Description

This repository contains PyTorch code to replicate the experiments given in NeurIPS 2019 paper

"One ticket to win them all: generalizing lottery ticket initializations across datasets and optimizers"

As finding the winning lottery tickets is computationally expensive, we also open-source winning tickets (pretrained and pruned models) we generated during our experiments. Link : Winning Tickets

How to Setup

# clone project   
git clone https://github.com/varungohil/Generalizing-Lottery-Tickets.git  

# install all dependencies   
cd Generalizing-Lottery-Tickets    
pip3 install -r requirements.txt

How to Run

There are 4 files in src folder:

• train.py : Use to train the neural network and find the winning ticket
• test.py : Use to test the accuracy of the trained model
• iterative_pruning.py : Use to iteratively prune the model.
• utils.py : Contains helper functions used in scripts mentioned above

To support more datasets and architectures, we need to add necessary code to utils.py

Using train.py

Mandatory arguments:

• --architecture : To specify the neural network architecture (vgg19 and resnet50)
• --dataset : The dataset to train on (cifar10, cifar100, fashionmnist, svhn, cifar10a, cifar10b)

  Optional arguments:

• --batch-size : To set the batch size while training
• --optimizer : The optimizer to use for training (sgd and adam). sgd used by default
• --seed : To set the ranodm seed
• --model-saving-path : Path to directory where trained model is saved.

The trained model will be saved for first 5 epochs. For VGG19 it will be saved for every 16^th epoch. FOr Resnet50, the model will be saved for every 9^th epoch. For our experiments, while pruning, we reinitialize te model with weights after epoch 2 (late resetting of 1).

# source folder
cd Generalizing-Lottery-Ticket/src   

# run train.py
python3 train.py --architecture=resnet50 --dataset=cifar10    

Using iterative_pruning.py

Mandatory arguments:

• --architecture : To specify the neural network architecture (vgg19 and resnet50)
• --target-dataset : The dataset to train on (cifar10, cifar100, fashionmnist, svhn, cifar10a, cifar10b)
• --source-dataset : The dataset using which winning ticket initialization was found (cifar10, cifar100, fashionmnist, svhn, cifar10a, cifar10b)
• --init-path : Path to model with winning ticket initialization

Optional arguments:

• --batch-size : To set the batch size while training
• --optimizer : The optimizer to use for training (sgd and adam). sgd used by default
• --seed : To set the ranodm seed
• --model-saving-path : Path to directory where trained model is saved.

The script will run 30 pruning iterations which will prune away 99.9% of the weights. The trained and pruned model will be saved at end of each pruning iteration

# source folder
cd Generalizing-Lottery-Ticket/src   

# run iterative_pruning.py
python3 iterative_pruning.py --architecture=resnet50 --source-dataset=cifar10 --target-dataset=cifar100 --model-saving-path=<path-to-dir-where-models-are-to-be-stored>

Using test.py

Mandatory arguments:

• --architecture : To specify the neural network architecture (vgg19 and resnet50)
• --dataset : The dataset to train on (cifar10, cifar100, fashionmnist, svhn, cifar10a, cifar10b)
• --model-path : The path to moedl whose accuracy needs to be evaluated.

Optional arguments:

• --batch-size : To set the batch size while training

Running this script will print the Fraction of pruned weights in the model and the Test Accuracy.

# source folder
cd Generalizing-Lottery-Ticket/src   

# run train.py
python3 test.py --architecture=resnet50 --dataset=cifar10 --model-path=<path-to-model>   

Results

The results of the replicated experiments can be found in plots folder.

Citation

If you use this repository, kindly cite the reproducibility report and the original paper. The bibtex is given below.

@article{Gohil:2020,
  author = {Gohil, Varun and Narayanan, S. Deepak and Jain, Atishay},
  title = {{[Re] One ticket to win them all: generalizing lottery ticket initializations across datasets and optimizers}},
  journal = {ReScience C},
  year = {2020},
  month = may,
  volume = {6},
  number = {2},
  pages = {{#4}},
  doi = {10.5281/zenodo.3818619},
  url = {https://zenodo.org/record/3818619/files/article.pdf},
  code_url = {https://github.com/varungohil/Generalizing-Lottery-Tickets},
  code_doi = {10.5281/zenodo.3700320},
  code_swh = {swh:1:dir:8a9e53bc8a9028428bbad6a4e77ae3fedae49d30},
  data_url = {},
  data_doi = {},
  review_url = {https://openreview.net/forum?id=SklFHaqG6S},
  type = {Replication},
  language = {Python},
  domain = {NeurIPS 2019 Reproducibility Challenge},
  keywords = {lottery ticket hypothesis, pytorch}
}

@incollection{NIPS2019_8739,
title = {One ticket to win them all: generalizing lottery ticket initializations across datasets and optimizers},
author = {Morcos, Ari and Yu, Haonan and Paganini, Michela and Tian, Yuandong},
booktitle = {Advances in Neural Information Processing Systems 32},
editor = {H. Wallach and H. Larochelle and A. Beygelzimer and F. d\textquotesingle Alch\'{e}-Buc and E. Fox and R. Garnett},
pages = {4932--4942},
year = {2019},
publisher = {Curran Associates, Inc.},
url = {http://papers.nips.cc/paper/8739-one-ticket-to-win-them-all-generalizing-lottery-ticket-initializations-across-datasets-and-optimizers.pdf}
}

Core Contributors

Varun Gohil, S. Deepak Narayanan, Atishay Jain

Development

We have a new branch dev in which pull requests are welcome. We will merge them after reviewing.

Contributors

fcorencot