An implementation of <Group Fisher Pruning for Practical Network Compression> based on pytorch and mmcv

• Pruning for fully-convolutional structures, such as one-stage detectors; (copied from the official code)

• Pruning for networks combining convolutional layers and fully-connected layers, such as faster-RCNN and ResNet;

• Pruning for networks which involve group convolutions, such as ResNeXt and RegNet.

torch
torchvision
mmcv / mmcv-full
mmcls 
mmdet 

This code is tested with

pytorch=1.3
torchvision=0.4
cudatoolkit=10.0
mmcv-full==1.3.14
mmcls=0.16 
mmdet=2.17

and

pytorch=1.8
torchvision=0.9
cudatoolkit=11.1
mmcv==1.3.16
mmcls=0.16 
mmdet=2.17

Download ImageNet and COCO, then extract them and organize the folders as

- detection
  |- tools
  |- configs
  |- data
  |   |- coco
  |   |   |- train2017
  |   |   |- val2017
  |   |   |- test2017
  |   |   |- annotations
  |
- classification
  |- tools
  |- configs
  |- data
  |   |- imagenet
  |   |   |- train
  |   |   |- val
  |   |   |- test 
  |   |   |- meta
  |
- ...

e.g. Classification

cd classification

1. Pruning

   # single GPU
   python tools/train.py configs/xxx_pruning.py --gpus=1
   # multi GPUs (e.g. 4 GPUs)
   python -m torch.distributed.launch --nproc_per_node=4 tools/train.py configs/xxx_pruning.py --launch pytorch

2. Fine-tune

   In the config file, modify the deploy_from to the pruned model, and modify the samples_per_gpu to 256/#GPUs. Then

   # single GPU
   python tools/train.py configs/xxx_finetune.py --gpus=1
   # multi GPUs (e.g. 4 GPUs)
   python -m torch.distributed.launch --nproc_per_node=4 tools/train.py configs/xxx_finetune.py --launch pytorch

3. Test

   In the config file, add the attribute load_from to the finetuned model. Then

   python tools/test.py configs/xxx_finetune.py --metrics=accuracy

The commands for pruning and finetuning of detection models are similar to that of classification models. Instructions will be added soon.

My project acknowledges the official code FisherPruning.