Pytorch-Pruning

This repository contains the implementations of neural network weight pruning methods.

Available Pruning Methods

Pruning has three stages:

1. Training
2. Pruning:
   • OneShotPruning
3. Retraining
   • No Retraining (none)
   • FineTuning (fine-tuning)
   • Learning Rate Rewinding (lr-rewinding)
   • Weight Rewinding (weight-rewinding)

Usage

Install the following python libraries: torch>=1.6.0 torchvision>=0.7.0 numpy tqdm livelossplot Run : $ mkdir saved_models

To get the pruned models you need:

1. A Model (torch.nn.module)
   • for which forward() takes input as a single variable "batch"
   • which returns loss function (by default) when the forward() is called, and returns the prediction when get_prediction argument is set to true in the forward.
2. Torch Data Loaders (train_dataloader, val_dataloader)

See the Jupyter Notebooks (.ipynb files) for a better idea about how to do the pruning. This repository contains implementations of the following Models: lenet, resnet20, resnet32, resnet44, resnet56, resnet110, resnet1202 and functions to load data loaders for MNIST, CIFAR-10

Directory Structure

.
|--- models
|    |--- __init__.py
|    |--- lenet.py
|    |--- resnet.py
|--- pruning
|    |--- unstructured
|    |    |--- __init__.py
|    |    |--- one_shot_pruning.py
|    |    |--- iterative_pruning.py
|    |--- structured
|    |    |--- __init__.py
|--- utils
|    |--- __init__.py
|    |--- data_utils.py
|    |--- train_utils.py
|--- Lisence
|--- README.md
|--- LISENCE

Todo

1. Write code for weight distribution visualizations for each layer.
2. Write code for iterative pruning.
3. Write code for ADMM sparsity regularization.