concat operation for the Cell in model_search.py

[tengteng95]

Hi, Thanks for your great work! I am a little confused by the line 58 in your model_search.py file, which writes as following:
return torch.cat(states[-self._multiplier:], dim=1)
Following the paper, the hidden states for concat operation should be the last self._steps states, so I think it should write as following:
return torch.cat(states[-self._steps:], dim=1)
However, I am not sure whether my understanding is right. Looking forward for your response.

[tengteng95]

I have figured it out. It is right for the implementation of the code.

[lorenmt]

Hi, I also found this question and haven't figured it out. Could you please kindly explain why it's multiplier rather than steps?

[StOnEGiggity]

I think multiplier should be the same as steps. As #13 said, If the input nodes have C channels, then there will be multiplier * C channels in the output of the cell. During each step in the loop, we add one new s to the list states with C channel. And we concat all state in the list except s0, s1, and we get multiplier*C channels in the output.
However, I am not sure whether my understanding is right :)

[lorenmt]

Yes. That’s what I thought. If so, It’s not necessary to introduce a new variable then, right? Anyway, we can just keep it as it is.

[giangtranml]

I don't think it's neccessary to create a new variable multiplier, for me it makes no sense. Since, by default s0 and s1 is the input/initial nodes in a cell (a Directed Acyclic Graph), why don't we just use the variable steps instead?