QIU023 / Lottery-Ticket-Hypothesis-in-Pytorch

This repository contains a Pytorch implementation of the paper "The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks" by Jonathan Frankle and Michael Carbin that can be easily adapted to any model/dataset.

Lottery Ticket Hypothesis in Pytorch

This repository contains a Pytorch implementation of the paper The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks by Jonathan Frankle and Michael Carbin that can be easily adapted to any model/dataset.

Requirements

pip3 install -r requirements.txt

How to run the code ?

Using datasets/architectures included with this repository :

python3 main.py --prune_type=lt --arch_type=fc1 --dataset=mnist --prune_percent=10 --prune_iterations=35

• --prune_type : Type of pruning
  • Options : lt - Lottery Ticket Hypothesis, reinit - Random reinitialization
  • Default : lt
• --arch_type : Type of architecture
  • Options : fc1 - Simple fully connected network, lenet5 - LeNet5, AlexNet - AlexNet, resnet18 - Resnet18, vgg16 - VGG16
  • Default : fc1
• --dataset : Choice of dataset
  • Options : mnist, fashionmnist, cifar10, cifar100
  • Default : mnist
• --prune_percent : Percentage of weight to be pruned after each cycle.
  • Default : 10
• --prune_iterations : Number of cycle of pruning that should be done.
  • Default : 35
• --lr : Learning rate
  • Default : 1.2e-3
• --batch_size : Batch size
  • Default : 60
• --end_iter : Number of Epochs
  • Default : 100
• --print_freq : Frequency for printing accuracy and loss
  • Default : 1
• --valid_freq : Frequency for Validation
  • Default : 1
• --gpu : Decide Which GPU the program should use
  • Default : 0

Using datasets/architectures that are not included with this repository :

• Adding a new architecture :
  • For example, if you want to add an architecture named new_model with mnist dataset compatibility.
    • Go to /archs/mnist/ directory and create a file new_model.py.
    • Now paste your Pytorch compatible model inside new_model.py.
    • IMPORTANT : Make sure the input size, number of classes, number of channels, batch size in your new_model.py matches with the corresponding dataset that you are adding (in this case, it is mnist).
    • Now open main.py and go to line 36 and look for the comment # Data Loader. Now find your corresponding dataset (in this case, mnist) and add new_model at the end of the line from archs.mnist import AlexNet, LeNet5, fc1, vgg, resnet.
    • Now go to line 82 and add the following to it :

       elif args.arch_type == "new_model":
       	model = new_model.new_model_name().to(device)
      

      Here, new_model_name() is the name of the model that you have given inside new_model.py.
• Adding a new dataset :
  • For example, if you want to add a dataset named new_dataset with fc1 architecture compatibility.
    • Go to /archs and create a directory named new_dataset.
    • Now go to /archs/new_dataset/and add a file namedfc1.py` or copy paste it from existing dataset folder.
    • IMPORTANT : Make sure the input size, number of classes, number of channels, batch size in your new_model.py matches with the corresponding dataset that you are adding (in this case, it is new_dataset).
    • Now open main.py and goto line 58 and add the following to it :

       elif args.dataset == "cifar100":
       	traindataset = datasets.new_dataset('../data', train=True, download=True, transform=transform)
       	testdataset = datasets.new_dataset('../data', train=False, transform=transform)from archs.new_dataset import fc1
      

      Note that as of now, you can only add dataset that are natively available in Pytorch.

How to combine the plots of various prune_type ?

• Go to combine_plots.py and add/remove the datasets/archs who's combined plot you want to generate (Assuming that you have already executed the main.py code for those dataset/archs and produced the weights).
• Run python3 combine_plots.py.
• Go to /plots/lt/combined_plots/ to see the graphs.

Kindly raise an issue if you have any problem with the instructions.

Datasets and Architectures that were already tested

	fc1	LeNet5	AlexNet	VGG16	Resnet18
MNIST	✔️	✔️	✔️	✔️	✔️
CIFAR10	✔️	✔️	✔️	✔️	✔️
FashionMNIST	✔️	✔️	✔️	✔️	✔️
CIFAR100	✔️	✔️	✔️	✔️	✔️

Repository Structure

Lottery-Ticket-Hypothesis-in-Pytorch
├── archs
│   ├── cifar10
│   │   ├── AlexNet.py
│   │   ├── densenet.py
│   │   ├── fc1.py
│   │   ├── LeNet5.py
│   │   ├── resnet.py
│   │   └── vgg.py
│   ├── cifar100
│   │   ├── AlexNet.py
│   │   ├── fc1.py
│   │   ├── LeNet5.py
│   │   ├── resnet.py
│   │   └── vgg.py
│   └── mnist
│       ├── AlexNet.py
│       ├── fc1.py
│       ├── LeNet5.py
│       ├── resnet.py
│       └── vgg.py
├── combine_plots.py
├── dumps
├── main.py
├── plots
├── README.md
├── requirements.txt
├── saves
└── utils.py

Interesting papers that are related to Lottery Ticket Hypothesis which I enjoyed

• Deconstructing Lottery Tickets: Zeros, Signs, and the Supermask

Acknowledgement

Parts of code were borrowed from ktkth5.

Issue / Want to Contribute ? :

Open a new issue or do a pull request incase you are facing any difficulty with the code base or if you want to contribute to it.