This repo contains some of the pruned models from paper OICSR: Out-In-Channel Sparsity Regularization for Compact Deep Neural Networks (CVPR 2019).
If you find the models useful, please kindly cite our paper:
@inproceedings{li2019oicsr,
title={OICSR: Out-In-Channel Sparsity Regularization for Compact Deep Neural Networks},
author={Li, Jiashi and Qi, Qi and Wang, Jingyu and Ge, Ce and Li, Yujian and Yue, Zhangzhang and Sun, Haifeng},
booktitle={Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition},
pages={7046--7055},
year={2019}
}
Download the pretrained models from here and put it in ./checkpoints.
We provide ResNet-50 model with various FLOPs pruned percents. The channel pruning results are showed as follows:
| Models | Top1 Acc (%) | Drop Top1 Acc (%) | Top5 Acc (%) | Drop Top5 Acc (%) | FLOPs (M) |
|---|---|---|---|---|---|
| resnet50 | 76.32 | 0.00 | 93.00 | 0.00 | 4089 |
| resnet50-37.3%FLOPs | 76.53 | -0.21 | 93.16 | -0.16 | 2563 |
| resnet50-44.4%FLOPs | 76.30 | 0.02 | 92.92 | 0.08 | 2274 |
| resnet50-50.0%FLOPs | 75.95 | 0.37 | 92.66 | 0.34 | 2046 |
To test the model, run:
python eval_prune_model.py --test_data /mnt/cephfs_wj/cv/common/datasets/ImageNet/ILSVRC2012_img_val --fpp 50.0
To contact the author:
Jiashi Li, lijiashi@bupt.edu.cn