fix(discriminators): CQT forward omits channel squeeze (+2 more)

[andrewwhitecdw]

This PR addresses 3 issues in discriminators.py.

Fix 1

fix: CQT forward omits channel squeeze before nnAudio CQT transform

Fix: Replace:

        x = self.resample(x)

        z = self.cqt_transform(x)

with:

        x = x.squeeze(1)
        x = self.resample(x)

        z = self.cqt_transform(x)

Fix 2

fix: CQT discriminator output not flattened and final fmap omitted

Fix: Replace:

        latent_z = self.conv_post(latent_z)

        return latent_z, fmap

with:

        latent_z = self.conv_post(latent_z)
        fmap.append(latent_z)
        latent_z = torch.flatten(latent_z, 1, -1)

        return latent_z, fmap

Fix 3

fix: DiscriminatorP period typed as List[int] but used as scalar

Fix: Replace:

        h: AttrDict,
        period: List[int],
        kernel_size: int = 5,

with:

        h: AttrDict,
        period: int,
        kernel_size: int = 5,

Files changed

• discriminators.py