Implement finalize for WandbLogger

[awaelchli]

What does this PR do?

Fixes #3002

The checkpoint artifacts will be uploaded at the end of the run:

[pep8speaks]

Hello @awaelchli! Thanks for updating this PR.

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2020-10-26 10:54:55 UTC

[codecov]

Codecov Report

Merging #4341 into master will decrease coverage by 0%.
The diff coverage is 67%.

@@          Coverage Diff           @@
##           master   #4341   +/-   ##
======================================
- Coverage      93%     93%   -0%     
======================================
  Files         111     111           
  Lines        8046    8049    +3     
======================================
+ Hits         7467    7469    +2     
- Misses        579     580    +1     

[borisdayma]

Thanks @awaelchli

I see from the commit history you probably intended to call self.logger.experiment.finish().

The reason I didn't add it in the logger is I was scared it would affect users who manually upload other charts (interpretation, etc) or files to their W&B run once training is complete.

I think we will be able to make the model upload cleaner through W&B artifacts.
I plan to look at it and propose something in the coming weeks.

[awaelchli]

@borisdayma Great to have you look into it more!
I tried self.experiment.finish() but this run object didn't have the finish method. The reason I ended up with wandb.save is because I saw it in the docs and it seemed to be a good way. For now, it is guarded by self._log_model so people can keep it off if there is an issue with it and we can look into your proposal :)

[borisdayma]

@awaelchli Thanks for the feedback. It would be easy add finish to self.experiment but I don't think it's a good idea.
For now let's keep your implementation and we'll try to find a better way with artifacts.
Thanks a lot for the help!