Unify algorithms part 2: mixup, cutmix, label smoothing

[dblalock]

Follow-up to #524 to continue addressing #343. This is mostly cleaning up the docs to be consistent across methods + match the code's behavior, fixing the doctests, and modifying the docstrings in accordance with some slight API changes (mostly param renaming). These changes are:

Mixup:

• n_classes -> num_classes in mixup_batch, for consitency with torch + rest of composer rather than sklearn.
• interpolation_lambda -> mixing in mixup_batch
• x -> input in mixup_batch for consistency with torch
• y -> target in mixup_batch for consistency with torch
• Removed the wrapper properties for attributes in the algorithm. These weren't documented and were seemingly (?) only there so that the tests didn't access private attrs

Cutmix:

• n_classes -> num_classes again
• X -> input in cutmix_batch for consistency with torch
• y -> target in cutmix_batch for consistency with torch
• cutmix_lambda -> length in cutmix_batch, matching cutout's length argument. As part of this change, I also added changed the logic to 1) correctly compute the internal cutmix_lambda when bbox is provided, and 2) throw if both length and bbox are provided.
• cutmix_batch now returns the permutation it used for consistency with mixup_batch. @coryMosaicML , do we need to return these? We apparently didn't here, but seemingly did in mixup_batch?
• Removed the wrapper properties for attributes for same reasons as in mixup

Label smoothing:

• alpha -> interpolation in smooth_labels and algorithm
• targets -> target in smooth_labels for consistency with torch

Parts I'm not sure about (updated after slack discussion):

• Whether to return the permutation in functional forms (see above)
• How best to handle cutmix allowing 3 different ways to specify the box to use (alpha, length / cutmix_lambda, and bbox). Right now alpha is just ignored if others are provided, while the other two now raise a ValueError if both are not None. Seems like the least bad way to do this, but a little inconsistent. Maybe we could remove bbox as a parameter?
• For removed wrapper properties, we'd ideally just not have the tests depend on class internals. I went with having tests access the private attrs rather than cluttering the code + public API, but we could also just make these attrs public and document them. Probably should have put this change in a separate PR, but I was already renaming the attrs and I was on a roll / changing a lot anyway, so I just went with it.

NOTE: need to fix jenkins error + double-check docstring output in the docs, but this is ready for feedback, esp regarding the API questions above.

[hanlint]

I like where this is going! Agreed on interpolation... how strictly do we want to maintain cross-functional consistency versus what the original paper/authors terminology. e.g. alpha for label smoothing?

[coryMosaicML]

cutmix_batch now returns the permutation it used for consistency with mixup_batch. @coryMosaicML , do we need to return these? We apparently didn't here, but seemingly did in mixup_batch?

I think this behavior in mixup_batch is a holdover from the old repo that got ported over. I don't think the original reason for doing that exists in composer, and not returning the permutation should be fine. It makes the functional interface feel a lot cleaner to me. However, two possible reasons to return permutations: 1) If someone wants to do further analysis/debugging of what mixup_batch and cutmix_batch are doing to the inputs, this could be handy, and 2) returning the permutation makes it possible to use the loss interpolation trick people often do to use mixup and cutmix with index labels after using mixup_batch and cutmix_batch if they so choose.

[hanlint]

I noticed two instances of significant test setup code to create the model and dataloaders. Are those needed, given that some of these are auto-loaded in https://github.com/mosaicml/composer/blob/dev/docs/source/doctest_fixtures.py ?

[hanlint]

Also -- update yamls?

[dblalock]

Also -- update yamls?

The algorithm yamls? The label_smoothing one is updated, but cutmix and mixup fortunately don't have to change anything since they only use alpha and num_classes, and these are unchanged.

[hanlint]

LGTM