A Closer Look at Structured Pruning for Neural Network Compression

Code used to reproduce experiments in https://arxiv.org/abs/1810.04622.

To prune, we fill our networks with custom MaskBlocks, which are manipulated using Pruner in funcs.py. There will certainly be a better way to do this, but we leave this as an exercise to someone who can code much better than we can.

Setup

This is best done in a clean conda environment:

conda create -n prunes python=3.6
conda activate prunes
conda install pytorch torchvision -c pytorch

Repository layout

-train.py: contains all of the code for training large models from scratch and for training pruned models from scratch
-prune.py: contains the code for pruning trained models
-funcs.py: contains useful pruning functions and any functions we used commonly

CIFAR Experiments

First, you will need some initial models.

To train a WRN-40-2:

python train.py --net='res' --depth=40 --width=2.0 --data_loc=<path-to-data> --save_file='res'

The default arguments of train.py are suitable for training WRNs. The following trains a DenseNet-BC-100 (k=12) with its default hyperparameters:

python train.py --net='dense' --depth=100 --data_loc=<path-to-data> --save_file='dense' --no_epochs 300 -b 64 --epoch_step '[150,225]' --weight_decay 0.0001 --lr_decay_ratio 0.1

These will automatically save checkpoints to the checkpoints folder.

Pruning

Once training is finished, we can prune our networks using prune.py (defaults are set to WRN pruning, so extra arguments are needed for DenseNets)

python prune.py --net='res'   --data_loc=<path-to-data> --base_model='res' --save_file='res_fisher'
python prune.py --net='res'   --data_loc=<path-to-data> --l1_prune=True --base_model='res' --save_file='res_l1'

python prune.py --net='dense' --depth 100 --data_loc=<path-to-data> --base_model='dense' --save_file='dense_fisher' --learning_rate 1e-3 --weight_decay 1e-4 --batch_size 64 --no_epochs 2600
python prune.py --net='dense' --depth 100 --data_loc=<path-to-data> --l1_prune=True --base_model='dense' --save_file='dense_l1'  --learning_rate 1e-3 --weight_decay 1e-4 --batch_size 64  --no_epochs 2600

Note that the default is to perform Fisher pruning, so you don't need to pass a flag to use it.
Once finished, we can train the pruned models from scratch, e.g.:

python train.py --data_loc=<path-to-data> --net='res' --base_file='res_fisher_<N>_prunes' --deploy --mask=1 --save_file='res_fisher_<N>_prunes_scratch'

Each model can then be evaluated using:

python train.py --deploy --eval --data_loc=<path-to-data> --net='res' --mask=1 --base_file='res_fisher_<N>_prunes'

Training Reduced models

This can be done by varying the input arguments to train.py. To reduce depth or width of a WRN, change the corresponding option:

python train.py --net='res' --depth=<REDUCED DEPTH> --width=<REDUCE WIDTH> --data_loc=<path-to-data> --save_file='res_reduced'

To add bottlenecks, use the following:

python train.py --net='res' --depth=40 --width=2.0 --data_loc=<path-to-data> --save_file='res_bottle' --bottle --bottle_mult <Z>

With DenseNets you can modify the depth or growth, or use --bottle --bottle_mult <Z> as above.

Acknowledgements

Jack Turner wrote the L1 stuff, and some other stuff for that matter.

Code has been liberally borrowed from many a repo, including, but not limited to:

https://github.com/xternalz/WideResNet-pytorch
https://github.com/bamos/densenet.pytorch
https://github.com/kuangliu/pytorch-cifar
https://github.com/ShichenLiu/CondenseNet

Citing this work

If you would like to cite this work, please use the following bibtex entry:

@article{crowley2018pruning,
  title={A Closer Look at Structured Pruning for Neural Network Compression},
  author={Crowley, Elliot J and Turner, Jack and Storkey, Amos and O'Boyle, Michael},
  journal={arXiv preprint arXiv:1810.04622},
  year={2018},
  }