How to define Dense layer between Conv layer and LSTM layer?

[hassaku]

I try to implement the following model called LSDNN.

CONVOLUTIONAL, LONG SHORT-TERM MEMORY, FULLY CONNECTED DEEP NEURAL NETWORKS http://static.googleusercontent.com/media/research.google.com/ja//pubs/archive/43455.pdf

Input -> Conv(+Pool) -> Dense (to reduce dimension) -> LSTM -> Dense -> Output

model = Sequential()
model.add(Convolution1D(..., input_shape=(timesteps, data_dim)))
model.add(MaxPooling1D(...))
model.add(Dense(...))  # to reduce dimension
model.add(LSTM(...))
model.add(Dense(...))
model.add(Dense(nb_classes, activation='sigmoid'))
model.compile(loss='binary_crossentropy', optimizer='rmsprop', metrics=['accuracy'])

model.fit(...)  # Exception: Input 0 is incompatible with layer dense_1: expected ndim=2, found ndim=3

If the first Dense layer is removed, exception not occurs. But dimension isn't reduced.

How to use Dense layer keeping input shape for LSTM? Thank you for any help someone can provide.

[carlthome]

LSTM() expects input of shape (nb_timesteps, nb_features) but Dense outputs 1D so use Reshape() to add a dummy dimension after the fully-connected layer or use TimeDistributed(Dense()).

[hassaku]

Thank you for your advice. It's solved by using TimeDistributed now.