Official implementation of the ECCV'24 paper Isomorphic Pruning for Vision Models.

Similar sub-structures presents more comparable distribution of importance scores.

Group sub-structure with the same computational topology and perform ranking within each isomorphic group.

pip install -r requirements.txt

Please prepare the ImageNet dataset as follows:

data
├── imagenet
│   ├── train
│   │   ├── n01440764
│   │   ├── n01443537
│   │   ├── n01484850
│   │   ├── n01491361
│   └── val
│   │   ├── n01440764
│   │   ├── n01443537
│   │   ├── n01484850
│   │   ├── n01491361

The full list of pruned models can be found here

mkdir pretrained && cd pretrained
wget https://github.com/VainF/Isomorphic-Pruning/releases/download/v0.1/deit_4.2G_isomorphic.pth

pretrained
├── convnext_small_8.5G_isomorphic.pth
├── convnext_tiny_4.2G_isomorphic.pth
├── deit_0.6G_isomorphic.pth
├── deit_1.2G_isomorphic.pth
├── deit_2.6G_isomorphic.pth
└── deit_4.2G_isomorphic.pth

You can evaluate the pruned models using the following command:

python evaluate.py --model pretrained/deit_4.2G_isomorphic.pth --interpolation bicubic

MACs: 4.1626 G, Params: 20.6943 M
Evaluating pretrained/deit_4.2G_isomorphic.pth...
100%|███████████████| 782/782 [01:57<00:00,  6.68it/s]
Accuracy: 0.8241, Loss: 0.8036

Evaluate the performance of the pre-trained models from Timm on ImageNet validation set.

bash scripts/evaluation/deit_small_distilled_patch16_224.fb_in1k.sh

MACs: 4.6391 G, Params: 22.4364 M
Evaluating deit_small_distilled_patch16_224.fb_in1k...
100%|█████████████| 782/782 [02:00<00:00,  6.51it/s]
Accuracy: 0.8117, Loss: 0.7511

Perform isomorphic pruning on the pre-trained models. We use data-driven method to estimate the importance of parameters. It will accumulate the importance scores over multiple batches.

bash scripts/pruning/deit_4.2G.sh 

Summary:
MACs: 17.69 G => 4.17 G
Params: 87.34 M => 20.72 M
Saving the pruned model to output/pruned/deit_4.2G.pth...

Finetune the pruned model and save the intermediate/latest/best checkpoints under output/finetuned.

bash scripts/finetuning/deit_4.2G.sh

The pruned model will be saved as a .pth file with the model definition. We can directly load the .pth to obain the pruned model. You can also pass the timm model name to the script to download the pre-trained model and evaluate it.

# bilinear for ResNet and bicubic for other models
python evaluate.py --model PATH_TO_PRUNED_MODEL --interpolation bicubic 

The ConvNext models in our paper were finetuned using the official implementation. We provide an isolated script convnext_train.py for ConvNext finetuning and will unify the training scripts in the future.

This implementation is based on Timm and Torchvision Reference.

@article{fang2024isomorphic,
  title={Isomorphic Pruning for Vision Models},
  author={Fang, Gongfan and Ma, Xinyin Tan and Mi, Michael Bi and Wang, Xinchao},
  journal={arXiv preprint arXiv:},
  year={2024}
}