[BugFix] Resolve DDP hanging bug within ModelCheckpoint and val_loss

[tchaton]

What does this PR do?

Currently, the training can hang under the following conditions:

• No monitor is being provided to ModelCheckPoint
• val_loss is being logged

Internally in ModelCheckpoint, self.monitor will become val_loss.

    def _add_backward_monitor_support(self, trainer):
        metrics = trainer.logger_connector.callback_metrics
        # backward compatibility... need to deprecate
        if self.monitor is None and 'val_loss' in metrics:
            self.monitor = 'val_loss'
        if self.monitor is None and 'checkpoint_on' in metrics:
            self.monitor = 'checkpoint_on'
        if self.save_top_k is None and self.monitor is not None:
            self.save_top_k = 1

And the code will diverge slightly later if check_monitor_top_k returns different value for both processes.

self.check_monitor_top_k(metrics.get(self.monitor))

Solution:

• val_loss and checkpoint_on should be reduced.

Fixes #5865

Before submitting

• Was this discussed/approved via a GitHub issue? (not for typos and docs)
• Did you read the contributor guideline, Pull Request section?
• Did you make sure your PR does only one thing, instead of bundling different changes together?
• Did you make sure to update the documentation with your changes? (if necessary)
• Did you write any new necessary tests? (not for typos and docs)
• Did you verify new and existing tests pass locally with your changes?
• Did you update the CHANGELOG? (not for typos, docs, test updates, or internal minor changes/refactorings)

PR review

Anyone in the community is free to review the PR once the tests have passed.
Before you start reviewing make sure you have read Review guidelines. In short, see the following bullet-list:

• Is this pull request ready for review? (if not, please submit in draft mode)
• Check that all items from Before submitting are resolved
• Make sure the title is self-explanatory and the description concisely explains the PR
• Add labels and milestones (and optionally projects) to the PR so it can be classified
• Check that target branch and milestone match!

Did you have fun?

Make sure you had fun coding 🙃

[pep8speaks]

Hello @tchaton! Thanks for updating this PR.

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

Comment last updated at 2021-02-17 13:12:13 UTC

[codecov]

Codecov Report

Merging #6004 (374466d) into master (5157ba5) will decrease coverage by 3%.
The diff coverage is 100%.

@@           Coverage Diff           @@
##           master   #6004    +/-   ##
=======================================
- Coverage      93%     90%    -3%     
=======================================
  Files         160     170    +10     
  Lines       11340   11786   +446     
=======================================
+ Hits        10550   10637    +87     
- Misses        790    1149   +359     

[justusschock]

Shouldn't we reduce it always?

[tchaton]

Metrics already perform reduction during compute or self.log.
I think the issue only happens with legacy metrics.
@awaelchli Should it be always be done ?

[awaelchli]

I'm not sure about this. I thought the metrics that the checkpoint gets from the trainer/logger connector are already reduced? We shouldn't let the checkpoint have the responsibility to reduce metrics or to assume how. the mean is not always correct.

[SeanNaren]

can we get a test case of this legacy metric? It will help debug and figure out why this reduction is necessary
loss should always be reduced I think, so I'm a bit puzzled how this would fix an issue

[carmocca]

Note that we should deprecate and remove _add_backward_monitor_support (we never got around to do it). I can do it in another PR

[tchaton]

Note that we should deprecate and remove _add_backward_monitor_support (we never got around to do it). I can do it in another PR

Sounds good !

[SeanNaren]

Seems like this didn't fix the issue in #5865. We're really missing a test case here for us to debug, so might be a good idea to get the user to help out here.