How to get probability score for each prediction from mlflow

[Zoumana]

I trained my model and was able to get the batch prediction from that model as specified below. But I want to also get the probability scores for each prediction. Do you have any idea?

Thank you!

logged_model = path_to_model

# Load model as a PyFuncModel.

loaded_model = mlflow.pyfunc.load_model(logged_model)

# Predict on a Pandas DataFrame.

import pandas as pd

loaded_model.predict(pd.DataFrame(data))

[Kaniz]

Hi @ Zoumana! My name is Kaniz, and I'm the technical moderator here. Great to meet you, and thanks for your question! Let's see if your peers in the community have an answer to your question first. Or else I will get back to you soon. Thanks.

[Zoumana]

Hi Kaniz,

Great to meet you too!

Thank you for replying to my question.

Best!

[Kaniz]

Hi @Zoumana Keita​ , Please refer to this doc here:-

For most models in sci-kit-learn, we can get the probability estimates for the classes through predict_proba.

https://scikit-learn.org/stable/modules/generated/sklearn.linear_model.LogisticRegression.html#sklea...

[SyedGhouri]

Hi @Kaniz Fatma​ 

Sorry for hijacking the post.

My question is - if I am reading a registered model from mlflow, I can only see the option of .predict method but not .predict_proba.

Do we have any straightforward solution to get the probabilities?

Thanks

Syed

[Kaniz]

Hi @Syed Ghouri​, Can you paste the error stack here?

@Syed Ghouri​, Also, I would suggest you go through this S.O thread. Please LMK how it goes.

[SyedGhouri · Accepted solution]

Hi @Kaniz Fatma​ 

The error said 'PyFuncModel' object has no attribute 'predict_proba'.

As shows above, I was using the following to load the model

loaded_model = mlflow.pyfunc.load_model(logged_model) and got the error.

After going through mlflow documentation, I changed it to

loaded_model = mlflow.sklearn.load_model(logged_model) and it is working fine.

It's all good now. Thanks for your time.

Syed

[Kaniz]

Hi @Syed Ghouri​ ​, It would mean a lot if you could select the "Best Answer" to help others find the correct answer faster.

This makes that answer appear right after the question, so it's easier to find within a thread.

It also helps us mark the question as answered so we can have more eyes helping others with unanswered questions.

Can I count on you?

[SyedGhouri]

Hi @Kaniz Fatma​ 

I do not see the option to select "Best Answer" but feel free to do anything that you think can help this community.

Thanks

Syed

[Kaniz]

Thank you for your response @Syed Ghouri​ !

I've selected the "Best Answer" on your behalf.

[OndrejHavlicek]

Now you can log the model using this parameter:

mlflow.sklearn.log_model(
    ...,  # the usual params
    pyfunc_predict_fn="predict_proba"
)

 which will return probabilities for the first class apparently when using the model for inference (e.g. when loading it using mlflow.pyfunc.spark_udf() ).