scikit-learn support for predict method not only predict_proba

[smolina74]

I was trying my own scikit-learn Iris model. In that one I'm using the Seldon wrapper (not s2i, since I'm getting the error described here kubeflow/example-seldon#13). Now, one thing I noticed there is that the module in charge of calling my model class is model_microservice.py. And for what I'm seeing, after calling the prediction method it is expecting that the result is a 2 dimensional array (prediction.shape[1] will fail if not), which is ok when you call a predict_proba, but not when you just call predict.
So, the question is, what can I do if I just need to call a predict method and not predict_proba?

[ukclivecox]

Its true the present python wrapper is assuming this.

seldon-core/wrappers/python/model_microservice.py

Lines 54 to 55 in befa040

	# TODO: check that predictions is 2 dimensional
	class_names = get_class_names(user_model, predictions.shape[1])

It is using this to get the class names. But we could make this only used if the predictions were 2 dimensional. The core data payload sent back doesn't place any requirements on the shape of the data.

[paul-crease]

@smolina74 - I had the same problem, and found a workaround with self.model.predict(X).reshape(1, -1) which then gets you a 2d array.

[ukclivecox]

Please reopen if still an issue with latest Python wrapper.