Features standardized with StandardScaler, fit on the training split only.
Dimensionality reduced with PCA (n_components=0.95 — 95% variance retained), reducing 64 features to ~40 components.

Model

Hard voting ensemble of three classifiers, each tuned via GridSearchCV:

Classifier	Best Params
SVM	`C`, `kernel` searched over `[0.1, 1, 10]` × `["linear", "rbf"]`
Random Forest	`n_estimators` searched over `[50, 100, 200]`
KNN	`n_neighbors` searched over `[3, 5, 7]`

Files

digit_classifier_artifact.joblib: dict with {"model", "scaler", "pca"}.
digit-image-classification.ipynb: full notebook (preprocessing, GridSearchCV, VotingClassifier, evaluation).

Usage

python

import joblib
import numpy as np
from huggingface_hub import hf_hub_download

path = hf_hub_download(repo_id="KubraParmak/digit-classifier-model", filename="digit_classifier_artifact.joblib")
artifact = joblib.load(path)

scaler = artifact["scaler"]
pca    = artifact["pca"]
model  = artifact["model"]

# X: numpy array of shape (n_samples, 64), pixel values in range 0–16
X_scaled = scaler.transform(X)
X_pca    = pca.transform(X_scaled)
predictions = model.predict(X_pca)

Performance

Test accuracy: 0.97 (VotingClassifier, hard voting, 5-fold CV).

Live Demo

See KubraParmak/digit-image-classification for an interactive Gradio demo.