mosaicml / streaming

A Data Streaming Library for Efficient Neural Network Training
https://streaming.docs.mosaicml.com
Apache License 2.0
1.09k stars 136 forks source link

Enable correct resumption from the end of an epoch #700

Closed snarayan21 closed 3 months ago

snarayan21 commented 3 months ago

Description of changes:

This PR enables correct resumption from the end of an epoch. For example, suppose the dataset has 512 samples and I want to resume from sample 512. The intended behavior should be to detect that we are at the end of the epoch and start a new one. There were two small bugs preventing this from happening correctly:

  1. The partition algorithms (orig and relaxed) both threw an error when num_samples <= drop_first instead of num_samples < drop_first. This means that if we wanted to create a sample partition of 512 samples but drop the first 512 samples, we would throw an error. This is not the right behavior -- we should return an empty sample partition instead.
  2. Writing an empty sample partition array to shared memory was not possible since we did not consider that edge case. This has been addressed.

Manual run that had earlier failed, succeeding with these changes: resumption-issue-testing-RnnDR4

Unit tests have been added for these cases.

Issue #, if available:

Merge Checklist:

Put an x without space in the boxes that apply. If you are unsure about any checklist, please don't hesitate to ask. We are here to help! This is simply a reminder of what we are going to look for before merging your pull request.

General

Tests