TSTNN
TSTNN copied to clipboard
key2miao
Time Delay of TSTNN
Hi, I have already read your paper and it is definitely an innovative paper. And I've also been working on the Demucs modelu recently and you also compared TSTNN with Demucs in your experiments. The Facebook AI lab has descripted taht the casual Demucs with a lookahead time of 37 ms and it can run real time on the specified cpu in their paper. So I have a question about your TSTNN, have you compute or estimate the time delay of the model or does it sastify the real time running? Hopefully to received your reply and we can communicate with each other!
Hi, I have already read your paper and it is definitely an innovative paper. And I've also been working on the Demucs modelu recently and you also compared TSTNN with Demucs in your experiments. The Facebook AI lab has descripted taht the casual Demucs with a lookahead time of 37 ms and it can run real time on the specified cpu in their paper. So I have a question about your TSTNN, have you compute or estimate the time delay of the model or does it sastify the real time running? Hopefully to received your reply and we can communicate with each other!
I think the attention mechanism of the model in this article utilizes the information of the whole sequence and should be non-real-time.I am also interested in the real-time transformer structure and am curious if it will affect the excellent performance of the model.
Thanks for your reply. I am looking forward to your work on the real-time Transformer structure.
Hi, sorry for late reply. If you want to test the real-time or casual performance of TSTNN, please try to do following modifications:
- Using masked attention in transformer for mask future information.
- In local transformer, keep non-casual setting in GRU (still maintain Bi-GRU). But in global transformer, replacing Bi-GRU by undirectional GRU and accordingly using casual normalization.
