Mlflow: Proposal:Add the function returning the result of predict_proba() to Model API

[masa26hiro]

System information

• OS Platform and Distribution (e.g., Linux Ubuntu 16.04):Windows10(Docker : Debian GNU/Linux 9)
• MLflow installed from (source or binary): source
• MLflow version (run mlflow --version):0.7.0
• Python version: 3.7.0

Describe the problem

Currently, the serve_model method of mlflow/sklearn.py returns only "model.predict(records)".
I think that there are times when we want to know the classification probability of each class when classifying.
So I suggest adding a function to return "model.predict_proba(records)" results to this method.

Thank you.

[MichaelMarien]

Hi,

a quick fix would be to overwrite the "predict" function of your sklearn model with "predict_proba" before saving it. This is a bit hacky though. Not sure what the view of the core dev team is, but if they want to adress this more nicely, I'd be happy to pick it up.

[masa26hiro]

Hi,

Thanks for your reply.
I used to rewrite predict () into predict_proba () by myself before too. I believe there is a demand for data scientists who wish to use predict_proba ().

[yassineAlouini]

I have a similar use case but with another custom function predict_distribution.
For now, I will override the predict method during the training but it would be better to offer a way to do it natively. Any help or suggestions are welcome. Thanks in advance. :)

[smurching]

Hi all, sorry for the delayed response - this should be possible in MLflow >= 1.0 by using custom Python models (see the docs), in particular this example - you can define a wrapper model class that loads the persisted sklearn model & returns the results of predict_proba. Please feel free to reopen with any questions!