Flatten of cnn-lstm

CNN and LSTM How to connect ？

I have seen it before , Flatten or not

The first way , Flattening

Suppose that the maximum length of your original sequence is 100, After word embedding , Dimension for 16, That's it [N, 100,16]. Suppose you CNN Same convolution and output 64 passageway , that CNN Then the output is [N, 100, 64]. Then you can put it flatten become [N, 6400] Input LSTM.

model = Sequential()
    model.add(Conv1D(filters=64, kernel_size=3, activation='relu',
                     input_shape=(n_timesteps, n_features)))
    model.add(Conv1D(filters=64, kernel_size=3, activation='relu'))
    model.add(MaxPooling1D(pool_size=2))
    model.add(Flatten())

    # If the input shape is （None,32）,
    #  After adding RepeatVector(3) After the layer ,
    #  The output changes to （None,3,32）,RepeatVector Don't change our step size ,
    #  Change the dimension of our every step （ namely ： Attribute length ）
    model.add(RepeatVector(n_outputs))

    model.add(LSTM(200, activation='relu', return_sequences=True))
    # TimeDistributed and Dense Used together ,
    #  Mainly used in one to many , Many to many situation .
    model.add(TimeDistributed(Dense(100, activation='relu')))
    # input_shape = (10, 16), Indicates that the step size is 10,
    #  The dimension of each step is 16,（ namely ： The attribute length of each data is 16））
    #  use first TimeDistributed（Dense（8）,input_shape = (10, 16)）
    #  The dimension of each step is 16 become 8, Do not change the size of the step 
    # eg： Enter a shape and (50, 10, 16), Then the output after this layer is (50, 10, 8)
    model.add(TimeDistributed(Dense(1)))

    model.compile(loss='mse', optimizer='adam', metrics=['accuracy'])
    print(model.summary())

The second method , No flattening

such as , You can also directly [N, 100, 64] Input LSTM.

model = Sequential()
    model.add(Conv1D(filters=64, kernel_size=3, activation='relu',
                     input_shape=(n_timesteps, n_features)))
    model.add(Conv1D(filters=64, kernel_size=6, activation='relu'))
    model.add(MaxPooling1D(pool_size=1))
    model.add(RepeatVector(n_outputs))

    model.add(LSTM(200, activation='relu', return_sequences=True))
    model.add(TimeDistributed(Dense(100, activation='relu')))
    model.add(TimeDistributed(Dense(1)))

    model.compile(loss='mse', optimizer='adam', metrics=['accuracy'])
    print(model.summary())

Results show

There's a difference ：