ICE-Pruning

The offical implmentations of the experiments of paper: "ICE-Pruning: An Iterative Cost-Efficient Pruning Pipeline for Deep Neural Networks".

The following table shows the links to the code that are used/modified in this repo for the paper:

Code Repo Link	Used/Modified for
https://github.com/jdg105/linearly-replaceable-filter-pruning	ResNet-152 model
https://github.com/kuangliu/pytorch-cifar	MobileNetV2 model
https://github.com/pytorch/vision/tree/main/torchvision/models	other models in this paper
https://github.com/kaiqi123/Automatic-Attention-Pruning/tree/main	AAP
https://github.com/tyui592/Pruning_filters_for_efficient_convnets/tree/master	L1 norm filter pruning and other pruning underlying code

The folders in this repo and experiments in the paper are one-to-one correspondence:

Folder	Experiments
freezing_motivation/smaller_scale	experiments in Section IV.B
compare_pruning_criteria	experiments in Section IV.C
ablation_study	experiments in Section IV.D
SOTAs	experiments in Section IV.E

Running

1. Install the requirement.txt.
2. All the *.sh in the folders are the scripts for running the corresponding experiments. Simply run ./*.sh for each expriment you want to run.

Notes

1. We fixed the bug for multi-objective optimization in https://github.com/kaiqi123/Automatic-Attention-Pruning/tree/main
   
2. The folders or files that have names as *_im are for ImageNet related experiments, and the folders or files that have names as * dense * are for DenseNet related experiments.
   
3. The SOTA/ICE folder does not contain the code of ICE_Pruning for ResNet-152 since this experiment already exist in the compare_pruning_criteria folder.

Citation

@misc{hu2025icepruningiterativecostefficientpruning,
      title={ICE-Pruning: An Iterative Cost-Efficient Pruning Pipeline for Deep Neural Networks}, 
      author={Wenhao Hu and Paul Henderson and José Cano},
      year={2025},
      archivePrefix={arXiv},
}