Home

Awesome

Neuron-level Structured Pruning using Polarization Regularizer

NeurIPS 2020 [Paper]

Introduction

Pipeline:

  1. Sparsity Training
  2. Pruning
  3. Fine-tuning

Running

We test our code on Python 3.6. Our code is incompatible with Python 2.x.

Install packages:

pip install -r requirements.txt

We recommend to run the code on PyTorch 1.2 and CUDA 10.0. The project is incompatible with PyTorch <= 1.0.

See README in ./imagenet or ./cifar for guidelines on running experiments on ImageNet (ILSVRC-12) or CIFAR10/100 datasets.

We upload the the pruned checkpoints on OneDrive.

Note

Pruning strategy

We introduce a novel pruning method in our paper (Fig. 2). We have implemented multiple pruning methods in our code (option --pruning-strategy).

Loss Type

Acknowledgement

We build our code based on rethinking-network-pruning. We'd like to thank their contribution to the research on structured pruning.