Network_Trimming_Pytorch

Implementation network trimming using pytorch

ImageNet

- datasets
    - imagenet - train
               - val
               - valprep.sh
- Prune_QTorch

Download : valprep.sh

./valprep.sh

How to use

compress rate	Conv 5-3	FC 6
1.00	512	4096
1.19	488	3477
1.45	451	2937
1.71	430	2479
1.96	420	2121
2.28	400	1787
2.59	390	1513

pruning

python prune.py --data_path ../datasets/imagenet \
                --save_path ./apoz_prune_model.pth.tar \
                --apoz_path ./vgg_apoz_fc.pkl \
                --select_rate 0

• pruning layer : Conv 5-3, FC 6

fine tune

python finetune.py --data_path ../datasets/imagenet \
                   --save_path ./apoz_fine_tune_model.pth.tar \
                   --prune_path ./apoz_prune_model.pth.tar \
                   --batch_size 128 \
                   --epoch 5

Benchmark

• prune

0 : 488, 3477

Before Pruning

Acc@1: 71.59 
Acc@5: 90.38

After Pruning

Acc@1: 70.37
Acc@5: 89.76

• finetune

Conv 5-3 : 512 -> 488
FC 6 : 4096 -> 3477

Before Fine tune

Acc@1: 70.37
Acc@5: 89.76

After Fine tune

Acc@1: 71.48
Acc@5: 90.26

Reference

• Network Trimming: A Data-Driven Neuron Pruning Approach towards Efficient Deep Architectures
• https://github.com/Eric-mingjie/rethinking-network-pruning