predict_proba does not return probabilities after load_model in VerticaPy >= 1.0.5

[okankcb]

Hello,

We encountered an issue using predict_proba after calling load_model in VerticaPy versions 1.0.5 and 1.1.1.

After loading a saved model using load_model, calling predict_proba executes without error, but does not return the expected probability columns.

Work fine in :

• VerticaPy 0.12

Issue reproducible in :

• Verticapy 1.0.5
• Verticapy 1.1.1

Reproducible example with VerticaPy 1.0.5 and 1.1.1 :

from verticapy.machine_learning.vertica import RandomForestClassifier
drop(name = schema +".model_titanic_test_rfc", method = "model")
model = RandomForestClassifier(name = schema +".model_titanic_test_rfc")
model.fit(titanic,["age","fare"], "survived")

from verticapy.machine_learning.vertica.model_management import load_model
model=load_model(schema + ".model_titanic_test_rfc")

model.predict_proba(titanic, 
                    X = ["age","fare"],
                    name = "predict_proba")

As a temporary solution while waiting for a fix, saving and loading the model using Python's pickle library works correctly. Example :

import pickle
from verticapy.machine_learning.vertica import RandomForestClassifier

drop(name = schema + ".model_titanic_recette_test_rfc", method = "model")
model = RandomForestClassifier(name = schema + ".model_titanic_recette_test_rfc")
model.fit(titanic_recette, ["age", "fare"], "survived")

with open("model_titanic_rfc.pkl", "wb") as f:
    pickle.dump(model, f)

with open("model_titanic_rfc.pkl", "rb") as f:
    model_loaded = pickle.load(f)

model_loaded.predict_proba(titanic_recette, ["age", "fare"])

Please let us know if :

• A specific parameter must be set to return the probability columns after load_model
• Or if a fix or patch is planned

Best regards,

Okan Kocabiyik

[Koumboss]

Hi Okan,

Indeed, the load_model function looks deprecated and undocumented in latest versions.
https://www.vertica.com/python/old/documentation-0.8.1/learn/load_model/

You can instead import models using import_models.

In my case, commenting out the load_model line returned the prediction column.

from verticapy.machine_learning.vertica import RandomForestClassifier
#drop(name = schema +".model_titanic_test_rfc", method = "model")
model = RandomForestClassifier(name = "public.model_titanic_test_rfc")
model.fit(titanic,["age","fare"], "survived")

#from verticapy.machine_learning.vertica.model_management import load_model
#model=load_model("public.model_titanic_test_rfc")

model.predict_proba(titanic, 
                    X = ["age","fare"],
                    name = "proba_predict")

Can you confirm this is what you needed please ? Thanks.

[okankcb]

Thank you for you response. Indeed, the predict_proba function returns the prediction column after a fit without using load_model. Whereas load_model allowed loading a saved model. How can we proceed now if load_model is obsolete for saving a model, loading it, and performing a predict_proba that returns probabilities after loading the model ?

Thanks

[mail4umar]

Good catch. There was a slight bug with how the models were loaded for RF and XGB classifiers. Creating a PR to fix this.

[mail4umar]

This is resolved in: #1358