Channel pruning for accelerating very deep neural networks

This repo contains the PyTorch implementation for paper channel pruning for accelerating very deep neural networks.

Prune

We just support vgg-series network pruning, you can type command as follow to execute pruning.

python3 main.py --model vgg16_bn \
                --batch_size 100 \
                --calib_batch 50 \
                --calib_dir /data/imagenet1k_calib \
                --valid_dir /data/ILSVRC2012/val \
                --pruner lasso \
                --ckpt ./vgg16_bn-6c64b313.pth

type	layer	sparsity	remain channal
conv1_2	5	0.5	32
conv2_1	9	0.5	64
conv2_2	12	0.5	64
conv3_1	16	0.5	130
conv3_2	19	0.5	128
conv3_3	22	0.5	205
conv4_1	26	0.2	409
conv4_2	29	0.2	415
conv4_3	32	0.2	512

After pruning:
  Top1 acc=59.728%
  Parameter: 135.452 M
  FLOPs: 7466.797M

Finetune

python3 train.py --data_root /data/ILSVRC2012 \
                 --model vgg16_bn_x \
                 --lr 1e-03 \
                 --lr_type cos \
                 --dataset imagenet \
                 --n_epoch 30 \
                 --batch_size 128 \
                 --ckpt_path vgg16_pruned.pth.tar \
                 --seed 1

Test

python3 test.py --model vgg16_bn_x \
                --checkpoint ./vgg16_pruned_finetune.pth.tar \
                --imagenet_path /data/ILSVRC2012/

After finetuning:
  Top1 acc=73.584%, Top5=91.490%

Orig:
  Top1 acc=73.476%, Top5=91.536%