LSTM Autoencoder - Encoder Output shape

1

Hi Community!

I hope it’s okay to just post my question here, as I couldn’t find a suitable subforum at first.
I have problems creating an autoencoder with LSTM layers.

An LSTM returns the following output: outputs, (hn, cn) = self.LSTM(...)

Since the last hidden state hn can be used as input for the decoder in an autoencoder I have to transform it into the right shape.
I have seen many different approaches on the Internet and am now unsure how to proceed.

According to the Pytorch documentation, the hidden state hn has the following shape: (1*num_layers, batch_size, hidden_size) # unidirectional and batch_first=True

How do I reshape the hidden state to pass it to the decoder as a compressed vector in the correct shape?

Thanks in advance for any help!

Kind regards,
Christopher

2

Welcome to the forums!

You won’t need to reshape anything if the hidden size of the LSTM decoder is the same size as the hidden of the encoder.

IMG_20240108_084052
IMG_20240108_084052920×573 98.6 KB

If the sizes aren’t the same, you can pass it through a Linear layer as is(no need to reshape or permute; Linear layers are only concerned with the final dim size).

hidden_nn = nn.Linear(enc_hidden, dec_hidden)

h0 = hidden_nn(hn)

The above is a simple case, but you may also try using multiple Linear layers with activations before continuing to the decoder.

1 Like
3

Thank you very much for your response!
Now it works as expected :blush:

1 Like