Need Help in Understanding RNN input sequences vs number of cells used in the layer

How can a RNN layer process a input sequence of length 100 with only 32 cells? Is the RNN output limited to 32 or how does it handle it.
Eg: It is machine translation task. Your sentence is 100 words long. Translated sentence must be 110, but you have only 32 rnn cells in the layer. How does the model output the 110 expected words.

An LSTM layer consists of a bunch of stacked LSTM cells.

You feed the 1st word to the 1st lstm cell. The 1st cell feeds the outputs to the next LSTM cell connected to it. This process continues till the last LSTM cell from where outputs is taken for feeding to the next layer or interpretation. In short, you can consider this entire arrangement as a single LSTM cell. Now that you get this perspective, the rest is the same as using an LSTM cell for multiple timesteps.

1 Like