Basic Serving of Sklearn Models

This topic describes how Sklearn models are handled inside of the Docker containers deployed on Seldon Core.

Preprocessing

All data are converted into a pandas DataFrame before being processed by the model. If the user passes in the list of names, then it will be applied to the columns of the DataFrame.

if names:     df = pd.DataFrame(X, columns=names) else:     df = pd.DataFrame(X)

 

The choice of whether to use the predict or predict_proba method on the model is chosen based on the meta passed in through the request. That is, the meta dictionary will be checked for the method key-value. If the value of method is one of predict or predict_proba, then that method will be run for the model. The default method is predict.

method = self.default_method if meta and isinstance(meta.get("method", 0), str):     method = meta["method"]

 

Predictions

Based on the method chosen by the user, the predict or predict_proba method of the model is executed. The total predictions are reshaped into a 2D numpy array.

if method == "predict_proba":     preds = self.model.predict_proba(df) else:     preds = self.model.predict(df)      result = preds.reshape(len(df), -1)

 

Methods for Different Model Types

Method Type	Valid for Regression?	Valid for Classification?	Valid for Clustering?
predict	Yes	Yes	Some**
predict_proba	No	Most*	No

 

* Some sklearn classification models do not have a predict_proba method

** Some sklearn clustering models do not have a predict method

The names list can be passed in the request at the user’s discretion. Generally speaking, if the model has any column-specific transformations, such as a column transformer from sklearn, then they will need to pass the names list in the request.