locuslab/TCN

Is too large of receptive field wasteful?

johnsyin opened this issue · 2 comments

johnsyin commented

In my data, the sequence are all <10k length. The problem now is that I want to stack many layers, this means I will have very large receptive field even though alot of them are just padded values.

I am wondering if I should stack TCN after one TCN (so the receptive field is smaller), instead of using one TCN with a long channel_size (many levels).

jerrybai1995 commented

Hi @johnsyin,

Thank you for your question. Yes, if you don't want a large receptive field, you can dynamically set the dilation factor (so not necessarily 1,2,4,8,16,...; you can do 1,1,1,2,2,2,3,3,3,4,4,4,...). You can also simply stack multiple TCNs as you suggested.

  • 👍2
johnsyin commented

I see, that might be a good way too.