Does F.lstm exist?

autograd
1

Is there a functional version of torch.nn.lstm?

2

I don’t think there is a functional API in torch.nn.functional, but torch._VF.lstm is used inside the nn.LSTM module as seen here so you might want to use it.

3

Thanks! @ptrblck

So If I wanted to use a Bidirectional LSTM, that receives as input a torch.tensor of shape (batch, timesteps, features):

hx should be a tensor of zeros as they do in the docs:

            num_directions = 2
            real_hidden_size = self.proj_size if self.proj_size > 0 else self.hidden_size
            h_zeros = torch.zeros(self.num_layers * num_directions,
                                  max_batch_size, real_hidden_size,
                                  dtype=input.dtype, device=input.device)
            c_zeros = torch.zeros(self.num_layers * num_directions,
                                  max_batch_size, self.hidden_size,
                                  dtype=input.dtype, device=input.device)
            hx = (h_zeros, c_zeros)

And but I’m having trouble understanding how they flatten the weights before passing it to the layer. Is it a list of the different LSTM variables flatten in a certain order?

4

The order of the parameters is given in the docs for nn.LSTM in the Variables section.