Hi, here is the speaker classifier. It might be slightly different from the one we used in this ariticle due to some code refactoring, but the result should be similar.

[dslllu]

Hi, here is the speaker classifier. It might be slightly different from the one we used in this ariticle due to some code refactoring, but the result should be similar. c_in: The dimension of the content/speaker embedding c_h: We set it to 128. c_out: The number of speakers, which is 80 in this paper.

import torch.nn as nn
from einops import rearrange

class Classifier(nn.Module):
    def __init__(self, c_in, c_h, c_out):
        super(Classifier, self).__init__()
        self.in_layer = nn.Linear(c_in, c_h)
        self.conv_relu_block = nn.Sequential(
            nn.Conv1d(c_h, c_h, 3),
            nn.ReLU(),
            nn.Conv1d(c_h, c_h, 3),
            nn.ReLU(),
            nn.Conv1d(c_h, c_h, 3),
            nn.ReLU(),
        )
        self.out_layer = nn.Linear(c_h, c_out)

    def forward(self, x):
        """
        x: (n, c, t)
        """
        x = rearrange(x, 'n c t -> n t c')
        y = self.in_layer(x)
        y = rearrange(y, 'n t c -> n c t')
        y = self.conv_relu_block(y)
        y = y.mean(-1)
        y = self.out_layer(y)
        return y

Originally posted by @KimythAnly in https://github.com/KimythAnly/AGAIN-VC/issues/10#issuecomment-854148318

[dslllu]

Hi，Can you provide more details on training the speaker classifier? such as how to set labels