为什么CNN-LSTM比LSTM更快？

Question

我对加速的原因感到困惑。训练和预测速度的提高都是巨大的，超过50倍。

下面是我创建LSTM模型的方法：

def create_model(learning_rate, num_LSTM_layers,
                 num_LSTM_nodes, dropout_rate):


    #CREATE THE LSTM NEURAL NETWORK
    model = Sequential()
    if num_LSTM_layers > 1:
        model.add(LSTM(num_LSTM_nodes, return_sequences=True ))
    if num_LSTM_layers == 1:
        model.add(LSTM(num_LSTM_nodes, return_sequences=False))
    model.add(Activation('relu'))
    model.add(Dropout(dropout_rate))

    if num_LSTM_layers > 1:
        for i in range(num_LSTM_layers-1):
            if i+1 == num_LSTM_layers-1:
                model.add(layers.LSTM(num_LSTM_nodes, return_sequences=False))
            else:
                model.add(layers.LSTM(num_LSTM_nodes, return_sequences=True))
            model.add(Activation('relu'))
            model.add(Dropout(dropout_rate))

    model.add(Dense(1))
    model.add(Activation('linear'))


    # Use the Adam method for training the network.
    # We want to find the best learning-rate for the Adam method.
    optimizer = Adam(lr=learning_rate)

    # In Keras we need to compile the model so it can be trained.
    model.compile(loss='mean_squared_error', optimizer=optimizer)

    return model

下面是我创建CNN-LSTM模型的方法：

def create_model_TD(learning_rate, num_conv_layers, num_LSTM_layers,
                 num_LSTM_nodes, dropout_rate, filter_size, kernel_height, pool_size):

    #CREATE THE LSTM NEURAL NETWORK
    model = Sequential()
    model.add(TimeDistributed(Conv1D(input_shape=(None, X_train.shape[2], X_train.shape[3]) , 
                                     filters= int(filter_size), kernel_size= int(kernel_height), activation='relu', padding='causal')))
    if num_conv_layers == 2:
        model.add(TimeDistributed(Conv1D(filters=int(filter_size), kernel_size= int(kernel_height), activation='relu', padding='causal')))
    model.add(TimeDistributed(MaxPooling1D(pool_size=int(pool_size))))
    model.add(TimeDistributed(Flatten()))
    if num_LSTM_layers > 1:
        model.add(LSTM(num_LSTM_nodes, return_sequences=True))
    if num_LSTM_layers == 1:
        model.add(LSTM(num_LSTM_nodes, return_sequences=False))
    model.add(Activation('relu'))
    model.add(Dropout(dropout_rate))

    if num_LSTM_layers > 1:
        for i in range(num_LSTM_layers-1):
            if i+1 == num_LSTM_layers-1:
                model.add(LSTM(num_LSTM_nodes, return_sequences=False))
            else:
                model.add(LSTM(num_LSTM_nodes, return_sequences=True))
            model.add(Activation('relu'))
            model.add(Dropout(dropout_rate))

    model.add(Dense(1))
    model.add(Activation('linear'))


    # Use the Adam method for training the network.
    # We want to find the best learning-rate for the Adam method.
    optimizer = Adam(lr=learning_rate)

    # In Keras we need to compile the model so it can be trained.
    model.compile(loss='mean_squared_error', optimizer=optimizer)

    return model

但当我查看可训练参数的数量时，CNN-LSTM似乎比经典的LSTM有更多的参数。有人知道原因吗？非常感谢你的帮助，谢谢。

Answer 1

CNN通过专注于关键功能，导致了各种不同的复杂性降低。卷积层的使用减少了张量的大小。此外，使用池化还会进一步减少。最后但并非最不重要的是，ReLu层降低了复杂性。由于训练时间的减少

cnn中更深层次的初学者观察

https://www.semanticscholar.org/paper/Introduction-to-Convolutional-Neural-Networks-Wu/450ca19932fcef1ca6d0442cbf52fec38fb9d1e5