sunggo / apib

Code for ICCV23 paper "Automatic network pruning via Hilbert Schmidt independence criterion lasso under information bottleneck principle"

Automatic Network Pruning via Hilbert-Schmidt Independence Criterion Lasso under Information Bottleneck Principle

Automatic Network Pruning via Hilbert-Schmidt Independence Criterion Lasso under Information Bottleneck Principle Song Guo, Lei Zhang, Xiawu Zheng, Yan Wang, Yuchao Li, Fei Chao, ShengChuan Zhang, Chenglin Wu, Rongrong Ji ICCV 2023

Model Pruning

1. VGG-16

pruning ratio (FLOPs): 66%

python main.py \
--model vgg16\
--dataset cifar10\
--target 107000000 \
--ckpt [pre-trained model dir] \
--data_path [dataset path]\
--omega 40\
--tolerance 0.01\
--alpha 5e-5

2. ResNet56

pruning ratio (FLOPs): 55%

python main.py \
--model resnet56\
--dataset cifar10\
--target 57000000 \
--ckpt [pre-trained model dir] \
--data_path [dataset path]\
--omega 5\
--tolerance 0.01\
--alpha 8e-4

3. ResNet110

pruning ratio (FLOPs): 63%

python main.py \
--model resnet110\
--dataset cifar10\
--target 96000000 \
--ckpt [pre-trained model dir] \
--data_path [dataset path]\
--omega 5\
--tolerance 0.01\
--alpha 8e-9

4. GoogLeNet

pruning ratio (FLOPs): 63%

python main.py \
--model googlenet\
--dataset cifar10\
--target 568000000 \
--ckpt [pre-trained model dir] \
--data_path [dataset path]\
--omega 9\
--tolerance 0.01\
--alpha 4e-8

5. ResNet50

pruning ratio (FLOPs): 62%

python main.py \
--model resnet50\
--dataset imagenet\
--target 1550000000 \
--ckpt [pre-trained model dir] \
--data_path [dataset path]\
--omega 1\
--tolerance 0.01\
--alpha 7e-5

Model Training

1. VGG-16

python train.py \
--model vgg16\
--dataset cifar10\
--lr 0.1\
--batch_size 256 \
--ckpt_path [pruned model dir]\
--data_path [dataset path]

2. ResNet-50

python train.py \
--model resnet50\
--dataset imagenet\
--lr 0.01\
--batch_size 128 \
--ckpt_path [pruned model dir]\
--data_path [dataset path]

Pre-trained Models

Additionally, we provide the pre-trained models used in our experiments.

CIFAR-10:

Vgg-16 | ResNet56 | ResNet110
| GoogLeNet

ImageNet:

ResNet50

Acknowledgments

Our implementation partially reuses Lasso's code | HRank's code | ITPruner's code.

apib's Issues

询问剪枝后精度

您好，非常感谢您的工作和开源代码，在测试代码过程时，由于main.py文件最后没有测试剪枝后的模型精度，我们在最后一行中添加了HSIClassopruner.metric()，但是我们发现得到的结果并不正确（vgg16在CIFAR10上的参数配置，top1 10% top5 50%)，请问如果想要测试剪枝后的模型精度，我们该怎么做？