RMNet: Equivalently Removing Residual Connection from Networks

This repository is the official implementation of "RMNet: Equivalently Removing Residual Connection from Networks". Welcome to discuss this paper with me on 知乎

Updates

Feb 18,2022, For better understanding, we implement a simpilify version RM Operation on ResNet and MobileNetV2.

Jan 25,2022, RM+AMC purning:

https://github.com/fxmeng/RMNet/blob/aec110b528c2646a19a20777bd5b93500e9b74a3/RM+AMC/README.md

Dec 24, 2021, RMNet Pruning:

python train_pruning.py --sr xxx --threshold xxx

python train_pruning.py --eval xxx/ckpt.pth

python train_pruning.py --finetune xxx/ckpt.pth

Nov 15, 2021, RM Opeartion now supports PreActResNet.

Nov 13, 2021, RM Opeartion now supports SEBlock.

Requirements

To install requirements:

pip install torch
pip install torchvision

Training

To train the models in the paper, run this command:

python train.py -a rmrep_69 --dist-url 'tcp://127.0.0.1:23333' --dist-backend 'nccl' --multiprocessing-distributed --world-size 1 --rank 0 --workers 32 [imagenet-folder with train and val folders]

Evaluation

To evaluate our pre-trained models trained on ImageNet, run:

python train.py -a rmrep_69 -e checkpoint/rmrep_69.pth.tar [imagenet-folder with train and val folders]

Results

Our model achieves the following performance on :

Help pruning achieve better performance drive.google

Method	Speed(Imgs/Sec)	Acc(%)
Baseline	3752	71.79
AMC(0.75)	4873	70.94
AMC(0.7)	4949	70.84
AMC(0.5)	5483	68.89
RM+AMC(0.75)	5120	73.21
RM+AMC(0.7)	5238	72.63
RM+AMC(0.6)	5675	71.88
RM+AMC(0.5)	6250	71.01

Help RepVGG achieve better performance even when the depth is large

Arch	Top-1 Accuracy(%)	Top-5 Accuracy(%)	Train FLOPs(G)	Test FLOPs(M)
RepVGG-21	72.508	90.840	2.4	2.1
RepVGG-21(RM 0.25)	72.590	90.924	2.1	2.1
RepVGG-37	74.408	91.900	4.4	4.0
RepVGG-37(RM 0.25)	74.478	91.892	3.9	4.0
RepVGG-69	74.526	92.182	8.6	7.7
RepVGG-69(RM 0.5)	75.088	92.144	6.5	7.7
RepVGG-133	70.912	89.788	16.8	15.1
RepVGG-133(RM 0.75)	74.560	92.000	10.6	15.1

Image Classification on ImageNet drive.google.

Model name	Top 1 Accuracy(%)	Top 5 Accuracy(%)
RMNeXt 41x5_16	78.498	94.086
RMNeXt 50x5_32	79.076	94.444
RMNeXt 50x6_32	79.57	94.644
RMNeXt 101x6_16	80.07	94.918
RMNeXt 152x6_32	80.356	80.356

Citation

If you find this code useful, please cite the following paper:

@misc{meng2021rmnet,
      title={RMNet: Equivalently Removing Residual Connection from Networks}, 
      author={Fanxu Meng and Hao Cheng and Jiaxin Zhuang and Ke Li and Xing Sun},
      year={2021},
      eprint={2111.00687},
      archivePrefix={arXiv},
      primaryClass={cs.CV}
}

Contributing

Our code is based on RepVGG and nni/amc pruning