Implementation:Kserve Kserve RFModel AIX MNIST

Knowledge Sources	Kserve_Kserve
Domains	Model Serving, Explainability
Last Updated	2026-02-13 00:00 GMT

Overview

Concrete tool for serving a Random Forest model for MNIST digit classification with AIX explainability support provided by the KServe sample code.

Description

This module defines two classes: PipeStep and RFModel. PipeStep is a simple sklearn pipeline wrapper that turns arbitrary functions into pipeline-compatible transforms (with no-op fitting). RFModel extends kserve.Model and loads a pickled Random Forest pipeline from disk. Its predict() method accepts image instances, converts them to NumPy arrays, reshapes them from 4D image arrays to 2D flattened vectors (n_samples x 2352), runs predict_proba on the model, and reformats the multi-class probability output into a per-class predictions structure.

Usage

Use this class when deploying a Random Forest MNIST classifier as a KServe inference service, particularly when paired with the AIX explainer for generating image-based explanations.

Code Reference

Source Location

• Repository: Kserve_Kserve
• File: docs/samples/explanation/aix/mnist/rfserver/rfserver/model.py
• Lines: 1-87

Signature

class PipeStep(object):
    def __init__(self, step_func):
        ...

    def fit(self, *args):
        ...

    def transform(self, X):
        ...


class RFModel(kserve.Model):
    def __init__(self, name: str):
        ...

    def load(self) -> bool:
        ...

    def predict(
        self,
        payload: Union[Dict, InferRequest, ModelInferRequest],
        headers: Dict[str, str] = None,
    ) -> Union[Dict, InferResponse, ModelInferResponse]:
        ...

Import

from rfserver.model import PipeStep, RFModel

I/O Contract

Inputs

PipeStep Constructor

Name	Type	Required	Description
step_func	Callable	Yes	A function to be used as a pipeline transform step

RFModel Constructor

Name	Type	Required	Description
name	str	Yes	The name of the model for KServe routing

predict()

Name	Type	Required	Description
payload	Union[Dict, InferRequest, ModelInferRequest]	Yes	Dictionary with "instances" key containing image data arrays
headers	Dict[str, str]	No	Optional HTTP headers from the request

Outputs

load()

Name	Type	Description
ready	bool	Returns True when the model is successfully loaded from pickle

predict()

Name	Type	Description
predictions	Dict	Dictionary with "predictions" key containing per-class probability lists

Usage Examples

Basic Usage

from rfserver.model import RFModel

# Create and load the Random Forest model
model = RFModel("mnist-rf")
model.load()

# Predict on MNIST-like image data (batch of images)
import numpy as np
sample_input = np.random.rand(1, 28, 28, 3).tolist()
result = model.predict({"instances": sample_input})
print(result)  # {"predictions": [[0.1, 0.9, ...], ...]}

Related Pages

• Environment:Kserve_Kserve_Python_Runtime