MNT Fixes failing test_keras

[thomasjpfan]

Fixes failing test on the main branch

[stsievert]

Thanks for the PR @thomasjpfan.

Looks like the tests are CI tests are failing because of an unrelated linting issue w/ mypy.

linux38 CI job: linting output

Checking flake8...
Checking flake8... DONE
Checking black...
black, version 19.10b0
All done! ✨ 🍰 ✨
100 files would be left unchanged.
Checking black... DONE
Checking isort...
4.3.21
Skipped 2 files
Checking isort... DONE
Checking mypy...
dask_ml/metrics/regression.py:102: error: "ndarray" has no attribute "chunks"
dask_ml/metrics/pairwise.py:30: error: "ndarray" has no attribute "to_delayed"
dask_ml/metrics/pairwise.py:37: error: "ndarray" has no attribute "chunks"
dask_ml/metrics/pairwise.py:42: error: "ndarray" has no attribute "chunks"
dask_ml/metrics/pairwise.py:58: error: "ndarray" has no attribute "chunks"
dask_ml/metrics/pairwise.py:59: error: "ndarray" has no attribute "map_blocks"
dask_ml/metrics/pairwise.py:136: error: Argument 2 to "check_pairwise_arrays" has incompatible type "Optional[ndarray]"; expected "ndarray"
dask_ml/metrics/pairwise.py:144: error: Argument 2 to "check_pairwise_arrays" has incompatible type "Optional[ndarray]"; expected "ndarray"
dask_ml/metrics/pairwise.py:161: error: Argument 2 to "check_pairwise_arrays" has incompatible type "Optional[ndarray]"; expected "ndarray"
dask_ml/metrics/pairwise.py:176: error: Argument 2 to "check_pairwise_arrays" has incompatible type "Optional[ndarray]"; expected "ndarray"
dask_ml/metrics/pairwise.py:211: error: Argument 2 to "check_pairwise_arrays" has incompatible type "Optional[ndarray]"; expected "ndarray"
Found 11 errors in 2 files (checked 5 source files)
dask_ml/preprocessing/label.py:200: error: Incompatible default for argument "uniques" (default has type "None", argument has type "ndarray")
dask_ml/preprocessing/label.py:202: error: "ndarray" has no attribute "cat"
dask_ml/preprocessing/label.py:212: error: "ndarray" has no attribute "cat"
dask_ml/preprocessing/data.py:39: error: Incompatible types in assignment (expression has type "float", variable has type "ndarray")
dask_ml/preprocessing/data.py:78: error: Need type annotation for 'n_features_in_'
dask_ml/preprocessing/data.py:145: error: Need type annotation for 'n_features_in_'
dask_ml/preprocessing/data.py:229: error: Argument 1 to "_handle_zeros_in_scale" has incompatible type "List[float]"; expected "ndarray"
dask_ml/preprocessing/data.py:230: error: Need type annotation for 'n_features_in_'
dask_ml/preprocessing/_encoders.py:119: error: Incompatible default for argument "dtype" (default has type "Type[floating[Any]]", argument has type "dtype[Any]")
dask_ml/preprocessing/_block_transformer.py:74: error: Need type annotation for 'func'
Found 19 errors in 4 files (checked 5 source files)
Checking mypy... DONE
##[error]Bash exited with code '1'.

    Finishing: Lint

[jsignell]

It seems like this one is good to merge then?

[stsievert]

This PR LGTM.

[jsignell]

Do you have merge access @stsievert? I don't :)

[stsievert]

Nope, no merge access. I know @TomAugspurger does though.

[jsignell]

And probably @jrbourbeau as well :) Seems like this one is good to go

[jrbourbeau]

Thanks @thomasjpfan! Merging as this is a definite improvement over our current CI failure situation. However I did want to point out that it appears this test isn't run on CI and is instead skipped here

dask-ml/tests/model_selection/test_keras.py

Line 30 in 8c46053

pytestmark = pytest.mark.skip(reason="Missing tensorflow or scikeras")

In the future it'd be good to make sure we have a CI build with tensorflow and scikeras installed

[jsignell]

Good point @jrbourbeau. It looks like keras and tensorflow are only installed when the branch is main. So the test ran once it was merged. I am not sure if that is intentional or not.

@thomasjpfan when it ran there it passed on all builds except earliest: https://dev.azure.com/dask-dev/6d1a2a76-7181-452f-943c-c034bbab2bfa/_apis/build/builds/1961/logs/39 this might just be an instance of we should bump the minimum requirements.

[stsievert]

Hm... SciKeras raises an error if loss is None, but KerasClassifier defaults to loss=None. More discussion will happen in adriangb/scikeras#206.