src/train.py

import os
import glob
import librosa
import numpy as np
import pandas as pd
import scipy.signal
import matplotlib.pyplot as plt
import seaborn as sns
from sklearn.model_selection import train_test_split
from sklearn.preprocessing import StandardScaler
from sklearn.linear_model import LogisticRegression
from sklearn.ensemble import RandomForestClassifier
from sklearn.svm import SVC
from sklearn.metrics import accuracy_score, confusion_matrix, roc_curve, auc, cohen_kappa_score
import joblib

# Numpy 2.0 compatibility for librosa
if not hasattr(np, 'trapz'):
    np.trapz = np.trapezoid
if not hasattr(np, 'in1d'):
    def in1d_patch(ar1, ar2, assume_unique=False, invert=False):
        return np.isin(ar1, ar2, assume_unique=assume_unique, invert=invert)
    np.in1d = in1d_patch

# Config
DATASET_DIR = "dataset"
TARGET_SR = 16000
AUDIO_LENGTH_SEC = 5
os.makedirs("weights", exist_ok=True)
os.makedirs("metrics", exist_ok=True)

def apply_clinical_bandpass(y, sr):
    nyq = 0.5 * sr
    low = 25.0 / nyq
    high = 400.0 / nyq
    b, a = scipy.signal.butter(4, [low, high], btype='band')
    return scipy.signal.filtfilt(b, a, y)

def extract_statistical_features(y, sr):
    """Extracts 1D interpretable statistical biomarkers."""
    features = {}
    
    mfccs = librosa.feature.mfcc(y=y, sr=sr, n_mfcc=13)
    for i in range(13):
        features[f'mfcc_{i}_mean'] = np.mean(mfccs[i])
        features[f'mfcc_{i}_std'] = np.std(mfccs[i])
        
    features['centroid_mean'] = np.mean(librosa.feature.spectral_centroid(y=y, sr=sr))
    features['zcr_mean'] = np.mean(librosa.feature.zero_crossing_rate(y))
    features['rms_mean'] = np.mean(librosa.feature.rms(y=y))
    
    prob = np.square(np.abs(librosa.stft(y)))
    prob = prob / np.sum(prob)
    features['entropy'] = -np.sum(prob * np.log2(prob + 1e-10))
    
    return features

def load_dataset():
    print("Scanning dataset directory...")
    files = glob.glob(os.path.join(DATASET_DIR, "*.wav"))
    
    if not files:
        print("ERROR: No .wav files found in dataset/")
        return None, None
        
    X_features = []
    y_labels = []
    
    for f in files:
        try:
            basename = os.path.basename(f).lower()
            label = 1 if 'murmur' in basename or 'abnormal' in basename else 0
            
            y, sr = librosa.load(f, sr=TARGET_SR, mono=True)
            y = librosa.util.normalize(y)
            y_clean = apply_clinical_bandpass(y, sr)
            
            target_length = TARGET_SR * AUDIO_LENGTH_SEC
            if len(y_clean) > target_length:
                y_clean = y_clean[:target_length]
            else:
                y_clean = np.pad(y_clean, (0, target_length - len(y_clean)))
                
            feats = extract_statistical_features(y_clean, sr)
            X_features.append(feats)
            y_labels.append(label)
        except Exception as e:
            print(f"Error processing {f}: {e}")
            
    df = pd.DataFrame(X_features)
    labels = np.array(y_labels)
    
    print(f"Successfully processed {len(df)} canine recordings.")
    return df, labels

def evaluate_model(y_true, y_pred):
    acc = accuracy_score(y_true, y_pred)
    cm = confusion_matrix(y_true, y_pred, labels=[0, 1])
    
    if cm.shape == (2, 2):
        tn, fp, fn, tp = cm.ravel()
        sensitivity = tp / (tp + fn) if (tp + fn) > 0 else 0.0
        specificity = tn / (tn + fp) if (tn + fp) > 0 else 0.0
    else:
        # Handle all one class cases for tiny datasets
        sensitivity = 0.0
        specificity = 0.0
        
    return acc, sensitivity, specificity, cm

def train_and_evaluate():
    X, y = load_dataset()
    if X is None: return
    
    # Feature Scaling is critical for SVM and Logistic Regression
    scaler = StandardScaler()
    feature_names = X.columns
    X_scaled = scaler.fit_transform(X)
    X_scaled = pd.DataFrame(X_scaled, columns=feature_names)
    joblib.dump(scaler, "weights/scaler.pkl")
    joblib.dump(list(feature_names), "weights/feature_columns.pkl")
    
