Why does nn.RNN's default input is (seq_len, batch, input_size) instead of (batch, seq_len, input_size)?


(Ilija Ilievski) #1

Is there some performance gain?
It’s more natural to me to have the batch size as the first dimension.


(Serdar) #2

I guess in RNN at every time insantance what will be update is
Data(t+1,:,:)=f(Data(t,:,:))
So more natural seems to be their choice.