hidden state not updated

[chrisliu54]

The hidden state of GridLSTM is not updated:

GridLSTM-PyTorch/gridlstm.py

Lines 191 to 201 in fe063f2

	for dim in range(self.dimensions):
	c = C[dim]
	if dim in self.lstm_dim_map:
	# LSTM update
	cell = self.lstms[self.lstm_dim_map[dim]]
	h, c = cell.forward(inp, (H, c))
	else:
	# Non-LSTM update
	layer = self.nonlstms[self.nonlstm_dim_map[dim]]
	h = layer.forward(H)
	local_state[dim] = (h, c)

BTW, pytorch will check the consistency of the input hidden tensor's size and the self.hidden_size:

https://github.com/pytorch/pytorch/blob/6ebcb4606f079b9152cb242b36e03b8eddcb6173/torch/nn/modules/rnn.py#L504-L513

But GridLSTM's input H is concatenated from all the h which is larger than your setting:

GridLSTM-PyTorch/gridlstm.py

Line 223 in fe063f2

self.hidden_size = output_hidden_channels

[kkleidal]

Hmm, I'll have to take a look. Tbh, I was planning on using this repo for some of research, but it never really panned out, so it is not tested and could have more bugs like this. I'll post a disclaimer for now on the README so people don't end up trying to use it out of the box and run into problems like this.

[chrisliu54]

Sure, happy coding : )