why still (3 7 11) about repgan HOT 7 CLOSED

in training, we also have three resnetblock 3, 7, 11, but in inference, we will fuse three kernel size into one, 11, u can check the inference function in rep_conv

from repgan.

in training, we also have three resnetblock 3, 7, 11, but in inference, we will fuse three kernel size into one, 11, u can check the inference function in rep_conv

ok, thanks, I will check it.

from repgan.

@azraelkuan do it this way, whether training and inference exist gap?

from repgan.

no quality gap, the weight is equal in training and inference

from repgan.

@azraelkuan thanks, another question, you generator does not have skip connetction(residual connect). This is different from hifigan, without it, does quailty decay?

def forward(self, x):
        x = self.input_conv(x)

        for i in range(len(self.upsamples)):
            x = self.upsamples[i](x)
            if self.training:
                x = self.blocks[i](x)
            else:
                x = self.blocks[i].inference(x)
            x = self.non_linear(x)
        x = self.output_conv(x)
        return 

and in repconv

def forward(self, x: torch.Tensor) -> torch.Tensor:
        for idx in range(len(self.convs1)):
            x = self.act(x)
            x = self.convs1[idx](x)
            if self.use_additional_convs:
                x = self.act(x)
                x = self.convs2[idx](x)
        return x

and resblock_dilations is different, origin len is 3, but yours is 4

from repgan.

check this:

from repgan.

Thank you very much!

from repgan.