becaliang/Music_Generation_Project

1import numpy as np2from keras.models import Sequential3from keras.layers import LSTM, Dense, Dropout, Activation4from keras.utils import to_categorical5 6def _one_hot_encode_sequences(X_data, y_data, token_to_index, sequence_length):7    vocab_size = len(token_to_index)8    X = np.zeros((len(X_data), sequence_length, vocab_size), dtype=np.float32)9    y = np.zeros((len(X_data), vocab_size), dtype=np.float32)10 11    for i, seq in enumerate(X_data):12        for t, token in enumerate(seq):13            if token in token_to_index:14                X[i, t, token_to_index[token]] = 1.015        if y_data[i] in token_to_index:16            y[i, token_to_index[y_data[i]]] = 1.017    return X, y18 19def build_multi_instrument_model(track_corpora, token_to_index, sequence_length=20, N_epochs=64):20    """21    Builds and trains an LSTM model for symbolic multi-instrument music generation.22 23    Parameters:24    - track_corpora: list of token sequences (each list = 1 instrument track)25    - token_to_index: dictionary mapping token -> index26    - sequence_length: number of tokens in each input sequence27    - N_epochs: number of training epochs28    """29    stride = 330    vocab_size = len(token_to_index)31    X_data, y_data = [], []32 33    # === Build training data ===34    for idx, corpus in enumerate(track_corpora):35        track_name = f"Track_{idx + 1}"36        if len(corpus) <= sequence_length:37            print(f"⚠️ Skipping {track_name}: too short ({len(corpus)} tokens).")38            continue39        for i in range(0, len(corpus) - sequence_length, stride):40            X_data.append(corpus[i:i + sequence_length])41            y_data.append(corpus[i + sequence_length])42 43    if len(X_data) == 0:44        # Fallback: pad with dummy "<PAD>" tokens45        print("⚠️ No valid sequences. Padding with dummy data to allow model structure initialization.")46        dummy_token = "<PAD>"47        if dummy_token not in token_to_index:48            token_to_index[dummy_token] = vocab_size49            vocab_size += 150        dummy_seq = [dummy_token] * sequence_length51        X_data = [dummy_seq, dummy_seq]52        y_data = [dummy_token, dummy_token]53 54    print(f"✅ Total sequences: {len(X_data)} | Vocabulary size: {vocab_size}")55 56    # === One-hot encode ===57    X, y = _one_hot_encode_sequences(X_data, y_data, token_to_index, sequence_length)58 59    # === Build LSTM model ===60    model = Sequential([61        LSTM(256, return_sequences=True, input_shape=(sequence_length, vocab_size)),62        Dropout(0.3),63        LSTM(256),64        Dropout(0.3),65        Dense(vocab_size),66        Activation("softmax")67    ])68 69    model.compile(70        loss="categorical_crossentropy",71        optimizer="rmsprop",72        metrics=["accuracy"]73    )74 75    model.fit(X, y, batch_size=64, epochs=N_epochs, verbose=1)76 77    return model