This repository is an extension of the repository of Network Slimming (Pytorch), an official pytorch implementation of the following paper: Learning Efficient Convolutional Networks Through Network Slimming (ICCV 2017).
This repository proposes a new proximal algorithm to perform Network Slimming, where it trains the CNN towards a sparse, accurate model. As a result, fine-tuning is an optional step.
Citation:
@InProceedings{Liu_2017_ICCV,
author = {Liu, Zhuang and Li, Jianguo and Shen, Zhiqiang and Huang, Gao and Yan, Shoumeng and Zhang, Changshui},
title = {Learning Efficient Convolutional Networks Through Network Slimming},
booktitle = {The IEEE International Conference on Computer Vision (ICCV)},
month = {Oct},
year = {2017}
}
You can refer to script.sh for examples to train and prune a CNN.
The dataset argument specifies which dataset to use: cifar10 or cifar100. The arch argument specifies the architecture to use: vgg,resnet or densenet. The depth is chosen to be the same as the networks used in the paper. The s parameter is the regularization parameter for the L1 norm. The beta parameter is the quadratic penalty term.
python main.py -sr --s 0.0045 --dataset cifar10 --arch vgg --depth 19 --beta 100 --name [MODEL_NAME] --save [DIRECTORY TO SAVE MODEL]The argument percent (has to be a value between 0 and 1) is the percentage of channels to be pruned. If it is set to zero, it will remove the channels that have zero scaling factors.
python vgg_prune_analyze.py --dataset cifar10 --depth 19 --percent 0.0 --model [NAME OF MODEL TO BE PRUNED] --save [DIRECTORY TO SAVE PRUNED MODEL]The pruned model will have pruned.pth.tar at the end of its name.
python main.py --refine [PRUNED_MODEL_NAME] --dataset cifar10 --arch vgg --depth 19 --epochs 160 --name [REFINED_MODEL_NAME] --save [DIRECTORY TO SAVE MODEL]torch v0.3.1, torchvision v0.2.0