classification_example.py

#!/usr/bin/env python3
"""
KortexDL Classification - Iris Dataset (Stable)
================================================

Conservative hyperparameters for stable training.

Usage:
    python classification_example.py
"""

import numpy as np
import kortexdl as bd

try:
    from sklearn.datasets import load_iris
    from sklearn.preprocessing import StandardScaler
    from sklearn.model_selection import train_test_split
    SKLEARN_AVAILABLE = True
except ImportError:
    SKLEARN_AVAILABLE = False


def main():
    print("🎯 KortexDL Classification - Iris Dataset")
    print("=" * 60)
    
    if not SKLEARN_AVAILABLE:
        print("❌ sklearn required: pip install scikit-learn")
        return 1
    
    # Load Iris dataset
    print("\n📁 Loading Iris dataset...")
    data = load_iris()
    X = data.data.astype(np.float32)
    y_raw = data.target
    
    # One-hot encode
    n_classes = 3
    y = np.zeros((len(y_raw), n_classes), dtype=np.float32)
    for i, label in enumerate(y_raw):
        y[i, label] = 1.0
    
    print(f"✅ Dataset: {len(X)} samples, {X.shape[1]} features, {n_classes} classes")
    
    # Normalize
    scaler = StandardScaler()
    X = scaler.fit_transform(X).astype(np.float32)
    
    # Split with stratification
    X_train, X_test, y_train, y_test = train_test_split(
        X, y, test_size=0.2, random_state=42, stratify=y_raw
    )
    print(f"✅ Split: {len(X_train)} train, {len(X_test)} test")
    
    # Simple network with Sigmoid (bounded outputs)
    print("\n🧠 Creating network...")
    net = bd.Network([4, 16, 8, 3], bd.ActivationType.Sigmoid)
    print("✅ Network: 4 -> 16 -> 8 -> 3 (Sigmoid)")
    
    # Conservative learning rate, many epochs
    print("\n🏋️ Training...")
    epochs = 1000
    learning_rate = 0.5
    
    X_flat = X_train.flatten().tolist()
    y_flat = y_train.flatten().tolist()
    
    for epoch in range(epochs):
        loss = net.train_batch(X_flat, y_flat, bd.LossType.MSE, learning_rate, len(X_train))
        
        if epoch % 200 == 0:
            correct = sum(1 for i in range(len(X_test)) 
                         if np.argmax(net.forward(X_test[i].tolist(), 1, False)) == np.argmax(y_test[i]))
            acc = correct / len(X_test) * 100
            print(f"  Epoch {epoch:4d}: Loss = {loss:.4f}, Acc = {acc:.1f}%")
    
    # Final evaluation
    print("\n📈 Final Evaluation...")
    correct = 0
    class_names = ['Setosa', 'Versicolor', 'Virginica']
    
    for i in range(len(X_test)):
        output = net.forward(X_test[i].tolist(), 1, False)
        pred = np.argmax(output)
        true = np.argmax(y_test[i])
        if pred == true:
            correct += 1
    
    accuracy = correct / len(X_test) * 100
    print(f"✅ Test Accuracy: {accuracy:.1f}%")
    
    print("\n📊 Sample Predictions:")
    for i in range(min(10, len(X_test))):
        output = net.forward(X_test[i].tolist(), 1, False)
        pred = np.argmax(output)
        true = np.argmax(y_test[i])
        status = "✓" if pred == true else "✗"
        print(f"  {class_names[true]:12s} -> {class_names[pred]:12s} {status}")
    
    print("\n" + "=" * 60)
    print("✅ Complete!")
    return 0


if __name__ == "__main__":
    exit(main())