Awesome

Sanity Checks for Lottery Tickets: Does Your Winning Ticket Really Win the Jackpot?

Sample code use for NeurIPS 2021 paper: Sanity Checks for Lottery Tickets: Does Your Winning Ticket Really Win the Jackpot?

CIFAR-10 and CIFAR-100

Requirements

python >= 3.6

PyTorch >= 1.6

TorchVision >= 0.7

Other required dependency: numpy, pyyaml, matplotlib, tensorboardX, opencv-python , sklearn, scikit-image.

Main pipeline

We prune globally. The layers that are considered "global" is defined in the corresponding .yaml files. Once we prune globally, the sparsity ratio in the .yaml file will be override by the global sparsity. There are five necessary settings for the LTH experimens (using resnet-20 as example):

pretraining
- pretrain a network from scratch, resulting a dense model.
```
  cd scripts/resnet20/pretrain
  bash run.sh
```
iterative magnitude-based pruning (LT-IMP)
- Prune model iteratively. At each round the initial weights are rewind to the same initial point as pretraining. In this case, specify the same seed used in the pretraining, which will give you the same initialization. You can also --resume a pre-saved initia model as the initial point in case different servers may produce varied results using same seed.
- Each round prunes 20% of the remaining weights.
```
  cd scripts/resnet20/iterative_pruning_lth_setting
  python run_imp_lr0.1.py
```
oneshot pruning + retraining (essentially the so called lr-rewinding technique)
- Oneshot prune the pretrained model by magnitude at the beginning, followed by retraining the remaining weights.
```
  cd scripts/resnet20/oneshot_pruning_retraining
  bash run_prune_lr_0.1.sh
```
oneshot magnitude-based pruning (LT-OMP)
- Must run oneshot pruning + retraining for 1 epochs to obtain the model with OMP masks.
- This setting is different from "oneshot pruning + retraining". In "oneshot pruning + retraining", the sparse weights are continuously trained from pretrained value (lr-rewinding), while in OMP, the LTH setting is: 1) using the mask obtained from oneshot pruning on the pretrained model, and 2) using the same initial weights as pretraining for the sparse weight training.
```
  cd scripts/resnet20/oneshot_pruning_lth_setting
  bash run_omp_lr_0.1.sh
```
ramdom re-initialization (RR-IMP/OMP)
- For RR-OMP, similar to the scripts of LT-OMP, with different seed (or load different weights). For RR-IMP, change the seed and mask model path of --sp-pre-defined-mask-dir to your LT-IMP checkpoints.

ImageNet-1k

Requirements

For easy implementation, we suggest to use nvidia-docker with CUDA-11 for the training environments. We have pre-built the ready-to-run nvidia-docker image here.

Load pre-built docker images (download or build):

docker load -i nvidia_rn50.tar
Rename the docker image:

docker image tag 4c5875fdd48859f69015c7ec7183e5d2e706ffe7dabcad177e39e041673dba82 nvidia_rn50:latest
Start nvidia-docker interactive session:

nvidia-docker run --rm -it -v /path/to/your/imagenet/:/data/imagenet -v /path/to/your/project:/workspace/rn50 --ipc=host nvidia_rn50

Main pipeline

Similar with CIFAR experiments, there are five necessary settings for the LTH experimens (using resnet-50 as example):

pretraining

  cd scripts/LTH/resnet50/pretraining
  bash run.sh

iterative magnitude-based pruning (LT-IMP)

  cd scripts/LTH/resnet50/iterative_pruning_lth_setting
  python run_imp.py

oneshot pruning + retraining (essentially the so called lr-rewinding technique)

  cd scripts/LTH/resnet50/oneshot_pruning_retraining
  bash run_prune.sh

oneshot magnitude-based pruning (LT-OMP)

  cd scripts/LTH/resnet50/oneshot_pruning_lth_setting
  bash run_omp.sh

ramdom re-initialization (RR-IMP/OMP)
- For RR-OMP, similar to the scripts of LT-OMP, with different seed (or load different weights). For RR-IMP, change the seed and mask model path of --sp-pre-defined-mask-dir to your LT-IMP checkpoints.

Citation

if you find this repo is helpful, please cite

@article{ma2021sanity,
  title={Sanity Checks for Lottery Tickets: Does Your Winning Ticket Really Win the Jackpot?},
  author={Ma, Xiaolong and Yuan, Geng and Shen, Xuan and Chen, Tianlong and Chen, Xuxi and Chen, Xiaohan and Liu, Ning and Qin, Minghai and Liu, Sijia and Wang, Zhangyang and others},
  journal={Advances in Neural Information Processing Systems},
  volume={34},
  year={2021}
}