Support gamma objective in LGBMRegressor

[janjagusch]

With this PR, I'm adding support for the gamma objective for the LGBMRegressor.

I don't think the failing Windows tests are related to my changes.

[xadupre]

I'll look into the failure. Could you add a unit test to check the conversion works with this objective?

[janjagusch]

I'll look into the failure. Could you add a unit test to check the conversion works with this objective?

@xadupre, I have added test cases for all objectives that we're currently supporting. It's also easy to add more test cases for more objectives (e.g. tweedie, which I might add in a future PR). Hope the test suite is okay for you, usually I use pytest so I'm not very used to unittest.

[xadupre]

This test is failing for some of the configuration. It should pass with lower decimal. Otherwise, it looks good.

[janjagusch]

Changing the number of decimals won't make this test less flaky unfortunately (because of the precision difference in the split rules, the prediction error could be arbitrarily large). I added a new test function that allows 1 in 10,000 rows to have a difference larger than the specified decimals.

[janjagusch]

@xadupre i think all tests that i can influence are succeeding now.

Two tests are failing that I don't think are related to my changes:

• This is a different test case:

tests/coreml/test_cml_GLMClassifierConverter.py::TestCoreMLGLMClassifierConverter::test_glm_classifier
  /home/vsts/work/1/s/onnxmltools/utils/tests_helper.py:181: UserWarning: Issue with 'CmlBinLinearSVC-NoProb' due to Model 'tests/temp/CmlBinLinearSVC-NoProb.model.onnx' has discrepencies.
  <class 'onnxmltools.utils.utils_backend.OnnxRuntimeAssertionError'>: Unexpected output in model 'tests/temp/CmlBinLinearSVC-NoProb.model.onnx'
  max diff(expected, output)=3.691214992322873
  
  Arrays are not almost equal to 5 decimals
  
  Mismatched elements: 150 / 150 (100%)
  Max absolute difference: 3.69121499
  Max relative difference: 34.80315894
   x: array([-1.31647, -0.49216, -1.24896, -1.17561, -1.69487, -1.76238,
         -1.85322, -1.24312, -1.02892, -0.71803, -1.31064, -1.54817,
         -0.64469, -1.40731, -1.37815, -2.43416, -1.76238, -1.31647,...
   y: array([0.21141, 0.37938, 0.22288, 0.23584, 0.15514, 0.14649, 0.13549,
         0.22389, 0.26329, 0.32783, 0.21238, 0.17535, 0.34419, 0.19666,
         0.20131, 0.0806 , 0.14649, 0.21141, 0.25371, 0.11983, 0.34683,...
    warnings.warn("Issue with '{0}' due to {1}".format(basename, e))

• And then one job is failing because the LgbmRegressor:TreeEnsembleRegressor is not implemented in the runtime.

onnxruntime.capi.onnxruntime_pybind11_state.NotImplemented: [ONNXRuntimeError] : 9 : NOT_IMPLEMENTED : Could not find an implementation for the node LgbmRegressor:TreeEnsembleRegressor(1)

Feel free to merge or let me know if you need anything else. 🙂

[xadupre]

Thanks, I'll have a look. The broken tests are not related. I noticed some weird failure too (coremltools 3.1 fails with scikit-learn 0.24.2 on Windows and python 3.9).

[janjagusch]

Thanks, I'll have a look. The broken tests are not related. I noticed some weird failure too (coremltools 3.1 fails with scikit-learn 0.24.2 on Windows and python 3.9).

@xadupre, considering that the failures are not related to the changes, do you think we could merge this? 🙂

[janjagusch]

Thanks, I'll have a look. The broken tests are not related. I noticed some weird failure too (coremltools 3.1 fails with scikit-learn 0.24.2 on Windows and python 3.9).

@xadupre, considering that the failures are not related to the changes, do you think we could merge this? 🙂

ping, @xadupre. 🙂

[xadupre]

It is possible to fix CI? The failing test fails because it is run with an old version of onnxruntime (1.0). Can you disable the test when onnxruntime is older than 1.3?

[janjagusch]

It is possible to fix CI? The failing test fails because it is run with an old version of onnxruntime (1.0). Can you disable the test when onnxruntime is older than 1.3?

Fixed the CI, ready to merge. 🙂