Missing training_step outputs in training_epoch_end

[mmiakashs]

Possible bug fix of #2320

[Borda]

mind add a test for this case? probably some simple example from #2320

[codecov]

Codecov Report

Merging #2327 into master will increase coverage by 0%.
The diff coverage is 100%.

@@          Coverage Diff           @@
##           master   #2327   +/-   ##
======================================
  Coverage      88%     88%           
======================================
  Files          70      70           
  Lines        5501    5503    +2     
======================================
+ Hits         4834    4836    +2     
  Misses        667     667           

[williamFalcon]

@mmiakashs mind trying master now? the solution in this PR wasn't quite 100% right and needed more testing.

This PR is likely not needed anymore but we need to add you as co-author to #2328 @Borda

[mergify]

This pull request is now in conflict... :(

[Borda]

This PR is likely not needed anymore but we need to add you as co-author to #2328 @Borda

to be done, this PR shall be merged to the other PR, but as this is closed and the other merged, there is nothing to do... pls ping me next time before close/merge 🐰

[williamFalcon]

yes... but this PR was incorrect

[mmiakashs]

@mmiakashs mind trying master now? the solution in this PR wasn't quite 100% right and needed more testing.

This PR is likely not needed anymore but we need to add you as co-author to #2328 @Borda

@williamFalcon Thanks a lot for the PR. One confusion: I just noticed that all the training_step end log metrics are combined with the dict key named 'log_metrics', however, the validation log metrics are combined with the dict key named 'log'. Is this variation intentional?

[mmiakashs]

@williamFalcon I debug again and found out that the issue #2320 still occurred only for training_step outputs. training_step outputs for the first optimizer iteration are missing, however, the second optimizer iteration outputs are merged properly.

[Borda]

@mmiakashs do you see a fix for it, mind send a PR?