feat(wandb): improve callback

[borisdayma]

Following changes are implemented:

• ensure callback is executed at the end -> we would get errors when the callback was executed before smooth_loss was defined
• time is not logged anymore as it is formatted as hh:mm which cannot be plotted (and create an error on associated graph). Wall time and relative time are already automatically logged by wandb
• we keep track of cumulative epoch -> this allows to handle the cases where the validation set changes between training stages (as n_iter could vary)

The changes allow to call learner(…, cbs=WandbCallback()) and multiple times .fit methods (even changing validation set) and add data to the same graphs as in this run.

[review-notebook-app[bot]]

Check out this pull request on 

You'll be able to see Jupyter notebook diff and discuss changes. Powered by ReviewNB.

[borisdayma]

Remaining TODO:

• prevent logging during lr_find such as this run -> could we update the no_logging function by adding a no_log attribute so it could be used in the callback?
• add tests -> I can look into this when we finalize base functionality

Possible improvements:

• log data such as confusion matrix and top losses samples (depending on data class type) after each epoch -> dataframes are nice to visualize like in this example

[borisdayma]

Slightly confused about the build error, when I run diff_nb_script I see no output.

[borisdayma]

I took advantage to add alternative use in the description: learn = learner(..., cbs=WandbCallback())

Feel free to remove entity from the docs after testing as it should not be needed anymore.

[sgugger]

Thanks!