Support drop option of OneHotEncoder

Hi,

I tried to use a OneHotEncoder with drop option set to first but when I do so, running the generated model with onnxruntime gives me the following error during prediction:

onnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL : Non-zero status code returned while running OneHotEncoder node. Name:'OneHotEncoder7' Status Message: Unknown Category and zeros = 0.

See https://github.com/onnx/sklearn-onnx/issues/321#issuecomment-584052819 for an example of code used to generate and convert the model.

[prabhat00155]

I see what the problem is now. If we set drop=‘first’, sk2onnx removes the first category from each feature and hence when you do transform with that feature value, skl2onnx give the error, whereas scikit keeps that category value, and simply hides that category from the output. This needs to be fixed, thanks for reporting. Here is a simpler example which shows this:

X = np.array([['male', 'first'], ['female', 'first'], ['male', 'second']])
model = OneHotEncoder(drop='first').fit(X)
print(model.transform(X).toarray())
onnx_model = convert_sklearn(model, 'ohe', [('input', StringTensorType([None, 2]))])
save_model(onnx_model, 'ohe.onnx')
sess = InferenceSession('ohe.onnx')
res = sess.run(None, input_feed={'input': X})

[[1. 0.]
 [0. 0.]
 [1. 1.]]
---------------------------------------------------------------------------
Fail                                      Traceback (most recent call last)
<ipython-input-57-66d5e90d517b> in <module>
      5 save_model(onnx_model, 'ohe.onnx')
      6 sess = InferenceSession('ohe.onnx')
----> 7 res = sess.run(None, input_feed={'input': X})
      8 res

~/Documents/MachineLearning/onnx_projects/skl_env/lib/python3.6/site-packages/onnxruntime/capi/session.py in run(self, output_names, input_feed, run_options)
    140             output_names = [output.name for output in self._outputs_meta]
    141         try:
--> 142             return self._sess.run(output_names, input_feed, run_options)
    143         except C.EPFail as err:
    144             if self._enable_fallback:

Fail: [ONNXRuntimeError] : 1 : FAIL : Non-zero status code returned while running OneHotEncoder node. Name:'OneHotEncoder1' Status Message: Unknown Category and zeros = 0.

If you print model.categories_, you still see all the categories in training data: [array([‘female’, ‘male’], dtype=‘<U6’), array([‘first’, ‘second’], dtype=‘<U6’)]

[victornoel]

@prabhat00155 tomorrow I will try to make a repro using one of the sklearn-provided datasets 😃