Lightning-AI / Lightning-AI/pytorch-lightning

Allow shuffling when overfit_batches is active

Open
#9,850 19 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

feature help wanted refactor
Dominant language
Python
Stars
31.4k
Forks
3.8k
Avg merge
6d 7h
Merged PRs (30d)
6

Description

Proposed refactoring or deprecation

Instead of disabling shuffle / replacing RandomSampler with SequentialSampler in the train dataloader, replace the train dataset with a fixed subset of it using torch.utils.data.Subset (eg. first N samples of the dataset, where N is given by overfit_batches. This gives the same dataset samples as with the previous implementation.)

Motivation

This prevents training batches to be the same for every epoch

Pitch

Added on 12 Oct 2021:
The current implementation for overfit_batches disables shuffling by replacing RandomSampler with SequentialSampler in the train dataloader, in order to restrict the training / overfit to the first N samples of the train dataset for every epoch. However, this gives the same sequence of batches & non-unique batches across epochs, which is undesirable.

We should instead allow shuffling within the N samples across epochs, according to the shuffle option of the train dataloader, in order to give a different sequence of batches across epochs & mostly unique batches throughout the training process.


If you enjoy Lightning, check out our other projects! ⚡
  • Metrics: Machine learning metrics for distributed, scalable PyTorch applications.

  • Flash: The fastest way to get a Lightning baseline! A collection of tasks for fast prototyping, baselining, finetuning and solving problems with deep learning

  • Bolts: Pretrained SOTA Deep Learning models, callbacks and more for research and production with PyTorch Lightning and PyTorch

  • Lightning Transformers: Flexible interface for high performance research using SOTA Transformers leveraging Pytorch Lightning, Transformers, and Hydra.

cc @borda @justusschock @awaelchli @akihironitta @rohitgr7

Contributor guide

Open the contributing guide

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start by tracing the overfit_batches handling in the train dataloader, focusing on RandomSampler, SequentialSampler, and torch.utils.data.Subset. Compare the current fixed-sample behavior with the requested shuffling across epochs. Done means only the intended first N samples are used while dataloader shuffling remains effective between epochs, with relevant tests passing.

Written by the indexing model from the issue text.

Assessment

Tech stack
python, pytorch
Domain
machine-learning
Issue type
Refactor
Difficulty
4/5
Estimated time
3-5 days
Activity status
Stale
Clarity
Mostly clear
Newbie friendliness
35/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.