rahulsmehta/sparsity-experiments

Code/experiments for Sparse Transfer Learning via Winning Lottery Tickets

Sparse Transfer Learning via Winning Lottery Tickets

Paper: https://arxiv.org/abs/1905.07785

This repository contains code for reproducing experiments in Sparse Transfer Learning via Winning Lottery Tickets.

Connecting to AWS

Ensure that the following environment variables are set (easiest way to do this is consolidate these in an ~/.aws_config file that is read in your bash_profile). For instance:

export AWS_KEY_PATH=/path/to/my_key.pem
export AWS_HOST=<my_host>.compute-<n>.amazonaws.com
export AWS_USER=ec2-user

Once this is complete run ./connect-ssh.sh [-o]. The optional flag -o forwards AWS_HOST:8888 to localhost:8888 to use Jupyter Lab.

sparsity-experiments's Issues

Fix early stopping

Actually return out and stop the entire training loop by returning a flag that the early stopping condition was met (rather than continuing for all epochs/pruning iterations)

Refactor lt-cifar-resnet18.py script to train a single initial network and checkpoint for each pruning

Path-dependent on #1 being fixed

Finish NORB dataloader/train basic ResNet on it