Why conv1d of pytorch is guranteed to be causal? about tcn HOT 4 CLOSED

@QiXuanWang The "causal" is simply achieved by shifting the output of ordinary conv1d by (k-1)/2, where k is the kernel size.

from tcn.

@huanglianghua Thanks for the comment. Where does (k-1)/2 come from? And why is 'dilation' missed in it? If it's done through chomp1d, then shouldn't shifting be "(k-1)*dilation"?

from tcn.

The Chomp1D class is to ensure causality. For more details, see #8.

from tcn.

@jerrybai1995 Thanks. That's confirmed my guess. I saw pytorch used C(i-1) for convolution of C(i) which somehow makes me think it's causal but indeed it should be Chomp1D.

from tcn.