Use "partial_dependence" function with Autogluon's TabularPredictor

[ArijitSinghEDA]

I am trying something like this:

from sklearn.inspection import partial_dependence

res = partial_dependence(model, train_data, train_data.columns)

This throws an error

ValueError: 'estimator' must be a fitted regressor or classifier.

Please note that model is a trained tabular predictor on the data train_data
I want the points only, I do not want to plot these value directly (else woukd have used autogluon.eda.auto.partial_dependence_plots)

I am using Autogluon 0.8.3 and scikit-learn 1.3.0

[adrinjalali]

Please leave a minimal reproducer (something we can copy/paste in its entirety and get the issue) or else we can't really diagnose the issue.

[ArijitSinghEDA]

@adrinjalali

from autogluon.tabular import TabularDataset, TabularPredictor
from sklearn.inspection import partial_dependence
from sklearn.datasets import load_wine
import pandas as pd

X = load_wine(as_frame=True)["data"]
y = load_wine(as_frame=True)["target"]
df = pd.concat([X, y], axis=1, join="inner")

data = TabularDataset(df)
label = "target"
model = TabularPredictor(label=label).fit(data)
print(partial_dependence(model, data.drop(columns=[label]), ["feat1"]))

[adrinjalali]

Please make sure the code only depends on scikit-learn, and that the data is generated in the script. We won't be downloading dataset files from the internet.

[ArijitSinghEDA]

@adrinjalali The code has been updated as required

[adrinjalali]

That still uses from autogluon.tabular import TabularDataset, TabularPredictor, if they're the culprit, then you need to open an issue on their side, not scikit-learn.