.tfrecord, train/test.tfrecord

  python src/prepare.py

2. Edit Configuration

src/configuration.py

• NUM_CLASSES must match number from previous steps
• MODEL choose model network architecture

3. Train Model

python src/train.py

• Training saves checkpoint saved/epoch-* every SAVE_N_EPOCH
• End of training saves final checkpoint as saved/model/*
• End of training exports model to saved/saved_model.pb

4. Evaluate Model Precision

Evaluate the model on the test dataset

python src/evalulate.py

Test on Single Image

python src/predict.py [filename]

Convert to TFJS Graph Model

Example: Quantize to F16 and save to saved/graph

tensorflowjs_converter --input_format=tf_saved_model --output_format=tfjs_graph_model
--strip_debug_ops=* --weight_shard_size_bytes=1073741824
--quantize_float16=* --control_flow_v2=true
saved/ saved/graph/

Optionally check model signature

Install TFJS-Utils

node signature.js saved/graph

2021-09-27 18:01:01 DATA:  created on: 2021-09-27T21:54:38.648Z
2021-09-27 18:01:01 INFO:  graph model: /home/vlado/dev/tf-cnn-classification/saved/graph/model.json
2021-09-27 18:01:01 INFO:  size: { numTensors: 51, numDataBuffers: 51, numBytes: 12780728 }
2021-09-27 18:01:01 DATA:  ops used by model: {
  graph: [ 'Const', 'Placeholder', 'Shape', 'Identity', [length]: 4 ],
  convolution: [ '_FusedConv2D', 'DepthwiseConv2dNative', 'AvgPool', [length]: 3 ],
  slice_join: [ 'GatherV2', 'ConcatV2', 'Pack', [length]: 3 ],
  reduction: [ 'Prod', [length]: 1 ],
  transformation: [ 'Reshape', [length]: 1 ],
  matrices: [ 'MatMul', [length]: 1 ],
  arithmetic: [ 'BiasAdd', [length]: 1 ],
  normalization: [ 'Softmax', [length]: 1 ]
}
2021-09-27 18:01:01 DATA:  inputs: [ { name: 'input_1', dtype: 'DT_FLOAT', shape: [ -1, 224, 224, 3 ] } ]
2021-09-27 18:01:01 DATA:  outputs: [ { id: 0, name: 'output_1', dytpe: 'DT_FLOAT', shape: [ -1, 1, 1, 4 ] } ]

---

References

Papers

• MobileNet_V1
• MobileNet_V2
• MobileNet_V3
• EfficientNet
• ResNeXt
• Inception_V4/Inception_ResNet_V1/Inception_ResNet_V2
• SENet
• SqueezeNet
• DenseNet
• ShuffleNetV2
• ResNet
• RegNet

Implementations

• MobileNet_V3
• EfficientNet
• ResNeXt
• InceptionV4
• DenseNet
• ResNet
• AlexNet and VGG
• InceptionV3
• ResNet