MobileV2 Pruning

This repository aim to try out different pruning-approaches on lightweight Backbones.

Usage

1. Training

   python main.py --arch MobileNetV2 (for l1norm pruner )
   python main.py --sr --arch MobileNetV2 (for slimming pruner) 
   python main.py --arch USMobileNetV2 (for Autoslim pruner )
   

2. Pruning (prune+finetune)

   python prune.py --arch MobileNetV2 --pruner l1normpruner --pruneratio 0.6
   python prune.py --arch MobileNetV2 --pruner SlimmingPruner --sr --pruneratio 0.6
   python prune.py --arch USMobileNetV2 --pruner AutoSlimPruner
   

Results on Cifar10

BackBone	Pruner	Prune Ratio	Original/Pruned/Finetuned Accuracy	FLOPs(M)	Params(M)
MobileV2	L1-Norm	0.6	0.937/0.100/0.844	313.5->225.5	2.24->1.15
MobileV2	Slimming	Optimal Thres	0.922/0.485/0.915	313.5->127.5	2.24->0.98
MobileV2	AutoSlim	<200 flops	0.922/0.795/0.919	313.5->137.5	2.24->1.037
VGG	Slimming	Optimal Thres	0.926/0.183/0.920	399.3->147.8	20.03->1.49
Resnet50	Slimming	Optimal Thres	0.926/0.665/0.921	3448->975	23.52->6.00
ShuffleNetV2	Slimming	Optimal Thres	0.897/0.894/0.895	348.6->188.5	2.22->1.24

NOTE:

1. args for VGG: --arch VGG --s 0.001 --sr --lr 0.02 --epochs 100
2. args for resnet: --arch resnet50 --s 0.001 --sr --lr 0.02 --epochs 100
3. args for shufflenet: --arch ShuffleNetV2 --s 0.007 --sr --lr 0.001 --epochs 100
   Try yourself with different arguments!

TODO

Pruning Methodsd

• l1-norm pruner
• Slimming pruner
• AutoSlim
• ThiNet
• Soft filter pruning
  ....

Backbones

• MobileV2
• ShuffleNet
• VGG
• ResNet
  ....

Reference

rethinking-network-pruning

Pruned-MobileNet_v2