Fix ModelCheckpoint(monitor=None, save_last=True) not saving checkpoints

[carmocca]

What does this PR do?

• If monitor=None and save_last=True, only the last checkpoints were getting saved.
• If save_top_k=0 and save_last=True, the last checkpoint wasn't getting saved

Fixes #6096
Closes #6344

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?
• [n/a] 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

• 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

[pep8speaks]

Hello @carmocca! Thanks for updating this PR.

• In the file pytorch_lightning/callbacks/model_checkpoint.py:

Line 558:13: W503 line break before binary operator
Line 559:13: W503 line break before binary operator
Line 560:13: W503 line break before binary operator

Comment last updated at 2021-03-07 21:50:57 UTC

[codecov]

Codecov Report

Merging #6136 (6182e18) into master (c7f30a2) will decrease coverage by 47%.
The diff coverage is 22%.

@@           Coverage Diff            @@
##           master   #6136     +/-   ##
========================================
- Coverage      94%     46%    -47%     
========================================
  Files         161     161             
  Lines       11476   11376    -100     
========================================
- Hits        10735    5288   -5447     
- Misses        741    6088   +5347     

[ananthsub]

if more options are added, the number of combinations will grow, and resolving the options inside the constructor will be error prone.

would it be safer to set a default value for save_top_k=-1 ? the last time this was raised, storage was a concern, but I see this as the safest option given that there's no default value available for monitor

[carmocca]

if more options are added, the number of combinations will grow, and resolving the options inside the constructor will be error prone.

I agree, ModelCheckpoint should be split up as discussed in #4335

would it be safer to set a default value for save_top_k=-1 ? the last time this was raised, storage was a concern, but I see this as the safest option given that there's no default value available for monitor

The current default is always save (monitor=None) but only keep one (save_top_k=None). Which is the same as setting (save_last=True, save_top_k=0). It doesn't seem like a bad default to me.

[tchaton]

LGTM !

[awaelchli]

isn't it safer to have separate test, so that we know they independently work?

[awaelchli]

in this test here we also have save_top_k=-1, but according to the PR description, I think we need a test for
monitor=None, save_last=True and the test should show that two files are saved:
epoch=X.ckpt and last.ckpt, where X is the last epoch and they are two identical files.
is this correct?

[carmocca]

so that we know they independently work?

Which things should be independent?

in this test here we also have save_top_k=-1, but according to the PR description, I think we need a test for
monitor=None, save_last=True

Top k -1 or >0 does not matter.

I think we need a test for
monitor=None, save_last=True and the test should show that two files are saved:
epoch=X.ckpt and last.ckpt, where X is the last epoch and they are two identical files.

That is the purpose of this (adapted) test, although it's not testing that the contents are exactly the same