keras-tcn icon indicating copy to clipboard operation
keras-tcn copied to clipboard

Published 20 hours ago • verified publisher icon philipperemy

fully convolution layer

Open HouKun-github opened this issue 4 years ago • 1 comments

Why is this version of the one-dimensional fully convolution layer gone? The previous version had a one-dimensional full convolution, at least the original paper used full convolution to ensure that the sequence is not missing, can you explain?

HouKun-github avatar May 31 '20 09:05 HouKun-github

@HouKun-github hey could it be linked to https://github.com/philipperemy/keras-tcn/pull/133?

philipperemy avatar Jun 03 '20 07:06 philipperemy

Yes, you are right. The 1-D fully convolutional layer is gone and I think it is because we found that removing it boosted the performance (or did not deteriorate the performance of the model, cf. Occam razor). I agree that we don't match the original paper 100% now because of this change.

philipperemy avatar Sep 25 '22 15:09 philipperemy