search

pytorch / opacus

Training PyTorch models with differential privacy
https://opacus.ai
Apache License 2.0
1.67k stars 332 forks source link

Fixed issue with missing argument in MNIST example #520

Closed gouthamuoc closed 1 year ago

gouthamuoc commented 1 year ago

Summary: sr stands for sampling rate, which is now legacy code. Now, it's just sample_rate = 1 / len(data_loader). This has been fixed in the example and values have been updated.

When running with DP, the following error is thrown

AttributeError: Can't pickle local object 'wrap_collate_with_empty.<locals>.collate'

For now, a temporary fix (based on https://github.com/IBM/Project_CodeNet/issues/21#issuecomment-864619383) is to make num_workers=0 in the dataset loaders. This commit does that.

Differential Revision: D40253037

facebook-github-bot commented 1 year ago

This pull request was exported from Phabricator. Differential Revision: D40253037

ffuuugor commented 1 year ago

@gouthamuoc Hi and thanks for your contribution! We've indeed missed on updating the documentation.

However, I would say that having a doc with MNIST classifier training to just 20% accuracy is the the point we want to make. We want to show that it is possible to train a decent classifier (90%+ accuracy) with a given privacy parameters.

--sr 0.004 would be equivalent to --batch-size 240 on MNIST (60000 * 0.004 = 240), so I would suggest updating the run commands, but keeping the loss/accuracy data. If you prefer round numbers, we can do batch size 256 (the loss/accuracy would probably still be within expected noise and won't need to be updated still)

gouthamuoc commented 1 year ago

Thanks, I updated accordingly and the test runs give similar accuracy.

facebook-github-bot commented 1 year ago

This pull request was exported from Phabricator. Differential Revision: D40253037