WandbCallback does not work with TfmdLists

[borisdayma]

When using TfmdLists, WandbCallback returns an error at this line.

The reason is that self.dls.valid_ds is TfmdLists so has no tls attribute (vs Datasets which is expected here).

At this step, we are trying to create a dataloader with specified idxs for future inference.

On another note, a few other issues are related to this callback:

• we sometimes run out of CUDA memory so maybe should do the prediction on CPU (since it won't happen too often).
• valid_dl may have a different batch size (maybe we should get the one from the fit loop) than the number of predictions (default to 36) so we would have to make sure we either run it a few times (if bs < n_preds) or check we have all results.

[borisdayma]

I tried a few things:

• find self.tlsor equivalent from TfmdLists but I didn't find an equivalent
• create a TfmdDL from TfmdLists using the dataloaders method but it seems a custom one would be needed so we get something like dl.new(self.subset(1)[idxs])

[borisdayma]

I looked at the predict function to process the items one at a time but I could not make it work with a learner built from TfmdLists.