[Metrics] Disable default reset after compute

[SkafteNicki]

What does this PR do?

Rework of #5193
Fixes #5166
Fixes #4641

This PR removes the automatic reset that happens after compute for metrics (solution proposed by @teddykoker here #5193 (comment)). This give the user more control over when to actually reset the state but also put a bit more responsibility on user as the they would need to reset the metric themself between epochs (except if they use metrics together with self.log, then we still auto reset between epochs for them).

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 @SkafteNicki! Thanks for updating this PR.

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

Comment last updated at 2021-01-13 08:42:24 UTC

[teddykoker]

hmm ddp tests seem to be failing, looking into it now

[teddykoker]

DDP tests broke as we were writing to self._cache multiple times before the cache is restored. Simple solution was store the cache in a local variable instead.

Now just need to fix tests/core/test_metric_result_integration.py

[teddykoker]

Second bug was a little more tricky. When the results object computes metrics, there can be duplicates of the same metric under a different name e.g. accuracy_step and accuracy_epoch. It is important not to reset() accuracy_step, if we need to compute accuracy_epoch afterwards. The simple solution here is to call compute() on epoch end for any step metrics (which we were doing before anyway to avoid any unnecessary accumulation of state.

[codecov]

Codecov Report

Merging #5409 (9508a4b) into release/1.2-dev (29bcf30) will increase coverage by 0%.
The diff coverage is 88%.

@@               Coverage Diff                @@
##           release/1.2-dev   #5409    +/-   ##
================================================
  Coverage               93%     93%            
================================================
  Files                  151     151            
  Lines                10645   10824   +179     
================================================
+ Hits                  9859   10043   +184     
+ Misses                 786     781     -5     

[tchaton]

Great job !

[tchaton]

Why reset is being added there ? self[k].compute() and self[k].reset() makes compute() useless no ?

[teddykoker]

compute() will save the computed value to self._computed, so if compute() is called again without any update() call, the cached value will be provided. We still call reset() so that we don't get too much memory usage by needlessly accumulating tensors over epochs.

[teddykoker]

This is important if we have a metric that is logged on both step and epoch. Since it will show up twice in the results meta dict, we want to reset the step metric on epoch end, but still allow the epoch metric to be computed.