CODEWITHSUNDEEP
machine-learningkeraslstmrecurrent-neural-network
Chakra
Chakra

Reputation: 746

RNN: What is the use of return_sequences in LSTM layer in Keras Framework

I am working in RNN. I have following lines of code from some site. If you observe second layer has no "returnSequence" parameter.

I am assuming return sequence is mandatory as it should return the sequences. Can you please tell why this is not defined.

First layer LSTM:

regressor.add(LSTM(units = 30, return_sequences = True))

Second layer LSTM:

regressor.add(LSTM(units = 30))

Upvotes: 7

Views: 6740

Answers (2)

Ali karimi
Ali karimi

Reputation: 500

when the LSTM layer is followed by a Dense layer you should set return_sequence=False, otherwise when you use several LSTM layers stacked together you should set it to TRUE

LSTM(units, return_sequences=True...)
LSTM(units, return_sequences=True...)
LSTM(units, return_sequences=False...)
Dense(...)

Upvotes: 0

KonstantinosKokos
KonstantinosKokos

Reputation: 3473

When the return_sequences argument is set to False (default), the network will only output hn, i.e. the hidden state at the final time step. Otherwise, the network will output the full sequence of hidden states, [h1, h2, ..., hn]. The internal equations of the layer are unchanged. Refer to the documentation.

Upvotes: 15

Related Questions