char-rnn.py 文件源码

def RNN(X_train,y_train,args):
    """
    Purpose -> Define and train the proposed LSTM network
    Input   -> Data, Labels and model hyperparameters
    Output  -> Trained LSTM network
    """
    #Sets the model hyperparameters
    #Embedding hyperparameters
    max_features = args[0]
    maxlen = args[1]
    embedding_size = args[2]
    # Convolution hyperparameters
    filter_length = args[3]
    nb_filter = args[4]
    pool_length = args[5]
    # LSTM hyperparameters
    lstm_output_size = args[6]
    # Training hyperparameters
    batch_size = args[7]
    nb_epoch = args[8]
    numclasses = args[9]
    test_size = args[10] 

    #Format conversion for y_train for compatibility with Keras
    y_train = np_utils.to_categorical(y_train, numclasses) 
    #Train & Validation data splitting
    X_train, X_valid, y_train, y_valid = train_test_split(X_train, y_train, test_size=test_size, random_state=42)

    #Build the sequential model
    # Model Architecture is:
    # Input -> Embedding -> Conv1D+Maxpool1D -> LSTM -> LSTM -> FC-1 -> Softmaxloss
    print('Build model...')
    model = Sequential()
    model.add(Embedding(max_features, embedding_size, input_length=maxlen))
    model.add(Convolution1D(nb_filter=nb_filter,
                            filter_length=filter_length,
                            border_mode='valid',
                            activation='relu',
                            subsample_length=1))
    model.add(MaxPooling1D(pool_length=pool_length))
    model.add(LSTM(lstm_output_size, dropout_W=0.2, dropout_U=0.2, return_sequences=True))
    model.add(LSTM(lstm_output_size, dropout_W=0.2, dropout_U=0.2, return_sequences=False))
    model.add(Dense(numclasses))
    model.add(Activation('softmax'))

    # Optimizer is Adamax along with categorical crossentropy loss
    model.compile(loss='categorical_crossentropy',
                optimizer='adamax',
                metrics=['accuracy'])


    print('Train...')
    #Trains model for 50 epochs with shuffling after every epoch for training data and validates on validation data
    model.fit(X_train, y_train, 
              batch_size=batch_size, 
              shuffle=True, 
              nb_epoch=nb_epoch,
              validation_data=(X_valid, y_valid))
    return model