Basic_CNNs_TensorFlow2

Basic_CNNs_TensorFlow2

A tensorflow2 implementation of some basic CNNs.

Networks included:

• MobileNet_V1
• MobileNet_V2
• MobileNet_V3
• EfficientNet
• ResNeXt
• InceptionV4, InceptionResNetV1, InceptionResNetV2
• SE_ResNet_50, SE_ResNet_101, SE_ResNet_152, SE_ResNeXt_50, SE_ResNeXt_101
• SqueezeNet
• DenseNet
• ShuffleNetV2
• ResNet
• RegNet

Other networks

For AlexNet and VGG, see : https://github.com/calmisential/TensorFlow2.0_Image_Classification
For InceptionV3, see : https://github.com/calmisential/TensorFlow2.0_InceptionV3
For ResNet, see : https://github.com/calmisential/TensorFlow2.0_ResNet

Train

1. Requirements:

• Python >= 3.9
• Tensorflow >= 2.7.0
• tensorflow-addons >= 0.15.0

2. To train the network on your own dataset, you can put the dataset under the folder original dataset, and the directory should look like this:

|——original dataset
   |——class_name_0
   |——class_name_1
   |——class_name_2
   |——class_name_3

3. Run the script split_dataset.py to split the raw dataset into train set, valid set and test set. The dataset directory will be like this:

|——dataset
  |——train
       |——class_name_1
       |——class_name_2
       ......
       |——class_name_n
  |——valid
       |——class_name_1
       |——class_name_2
       ......
       |——class_name_n
  |—-test
       |——class_name_1
       |——class_name_2
       ......
       |——class_name_n

4. Run to_tfrecord.py to generate tfrecord files.
5. Change the corresponding parameters in config.py.
6. Run show_model_list.py to get the index of model.
7. Run python train.py --idx [index] to start training.
   If you want to train the EfficientNet, you should change the IMAGE_HEIGHT and IMAGE_WIDTH before training.

• b0 = (224, 224)
• b1 = (240, 240)
• b2 = (260, 260)
• b3 = (300, 300)
• b4 = (380, 380)
• b5 = (456, 456)
• b6 = (528, 528)
• b7 = (600, 600)

Evaluate

Run python evaluate.py --idx [index] to evaluate the model’s performance on the test dataset.

Different input image sizes for different neural networks

Type	Neural Network	Input Image Size (height * width)
MobileNet	MobileNet_V1	(224 * 224)
	MobileNet_V2	(224 * 224)
	MobileNet_V3	(224 * 224)
EfficientNet	EfficientNet(B0~B7)	/
ResNeXt	ResNeXt50	(224 * 224)
	ResNeXt101	(224 * 224)
SEResNeXt	SEResNeXt50	(224 * 224)
	SEResNeXt101	(224 * 224)
Inception	InceptionV4	(299 * 299)
	Inception_ResNet_V1	(299 * 299)
	Inception_ResNet_V2	(299 * 299)
SE_ResNet	SE_ResNet_50	(224 * 224)
	SE_ResNet_101	(224 * 224)
	SE_ResNet_152	(224 * 224)
SqueezeNet	SqueezeNet	(224 * 224)
DenseNet	DenseNet_121	(224 * 224)
	DenseNet_169	(224 * 224)
	DenseNet_201	(224 * 224)
	DenseNet_269	(224 * 224)
ShuffleNetV2	ShuffleNetV2	(224 * 224)
ResNet	ResNet_18	(224 * 224)
	ResNet_34	(224 * 224)
	ResNet_50	(224 * 224)
	ResNet_101	(224 * 224)
	ResNet_152	(224 * 224)

References

1. MobileNet_V1: Efficient Convolutional Neural Networks for Mobile Vision Applications
2. MobileNet_V2: Inverted Residuals and Linear Bottlenecks
3. MobileNet_V3: Searching for MobileNetV3
4. EfficientNet: EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks
5. The official code of EfficientNet: https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet
6. ResNeXt: Aggregated Residual Transformations for Deep Neural Networks
7. Inception_V4/Inception_ResNet_V1/Inception_ResNet_V2: Inception-v4, Inception-ResNet and the Impact of Residual Connectionson Learning
8. The official implementation of Inception_V4: https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_v4.py
9. The official implementation of Inception_ResNet_V2: https://github.com/tensorflow/models/blob/master/research/slim/nets/inception_resnet_v2.py
10. SENet: Squeeze-and-Excitation Networks
11. SqueezeNet: SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size
12. DenseNet: Densely Connected Convolutional Networks
13. https://zhuanlan.zhihu.com/p/37189203
14. ShuffleNetV2: ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design
15. https://zhuanlan.zhihu.com/p/48261931
16. ResNet: Deep Residual Learning for Image Recognition
17. RegNet: Designing Network Design Spaces