Example of K-fold Cross Validation with Fabric

[shenoynikhil]

What does this PR do?

Resolves #16902

Before submitting

• Was this discussed/agreed via a GitHub issue? (not for typos and docs)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests? (not for typos and docs)
• Did you verify new and existing tests pass locally with your changes?
• Did you list all the breaking changes introduced by this pull request?
• Did you update the CHANGELOG? (not for typos, docs, test updates, or minor internal changes/refactors)

PR review

Anyone in the community is welcome to review the PR.
Before you start reviewing, make sure you have read the review guidelines. In short, see the following bullet-list:

Reviewer checklist

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified

[peterchristofferholm]

Instead of having the k-folds in outer loop, it would be better to loop over the folds in each epoch. Otherwise you will not be able to things like early stopping or pruning doing hyperparameter sweeps.

for epoch in range(max_epochs):
    losses = []
    for k in range(k_folds):
        yhat = model[k](data[k])
        losses[k] = ...
    epoch_loss = mean(losses)

[shenoynikhil]

Instead of having the k-folds in outer loop, it would be better to loop over the folds in each epoch. Otherwise you will not be able to things like early stopping or pruning doing hyperparameter sweeps.

That makes sense. I'm looking into it.

[shenoynikhil]

@peterchristofferholm I made the changes accordingly. I believe I saw K-Fold implemented in both ways, one where the k-fold loop is the outermost one and one where the epoch loop is. Not sure if there is one correct one, because eventually, you would want to aggregate validation score across the models.

@awaelchli Can you maybe also review how Fabric has been used?

[justusschock]

Instead of having the k-folds in outer loop, it would be better to loop over the folds in each epoch. Otherwise you will not be able to things like early stopping or pruning doing hyperparameter sweeps.

The downside of this is, that with the current implementation you have K-times your model in GPU-memory (which often fits only a single model), or you have to do manual device management

[shenoynikhil]

@awaelchli @justusschock @peterchristofferholm I guess I have both the implementations in the git history. Let me know which one is preferred, and I'll keep that.

[awaelchli]

LGTM, minor comments.
Thanks for jumping on this so quickly. Impressive!