import numpy as np from sklearn.datasets import fetch_openml from sklearn.model_selection import train_test_split from sklearn.preprocessing import StandardScaler RANDOM_STATE = 42 def load_data(n_dev=10000): print("Loading MNIST (this may take a minute the first time)...") mnist = fetch_openml('mnist_784', version=1, as_frame=False, parser='auto', data_home='d:/sklearn_data') X, y_digits = mnist.data, mnist.target.astype(int) # Binary relabeling: +1 if digit is 6, -1 otherwise y = np.where(y_digits == 6, 1, -1) print(f"Total samples: {X.shape[0]}") print(f"Class +1 (is 6): {(y == 1).sum():5d} ({(y == 1).mean()*100:.1f}%)") print(f"Class -1 (not 6): {(y == -1).sum():5d} ({(y == -1).mean()*100:.1f}%)") # Normalize pixels to [0, 1] X = X / 255.0 # Train / Validation / Test split (60 / 20 / 20) X_trainval, X_test, y_trainval, y_test = train_test_split( X, y, test_size=0.20, stratify=y, random_state=RANDOM_STATE) X_train, X_val, y_train, y_val = train_test_split( X_trainval, y_trainval, test_size=0.25, stratify=y_trainval, random_state=RANDOM_STATE) print(f"Train: {X_train.shape[0]}, Val: {X_val.shape[0]}, Test: {X_test.shape[0]}") # Standardize — fit on train only to avoid data leakage scaler = StandardScaler() X_train_s = scaler.fit_transform(X_train) X_val_s = scaler.transform(X_val) X_test_s = scaler.transform(X_test) # Dev subsample for faster training during development rng = np.random.RandomState(RANDOM_STATE) dev_idx = rng.choice(len(X_train_s), size=n_dev, replace=False) X_dev, y_dev = X_train_s[dev_idx], y_train[dev_idx] print(f"Dev set: {len(y_dev)} samples " f"({(y_dev == 1).sum()} sixes, {(y_dev == -1).sum()} non-sixes)") return X_train_s, X_val_s, X_test_s, y_train, y_val, y_test, X_dev, y_dev, scaler