Carrying out CNN Channel Pruning in a White Box (IEEE TNNLS 2022) (Paper Link)

Requirements

• python3.7.4, pytorch 1.5.1, torchvision 0.4.2, thop 0.0.31

Reproduce the Experiment Results

Run the following scripts to reproduce the results reported in paper (change your data path in the corresponding scripts).

• VGGNet-16-CIFAR10 ./scripts/vgg.sh
• ResNet-56-CIFAR10 ./scripts/resnet56.sh
• ResNet-110-CIFAR10 ./scripts/resnet110.sh
• MobileNet-v2-CIFAR10 ./scripts/mobilenetv2.sh
• ResNet-50-ImageNet(FLOPs:2.22B) ./scripts/resnet50-1.sh
• ResNet-50-ImageNet(FLOPs:1.50B) ./scripts/resnet50-2.sh

Evaluate Our Pruned Models

Run the following scripts to test our results reported in the paper (change your data path and input the pruned model path in the corresponding scripts. The pruned model can be downloaded from the links in the following table).

• VGGNet-16-CIFAR10 ./scripts/test-vgg.sh
• ResNet-56-CIFAR10 ./scripts/test-resnet56.sh
• ResNet-110-CIFAR10 ./scripts/test-resnet110.sh
• MobileNet-v2-CIFAR10 ./scripts/test-mobilenetv2.sh
• ResNet-50-ImageNet(FLOPs:2.22B) ./scripts/test-resnet50-1.sh
• ResNet-50-ImageNet(FLOPs:1.50B) ./scripts/test-resnet50-2.sh

CIFAR-10

Full Model	Flops ↓	Accuracy	Pruned Model
VGG16	76.4%	93.47%	Modellink
ResNet56	55.6%	93.54%	Modellink
ResNet110	66.0%	94.12%	Modellink
MobileNet-V2	29.2%	95.28%	Modellink

ImageNet

Network	Flops ↓	Top-1 Acc.	Top-5 Acc.	Pruned Model
ResNet50	45.6%	75.32%	92.43%	Modellink
ResNet50	63.5%	74.21%	92.01%	Modellink