BiRNN in the spectral filtering layer?

Question

BiRNN in the spectral filtering layer?

Closed this issue 3 years ago · 4 comments

Thank you for publishing the repository. I noticed that you commented the RNN layer in the bijectors of spectral flow. Did they lead to poor performance? Or any other reason to do so? Thank you!

Answer 1 · 2022-01-17T04:13:35.000Z

Hi ZhuangweiKang, No, they will generally perform better at the expense of compute time since we cascade multiple flows each with its RNN layer. I recommend uncommenting the RNN layer if you are running the model on GPU.

Answer 2 · 2022-01-17T04:42:17.000Z

Thanks for your quick response. Also, does BiRNN and RNN have different performance? And does the different masking rule influence performance (such as [0,1,0,1,...] and [0,0,0,...,1,1,1])?

Answer 3 · 2022-01-17T04:49:22.000Z

I did not find them to have diff performance on simple 1D time series. The masks are introduced to handle variable length sequences and are not arbitrary, they have to take the form [1, 1, 1, 1, 0, 0, ..., 0].

Answer 4 · 2022-01-17T04:50:35.000Z

I see. Thank you very much.