    # Strictly 70/30 split
    X_train, X_test, y_train, y_test = train_test_split(X_scaled, y, test_size=0.3, random_state=42)
    
    print(f"\n--- Training on {len(X_train)} samples, Testing on {len(X_test)} samples (70/30 Split) ---")

    models = {
        "Logistic Regression": LogisticRegression(max_iter=1000, random_state=42),
        "Random Forest": RandomForestClassifier(n_estimators=100, max_depth=10, random_state=42),
        "SVM (RBF)": SVC(kernel='rbf', probability=True, random_state=42)
    }

    results = {}
    y_preds_all = {}
    y_proba_all = {}

    for name, model in models.items():
        print(f"\nTraining {name}...")
        model.fit(X_train, y_train)
        
        y_pred = model.predict(X_test)
        y_proba = model.predict_proba(X_test)[:, 1]
        
        y_preds_all[name] = y_pred
        y_proba_all[name] = y_proba
        
        acc, sens, spec, cm = evaluate_model(y_test, y_pred)
        results[name] = {
            "Accuracy": acc,
            "Sensitivity": sens,
            "Specificity": spec,
            "CM": cm
        }
        
        print(f"Accuracy:    {acc*100:.1f}%")
        print(f"Sensitivity: {sens*100:.1f}%")
        print(f"Specificity: {spec*100:.1f}%")
        
        filename = name.lower().replace(" ", "_").replace("(", "").replace(")", "")
        joblib.dump(model, f"weights/canine_{filename}.pkl")

    # 1. Output ROC Curve Plot
    plt.figure(figsize=(8, 6))
    for name, y_proba in y_proba_all.items():
        fpr, tpr, _ = roc_curve(y_test, y_proba)
        roc_auc = auc(fpr, tpr)
        plt.plot(fpr, tpr, lw=2, label=f'{name} (AUC = {roc_auc:.2f})')
        
    plt.plot([0, 1], [0, 1], color='navy', lw=2, linestyle='--')
    plt.xlim([0.0, 1.0])
    plt.ylim([0.0, 1.05])
    plt.xlabel('False Positive Rate (1 - Specificity)')
    plt.ylabel('True Positive Rate (Sensitivity)')
    plt.title('Receiver Operating Characteristic (ROC) Comparison')
    plt.legend(loc="lower right")
    plt.grid(True, alpha=0.3)
    plt.savefig('metrics/roc_curve.png')
    plt.close()
    
    # 2. Confusion Matrices Plot
    fig, axes = plt.subplots(1, 3, figsize=(15, 4))
    for ax, (name, res) in zip(axes, results.items()):
        sns.heatmap(res["CM"], annot=True, fmt='d', cmap='Blues', ax=ax, cbar=False)
        ax.set_title(f'{name}\nAcc: {res["Accuracy"]:.2f}')
        ax.set_xlabel('Predicted Label')
        ax.set_ylabel('True Label')
        ax.set_xticklabels(['Normal (0)', 'Murmur (1)'])
        ax.set_yticklabels(['Normal (0)', 'Murmur (1)'])
    plt.tight_layout()
    plt.savefig('metrics/confusion_matrix.png')
    plt.close()

    # 3. Random Forest Feature Importance Plot
    rf_model = models["Random Forest"]
    importances = rf_model.feature_importances_
    indices = np.argsort(importances)[::-1][:15] # Top 15 features
    
    plt.figure(figsize=(10, 6))
    plt.title("Top 15 Feature Importances (Random Forest)")
    plt.bar(range(15), importances[indices], align="center", color='skyblue', edgecolor='black')
    plt.xticks(range(15), [feature_names[i] for i in indices], rotation=45, ha='right')
    plt.xlim([-1, 15])
    plt.tight_layout()
    plt.savefig('metrics/feature_importance.png')
    plt.close()
    
    # 4. Model Agreement (Kappa between RF and SVM)
    kappa = cohen_kappa_score(y_preds_all["Random Forest"], y_preds_all["SVM (RBF)"])
    print(f"\n--- Model Agreement ---")
    print(f"Cohen's Kappa (Random Forest vs SVM): {kappa:.3f}")
    
    print("\nTraining Pipeline Complete.")
    print("Interpretable Models saved to weights/")
    print("Clinical visual metrics saved to metrics/")

if __name__ == "__main__":
    train_and_evaluate()