I'm not sure whether you will revisit this repo or not, but I really think it's a perfect idea behind your paper, I keen to reproduce it as soon as possible, when I read through your model , I raise a question for your consideration, original code is as following:
rnninput = y[i, :, 1:]
rnn_input = torch.cat((rnn_input, output), dim=1)
rnn_input = rnn_input.unsqueeze(0)
But according to the Decoder logic, it maybe better to swipe it as:
rnn_input = torch.cat((output,rnn_input), dim=1)
Otherwise, the sequence will not align with the first row of data.
I'm not sure whether you will revisit this repo or not, but I really think it's a perfect idea behind your paper, I keen to reproduce it as soon as possible, when I read through your model , I raise a question for your consideration, original code is as following: rnninput = y[i, :, 1:] rnn_input = torch.cat((rnn_input, output), dim=1) rnn_input = rnn_input.unsqueeze(0)
But according to the Decoder logic, it maybe better to swipe it as: rnn_input = torch.cat((output,rnn_input), dim=1) Otherwise, the sequence will not align with the first row of data.
Hope to hear from you.