juyongjiang/spconv.pytorch

[ IJCAI-20 ] Split to Be Slim: An Overlooked Redundancy in Vanilla Convolution

SPConv.pytorch

[ IJCAI-20, 12.9% Accepted] Split to Be Slim: An Overlooked Redundancy in Vanilla Convolution

This repo provides Pytorch implementation of IJCAI 2020 paper
Split to Be Slim: An Overlooked Redundancy in Vanilla Convolution

Requirements

Basic code is borrowed from NVIDIA DALI tutorials

• Python 3
• Pytorch 1.1
• NVIDIA DALI for GPU dataloader
• NVIDIA APEX for mixed precision

Introduction of SPConv

Redundancy in Feature Maps

Abstract

We reveal that many feature maps within a layer share similar but not identical patterns. However, it is difficult to identify if features with similar patterns are redundant or contain essential details. Therefore, instead of directly removing uncertain redundant features, we propose a split based convolutional operation, namely SPConv, to tolerate features with similar patterns but require less computation.

Specifically, we split input feature maps into the representative part and the uncertain redundant part, where intrinsic information is extracted from the representative part through relatively heavy computation while tiny hidden details in the uncertain redundant part are processed with some light-weight operation.

SPConv Module

Performance

Outperforms SOTA baselines in both accuracy and inference time on GPU, with FLOPs and parameters dropped sharply.

Small Scale Classification

CIFAR_10 - VGG_16
Model	FLOPs	FLOPs Reduced	Params	Params Reduced	Acc@1
VGG_16-Baseline	349.51M	-	16.62M	-	94.00%
Ghost-VGG_16-s2	158M	45.20%	7.7M	53.67%	93.70%
SPConv-VGG_16-α1/2	118.27M	66.24%	5.6M	66.30%	94.40%
HetConv-VGG_16-P4	105.98M	69.67%	5.17M	68.89%	93.92%
SPConv-VGG_16-α1/4	79.34M	77.29%	3.77M	77.31%	93.94%
HetConv-VGG_16-P8	76.89M	78.00%	3.54M	78.70%	93.86%
SPConv-VGG_16-α1/8	59.87M	82.87%	2.85M	82.85%	93.77%
SPConv-VGG 16-α1/16	55.14M	84.22%	2.39M	85.62%	93.43%
CIFAR_10 - ResNet_20
ResNet_20-Baseline	41.62M	-	0.27M	-	92%
SPConv-ResNet_20-α1/2	17.91M	56.96%	0.10M	63.00%	92.23%
SPConv-ResNet_20-α1/4	12.89M	75.88%	0.071M	73.70%	91.15%

Large Scale Classification

ImageNet2012-ResNet50
Model	FLOPs	FLOPs Reduced	Params	Params Reduced	Acc@1	Acc@5	Inference Time on GPU	Download
ResNet50-Baseline	4.14G	-	25.56M	-	75.91%	92.78%	1.32 ms	-
SPConv-ResNet50-α1/2	2.97G	28.26%	18.34M	28.24%	76.26%	93.05%	1.23 ms	model
HetConv-ResNet50-P4	2.85G	30.32%	-	-	76.16%		-	-
SPConv-ResNet50-α1/4	2.74G	33.82%	16.93M	33.76%	75.95%	92.99%	1.19 ms	model
SPConv-ResNet50-α1/8	2.62G	36.72%	16.22M	36.54%	75.40%	92.77%	1.17 ms	model
OctConv-ResNet50-α0.5†	2.40G	42.00%	25.56M	0.00%	76.40%	93.14%	3.51 ms	-
Ghost-ResNet50-s2	2.20G	46.85%	13.0M	49%	75%	92.3%	-	-

Citation

If you find this work or code is helpful in your research, please cite:

@inproceedings{zhang2020spconv,
  title={Split to Be Slim: An Overlooked Redundancy in Vanilla Convolution},
  author={Zhang, Qiulin and Jiang, Zhuqing and Lu, Qishuo and Han, Jia'nan and Zeng, Zhengxin and Gao, Shang-Hua and Men, Aidong},
  booktitle={International Joint Conference on Artificial Intelligence (IJCAI)},
  year={2020}
}