Example of Many-to-One LSTM

[tom (Thomas V)]

Hello @osm3000,

Why the output shape has a hidden_size as a factor?

The output per time step has dimension hidden_size per direction (it’s the hidden layer). In LSTM the output is “modulated” cell state.

What is the num_directions?

This is 1 for “usual” LSTM and 2 for bidirectional ones.

What if I want the output sequence to have a specific size (number of features)?

Commonly, you would then use hidden_size as the target size and use the last (time direction) output per batch item, i.e.

x = output[-1] # this is batch * hidden_size for unidirectional LSTM

You can then use x as input into whatever layer you want to have above the LSTM.

Best regards

Thomas