Open linfeng-du opened 1 year ago
Ah I seem to get it. Isometric convolution is applied after downsampling so the sequence length is already reduced. Here is to calculate the sequence length after downsampling. May the author confirm if that's correct? Thanks!
In the paper it is said that the kernel size of the isometric convolution is set to equal the input sequence length. However from the code,
kernel_size=(seq_len + pred_len + ksize - ksize % 2) // ksize
. Is this intended?