Some fastai Loss Functions Cannot be Exported due to Pickling Error

[warner-benjamin]

Some fastai loss functions, such as LabelSmoothingCrossEntropyFlat are not picklable, and thus not exportable via Learner.export. But most, including LabelSmoothingCrossEntropy and other flattened losses, appear pickle without issue.

This minimal reproduction on Python 3.10 with the Learner.export default protocol=2

try:
    pickle.dumps(LabelSmoothingCrossEntropyFlat(), protocol=2)
    print("The class is picklable")
except pickle.PicklingError as e:
    print(e)

results in TypeError: cannot pickle 'code' object for LabelSmoothingCrossEntropyFlat but no errors for LabelSmoothingCrossEntropy or CrossEntropyLossFlat.

We should probably add a test for pickling LabelSmoothingCrossEntropyFlat to catch future changes to pickle.

[nglillywhite]

Hey @warner-benjamin, let me know if this still needs picking up, I'll put in a PR if so.

[warner-benjamin]

@nglillywhite I am unaware of this being worked on

[nglillywhite]

Great, I'll have a look and put in a PR, thanks

[nglillywhite]

PR 3910 adds these guards for current loss functions, I'd like to investigate why these loss functions can't be pickled with this issue but happy to open a new issue & PRs if you'd like to keep them separate.

[dharik-arsath]

Hi, I am facing a same issue on Focal Loss, unable to pickle focal loss. am I missing out something, is there any solution to this problem.

[jordiju]

Bump. Same issue here. Would like to export a Learner with FocalLossFlat