Reset all results on epoch end

[carmocca]

What does this PR do?

Fixes #13554

Does your PR introduce any breaking changes? If yes, please list them.

Epoch-end results will not be logged again in subsequent epochs.

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 list all the breaking changes introduced by this pull request?
• Did you update the CHANGELOG? (not for typos, docs, test updates, or minor internal changes/refactors)

PR review

• 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

cc @Borda @carmocca @edward-io @ananthsub @rohitgr7 @kamil-kaczmarek @Raalsky @Blaizzy @akihironitta

[codecov]

Codecov Report

Merging #14061 (b46bc87) into master (b25275c) will increase coverage by 15%.
The diff coverage is 100%.

@@            Coverage Diff            @@
##           master   #14061     +/-   ##
=========================================
+ Coverage      61%      76%    +15%     
=========================================
  Files         324      324             
  Lines       26384    26420     +36     
=========================================
+ Hits        16169    20147   +3978     
+ Misses      10215     6273   -3942     

[awaelchli]

@carmocca I'm quite confused as to how the on_step=False, on_epoch=True|False affects whether the special on_train_start value is being logged or not. Is this expected?

class BugModel(BoringModel):
    def on_train_start(self):
        print("ON_TRAIN_START")
        self.log("train_loss", 5000, on_step=False, on_epoch=True)

    def training_step(self, batch, batch_idx):
        loss = self(batch).sum()
        self.log("train_loss", loss, on_step=False, on_epoch=True)
        return loss

[carmocca]

@carmocca I'm quite confused as to how the on_step=False, on_epoch=True|False affects whether the special on_train_start value is being logged or not. Is this expected?

This question is totally unrelated to the PR, but to answer it. It's because we differentiate results by the {fx_name}.{logged_key}, meaning that the values self.logged in each hook don't get reduced together.

However, since they just get added to a simple dictionary (logged_metrics), and we only use logged_key as the key, the latter overwrites the former. It's a legacy design limitation

[tchaton]

LGTM !