Deprecate terminate_on_nan Trainer argument in favor of detect_anomaly

[yopknopixx]

What does this PR do?

Follow up to #8313
Fixes #9849
Closes #9849

Does your PR introduce any breaking changes? If yes, please list 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 list all the breaking changes introduced by this pull request?
• Did you update the CHANGELOG? (not for typos, docs, test updates, or internal minor changes/refactorings)

PR review

Anyone in the community is welcome to review the PR.
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

Did you have fun?

Make sure you had fun coding 🙃

[carmocca]

2 questions regarding the difference between terminate_on_nan and detect_anomaly:

• Does anomaly detection raise an error on +- inf? terminate_on_nan does:
  https://github.com/PyTorchLightning/pytorch-lightning/blob/46b00a7de44d226a3da22b258c74cc2b3cd92bec/pytorch_lightning/loops/utilities.py#L35
• Will we keep the loss value check when terminate_on_nan is removed?
  https://github.com/PyTorchLightning/pytorch-lightning/blob/46b00a7de44d226a3da22b258c74cc2b3cd92bec/pytorch_lightning/loops/batch/training_batch_loop.py#L279-L281

[yopknopixx]

2 questions regarding the difference between terminate_on_nan and detect_anomaly:

• Does anomaly detection raise an error on +- inf? terminate_on_nan does:
  https://github.com/PyTorchLightning/pytorch-lightning/blob/46b00a7de44d226a3da22b258c74cc2b3cd92bec/pytorch_lightning/loops/utilities.py#L35
• Will we keep the loss value check when terminate_on_nan is removed?
  https://github.com/PyTorchLightning/pytorch-lightning/blob/46b00a7de44d226a3da22b258c74cc2b3cd92bec/pytorch_lightning/loops/batch/training_batch_loop.py#L279-L281

Anomaly detection doesn't raise an error for infs. Also, since detect_anomaly doesn't check the loss I think the loss check should be kept after terminate_on_nan is removed.

[carmocca]

Anomaly detection doesn't raise an error for infs. Also, since detect_anomaly doesn't check the loss I think the loss check should be kept after terminate_on_nan is removed.

In that case, I don't think we should deprecate terminate_on_nan but instead add the anomaly detection functionality to the existing flag.

[cowwoc]

Per #9799 we only want to turn on detect_anomaly after the scaler value converges to a stable value:

At the beginning of training, in the first few iterations it can happen that the grads become inf or nan. AMP will skip the optimizer.step() in that case and adjust the scale factor iteratively until it converges to a point where that doesn't happen.

[awaelchli]

@yopknopixx After #9848 got merged, I am repurposing your PR here for the deprecation of the old terminate_on_nan argument so we can include your contribution. Thanks for your patience!