Pruning Filter in Filter

Introduction

This is the PyTorch implementation of our NeurIPS 2020 paper "Pruning Filter in Filter". In this paper:

We propose a new pruning paradigm called Stripe-wise-Pruning(SP), which can be seen as a general case of the Filter-Pruning (FP). SP treats a filter $F \in \mathbb{R}^{C\times K\times K}$as $K\times K$ stripes (i.e., $1\times 1$ filters $\in \mathbb{R}^c $ ), and perform pruning at the unit of stripe instead of the whole filter. Compared to the existing methods, SP achieves finer granularity than traditional FP while being hardware friendly than Weight-Pruning and keeping independence between filters than Group-wise-Pruning, leading to state-of-the-art pruning ratio on CIFAR-10 and ImageNet.
More arousingly, by applying SP, we find another important property of filters alongside their weight: shape. Start with a random initialized ResNet56, we train and trim the shape of filters and boost the test accuracy to 80.58% on CIFAR-10 without updating the filter weights. The optimal shape of filters are learned by the proposed Filter-Skeleton (FS) in the paper and we believe FS could inspire further research towards the essence of network pruning.

requirements

python3.7

pytorch1.4.0

torchvision0.5.0

Baseline

mkdir -p checkpoint/VGG/baseline
python main.py --arch VGG --data_path ../data --save checkpoint/VGG/baseline
mkdir -p checkpoint/ResNet56/baseline
python main.py --arch ResNet56 --data_path ../data --save checkpoint/ResNet56/baseline

Training with Filter Skeleton & pruning Filter in Filter

mkdir -p checkpoint/VGG/sr0.00001_threshold_0.01
python main.py --arch VGG --data_path ../data --sr 0.00001 --threshold 0.01 --save checkpoint/VGG/sr0.00001_threshold_0.01
mkdir -p checkpoint/ResNet56/sr0.00001_threshold_0.01
python main.py --arch ResNet56 --data_path ../data --sr 0.00001 --threshold 0.01 --save checkpoint/ResNet56/sr0.00001_threshold_0.01

Results

Backbone Metrics Params(%) $\downarrow$ FLOPS(%) $\downarrow$ Accuracy(%)$\downarrow$

VGG16 L1 64 34.3 -0.15

ThiNet 63.95 64.02 2.49

SSS 73.8 41.6 0.23

SFP 63.95 63.91 1.17

GAL 77.6 39.6 1.22

Hinge 80.05 39.07 -0.34

HRank 82.9 53.5 -0.18

Ours 92.66 71.16 -0.4

ResNet56 L1 13.7 27.6 -0.02

CP --- 50 1.00

NISP 42.6 43.6 0.03

DCP 70.3 47.1 -0.01

IR --- 67.7 0.4

C-SGD --- 60.8 -0.23

GBN 66.7 70.3 0.03

HRank 68.1 74.1 2.38

Ours 77.7 75.6 0.12

Backbone Metrics FLOPS(%) $\downarrow$ Top-1(%) $\downarrow$ Top-5(%)$\downarrow$

ResNet18 LCCL 35.57 3.43 2.14

SFP 42.72 2.66 1.3

FPGM 42.72 1.35 0.6

TAS 43.47 0.61 -0.11

DMCP 42.81 0.56 ---

Ours($\alpha =5e-6$) 50.48 -0.23 -0.22

Ours($\alpha =2e-5$) 54.58 0.17 0.04

Backbone	Metrics	Params(%) $\downarrow$	FLOPS(%) $\downarrow$	Accuracy(%)$\downarrow$
VGG16	L1	64	34.3	-0.15
	ThiNet	63.95	64.02	2.49
	SSS	73.8	41.6	0.23
	SFP	63.95	63.91	1.17
	GAL	77.6	39.6	1.22
	Hinge	80.05	39.07	-0.34
	HRank	82.9	53.5	-0.18
	Ours	92.66	71.16	-0.4
ResNet56	L1	13.7	27.6	-0.02
	CP	---	50	1.00
	NISP	42.6	43.6	0.03
	DCP	70.3	47.1	-0.01
	IR	---	67.7	0.4
	C-SGD	---	60.8	-0.23
	GBN	66.7	70.3	0.03
	HRank	68.1	74.1	2.38
	Ours	77.7	75.6	0.12
Backbone	Metrics	FLOPS(%) $\downarrow$	Top-1(%) $\downarrow$	Top-5(%)$\downarrow$
ResNet18	LCCL	35.57	3.43	2.14
	SFP	42.72	2.66	1.3
	FPGM	42.72	1.35	0.6
	TAS	43.47	0.61	-0.11
	DMCP	42.81	0.56	---
	Ours($\alpha =5e-6$)	50.48	-0.23	-0.22
	Ours($\alpha =2e-5$)	54.58	0.17	0.04

Citation

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

@article{Meng2020PruningFI,
  title={Pruning Filter in Filter},
  author={Fan-Xu Meng and Hao Cheng and Ke Li and Huixiang Luo and Xiao-wei Guo and G. Lu and Xing Sun},
  journal={ArXiv},
  year={2020},
  volume={abs/2009.14410}
}

References

For CIFAR, our code is based on https://github.com/Eric-mingjie/rethinking-network-pruning

For ImageNet, our code is based on https://github.com/pytorch/examples/tree/master/imagenet

moranxiachong/Pruning-PFF