Metric aggregation

[justusschock]

Allows aggregation on the same device as well as across multiple devices in DDP

Fixes #3245

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? Otherwise, we ask you to create a separate PR for every change.
• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?
• 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

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 🙃

[pep8speaks]

Hello @justusschock! Thanks for updating this PR.

There are currently no PEP 8 issues detected in this Pull Request. Cheers! 🍻

Comment last updated at 2020-09-13 15:02:08 UTC

[justusschock]

@SkafteNicki Do we need a default reduce op for base classes? I don't think so, since currently it's all gathered and no op is done during ddp sync

[SkafteNicki]

Really nice changes:
I completely agree that we do not need the reduce_op argument anymore for the base class.

[justusschock]

@SkafteNicki I fixed the tests to be compliant to our current behaviour. All changes to aggregation will be done in following PRs :)

[awaelchli]

a bit much with all the black formatting :)
just two feedbacks:

1. reduce op docstring is left over
2. stacking only works if all the dims are the same, is this assumption correct?

[SkafteNicki]

we have one test that are failing: test_model_pickable[DummyNumpyMetric] in file tests/metrics/test_metrics.
The test succeeds when we are using DummyTensorMetric but fails with DummyNumpyMetric. I wonder if the error is fixed, if the the special aggregation for TensorMetric also is implemented for NumpyMetric?

[codecov]

Codecov Report

Merging #3321 into master will decrease coverage by 5%.
The diff coverage is 93%.

@@           Coverage Diff            @@
##           master   #3321     +/-   ##
========================================
- Coverage      85%     80%     -5%     
========================================
  Files         104     118     +14     
  Lines        8149    9750   +1601     
========================================
+ Hits         6950    7844    +894     
- Misses       1199    1906    +707     

[Borda]

we have one test that are failing: test_model_pickable[DummyNumpyMetric] in file tests/metrics/test_metrics.
The test succeeds when we are using DummyTensorMetric but fails with DummyNumpyMetric. I wonder if the error is fixed, if the the special aggregation for TensorMetric also is implemented for NumpyMetric?

it seems to be appearing just in this PR, it is not presented in master... it shall be patched so we do not break master...

[Borda]

let's fix the failing test...

[mergify]

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

[s-rog]

Should lightning support variable sizes (in batch dimension) in gather all?

[SkafteNicki]

@Borda I fixed the failing test, can I merge this?