Fix unfreeze strategies with onecyclelr and reduced lr

[ethanwharris]

What does this PR do?

Fixes #1321

Note: This also changes the behaviour to no longer divide the learning rate by 10 when unfreezing. This could be seen as a breaking change, but also as a bug since this behaviour is undocumented and doesn't normally make sense (generally it's a good idea to use a smaller LR after unfreezing for the whole model but this should be handled by an LR scheduler not done silently by us).

Before submitting

• Was this discussed/approved via a Github issue? (no need for typos and docs improvements)
• 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?
• Did you write any new necessary tests? [not needed for typos/docs]
• Did you verify new and existing tests pass locally with your changes?
• If you made a notable change (that affects users), did you update the CHANGELOG?

PR review

• Is this pull request ready for review? (if not, please submit in draft mode)

Anyone in the community is free to review the PR once the tests have passed.
If we didn't discuss your PR in Github issues there's a high chance it will not be merged.

Did you have fun?

Make sure you had fun coding 🙃

[codecov]

Codecov Report

Merging #1329 (3bd7963) into master (48a2500) will increase coverage by 1.62%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master    #1329      +/-   ##
==========================================
+ Coverage   89.84%   91.47%   +1.62%     
==========================================
  Files         287      287              
  Lines       12894    12900       +6     
==========================================
+ Hits        11585    11800     +215     
+ Misses       1309     1100     -209     

Flag	Coverage Δ
unittests	91.47% <100.00%> (+1.62%)	⬆️

Flags with carried forward coverage won't be shown. Click here to find out more.

Impacted Files	Coverage Δ
flash/core/finetuning.py	90.69% <100.00%> (+0.69%)	⬆️
flash/core/serve/dag/task.py	96.82% <0.00%> (-1.06%)	⬇️
flash/text/question_answering/input.py	95.06% <0.00%> (+2.46%)	⬆️
flash/core/optimizers/optimizers.py	100.00% <0.00%> (+3.03%)	⬆️
flash/text/seq2seq/translation/model.py	63.63% <0.00%> (+4.54%)	⬆️
flash/text/question_answering/collate.py	97.64% <0.00%> (+7.05%)	⬆️
flash/text/question_answering/model.py	92.51% <0.00%> (+7.48%)	⬆️
flash/text/classification/collate.py	100.00% <0.00%> (+8.33%)	⬆️
flash/text/seq2seq/translation/cli.py	70.00% <0.00%> (+10.00%)	⬆️
flash/text/seq2seq/core/collate.py	100.00% <0.00%> (+12.50%)	⬆️
... and 13 more

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update 48a2500...3bd7963. Read the comment docs.

[krshrimali]

Nice! I agree that we should not change the LR silently. Should we also update the docs here? (in the code sample)

def finetune_function(self, pl_module, current_epoch, optimizer, opt_idx):
    # unfreeze any module you want by overriding this function

    # When ``current_epoch`` is 5, backbone will start to be trained.
    if current_epoch == self._unfreeze_epoch:
        self.unfreeze_and_add_param_group(
             pl_module.backbone,
             optimizer,
        )