This repository is for SPSRC method introduced in the following paper accepted by WACV2024:
Towards Better Structured Pruning Saliency by Reorganizing Convolution
Xinglong Sun, Humphrey Shi
We present SPSRC, a novel, simple and effective framework to extract enhanced Structured Pruning Saliency scores by Reorganizing Convolution. We observe that performance of pruning methods have gradually plateaued recently and propose to make better use of the learned convolutional kernel weights simply after a few steps of transformations. We firstly re-organize the convolutional operations between layers as matrix multiplications and then use the singular values and the matrix norms of the transformed matrices as saliency scores to decide what channels to prune or keep. We show both analytically and empirically that the long-standing kernel-norm-based channel importance measurement, like L1 magnitude, is not precise enough possessing a very obvious weakness of lacking spatial saliency but can be improved with SPSRC. We conduct extensive experiments across different settings and configurations and compare with the counterparts without our SPSRC along with other popular methods, observing obvious improvements. Our code is available at: https://github.com/AlexSunNik/SPSRC/tree/main.
Overview of our method.
In order to compute the importance scores following our proposed reorganization steps, run the compute_saliency.py file as follows:
python3 compute_saliency.py --model [MODEL_ARCH] --data-set [DATASET]
The script will compute all of Spectral, Nuclear, and Frobenious norm for the specified model architecture and dataset.
If you want to learn detail of how the transformation and the reorganization from Convolution to Matrix Multiplication is done, check the functions: conv_to_mat(.), calculate_eigvs(.), calculate_nucs(.), and calculate_fros(.) from reconv.py.
We also provide collected scores by us in the saliency folder
If you need to prune and finetune the model following the measured saliency scores, run one of the prune_*.py files. For example, to prune ResNet34, you do:
python3 prune_resnet34.py --metric [METRIC] --model [MODEL_ARCH] --data-set [DATASET] --prune-cfg [PRUNE_CFG] --save-path [SAVE_PATH]
METRIC: Pick one from [spec, nuc, fro]. Choose saliency from Spectral Norm, Nuclear Norm, or Frobenious Norm after reorganization.
MODEL_ARCH: Pick one from [vgg16_bn, resnet34, resnet56, resnet101]
DATASET Pick one from [imagenet, CIFAR10, CIFAR100]
SAVE_PATH Specify where you want to save the pruned model
Notice that finetuning is involved when you run the above files for ResNet56, ResNet101, and VGG for the CIFAR datasets.
In order to finetune ResNet34 on ImageNet, you need to additionally run:
python3 finetune_resnet34.py --refine [PATH_TO_PRUNED_MODEL]
If you find this repo useful to your project or research, please cite our paper below:
@inproceedings{sun2024towards,
title={Towards Better Structured Pruning Saliency by Reorganizing Convolution},
author={Sun, Xinglong and Shi, Humphrey},
booktitle={Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision},
pages={2204--2214},
year={2024}
}