Regarding ECAPA-TDNN implementation

clovaai / voxceleb_trainer

In defence of metric learning for speaker recognition

MIT License

1.03k stars 272 forks source link

Regarding ECAPA-TDNN implementation #86

Closed 009deep closed 3 years ago

009deep commented 3 years ago

Hi @joonson, In you your voxsrc 20 summary paper, you have mentioned you implemented ECAPA-TDNN network with help of Brecht. Could you provide that network implementation details? It's hard to understand from paper.

joonson commented 3 years ago

Hi, here is our implementation of the network. Let me know if it works -- I might merge this into the repository in a future update.

ECAPA.zip

009deep commented 3 years ago

thank you so much.

lawlict commented 3 years ago

Hi, @joonson. I find that the ECAPA_TDNN model runs slow. When I set the bias of Conv1d in Line 46 to False, it works normally:

for i in range(self.nums):
    convs.append(nn.Conv1d(width, width, kernel_size=kernel_size, dilation=dilation, padding=num_pad, bias=False))
    bns.append(nn.BatchNorm1d(width))

Do you meet the same problem? I'm just confused by the phenomenon.

Reptilefury commented 2 years ago

is there an implementation using Tensorflow

Jungjee commented 2 years ago

Hi @Reptilefury, at this moment, we are not considering the support of Tensorflow in this repository.