search

JusperLee / TDANet

An efficient speech separation method
https://cslikai.cn/project/TDANet
Apache License 2.0
262 stars 29 forks source link

Might need to change one line in TDANet.py. #39

Open julyanghar opened 5 days ago

julyanghar commented 5 days ago

According to the original paper, the 367th line in TDANet seems to be changed as follow:

# expanded = self.last_layer[i](x_fused[i], x_fused[i - 1])
expanded = self.last_layer[i](x_fused[i], x_fused[i + 1])

Because the first embedding in the decoder should be produced using the top global feature and the upsampled one by factor 2. image

JusperLee commented 5 days ago

Yes, you are right. In the tdanet v2, we have changed this code. https://github.com/JusperLee/TDANet/blob/565af18692e18bf695e5bb0ca54ba466c4a86a2a/look2hear/models/TDANet-v2.py#L377

julyanghar commented 5 days ago

Good to know!