kevinmartinjos / lottery-ticket-experiments

Experiments on the lottery ticket hypothesis for finding sparse trainable neural networks

lottery-ticket-experiments

Experiments on the lottery ticket hypothesis for finding sparse trainable neural networks

Objectives:

1. Reproduce winning lottery tickets on MNIST dataset (Lecun FCCN)
2. See if we can find winning lotter tickets for the ShuffleNet on CIFAR-10 (kinda)
3. Can we discover winning tickets faster by simply using less data? Will the tickets thus discovered continue to exhibit the lottery ticket pattern when retrained on the whole dataset? (yes!)

Please see our report for more details on our experiment and the results we obtained

How-To:

1. Install pytorch and torchvision. TODO: create a requirements.txt
2. python experiment.py --experiment=mnist to reproduce the MNIST baseline
3. python experiment.py --experiment=shufflenet to iteratively prune shufflenet using the CIFAR-10 dataset