lt-tensor 0.0.1a35__py3-none-any.whl → 0.0.1a37__py3-none-any.whl

lt_tensor/model_zoo/losses/discriminators.py

@@ -726,3 +726,235 @@ class MultiResolutionDiscriminator(_MultiDiscriminatorT):
             y_d_gs.append(y_d_g)
             fmap_gs.append(fmap_g)
         return y_d_rs, y_d_gs, fmap_rs, fmap_gs
+
+
+class DiscriminatorCQT(ConvNets):
+    """Adapted from https://github.com/open-mmlab/Amphion/blob/main/models/vocoders/gan/discriminator/mssbcqtd.py under the MIT license."""
+
+    def __init__(
+        self,
+        hop_length: int,
+        n_octaves: int,
+        bins_per_octave: int,
+        sampling_rate: int,
+        cqtd_filters: int = 128,
+        cqtd_max_filters: int = 1024,
+        cqtd_filters_scale: int = 1,
+        cqtd_dilations: list = [1, 2, 4],
+        cqtd_in_channels: int = 1,
+        cqtd_out_channels: int = 1,
+        cqtd_normalize_volume: bool = False,
+    ):
+        super().__init__()
+        self.filters = cqtd_filters
+        self.max_filters = cqtd_max_filters
+        self.filters_scale = cqtd_filters_scale
+        self.kernel_size = (3, 9)
+        self.dilations = cqtd_dilations
+        self.stride = (1, 2)
+
+        self.fs = sampling_rate
+        self.in_channels = cqtd_in_channels
+        self.out_channels = cqtd_out_channels
+        self.hop_length = hop_length
+        self.n_octaves = n_octaves
+        self.bins_per_octave = bins_per_octave
+
+        # Lazy-load
+        from lt_tensor.model_zoo.losses.CQT.transforms import CQT2010v2
+
+        self.cqt_transform = CQT2010v2(
+            sr=self.fs * 2,
+            hop_length=self.hop_length,
+            n_bins=self.bins_per_octave * self.n_octaves,
+            bins_per_octave=self.bins_per_octave,
+            output_format="Complex",
+            pad_mode="constant",
+        )
+
+        self.conv_pres = nn.ModuleList()
+        for _ in range(self.n_octaves):
+            self.conv_pres.append(
+                nn.Conv2d(
+                    self.in_channels * 2,
+                    self.in_channels * 2,
+                    kernel_size=self.kernel_size,
+                    padding=self.get_2d_padding(self.kernel_size),
+                )
+            )
+
+        self.convs = nn.ModuleList()
+
+        self.convs.append(
+            nn.Conv2d(
+                self.in_channels * 2,
+                self.filters,
+                kernel_size=self.kernel_size,
+                padding=self.get_2d_padding(self.kernel_size),
+            )
+        )
+
+        in_chs = min(self.filters_scale * self.filters, self.max_filters)
+        for i, dilation in enumerate(self.dilations):
+            out_chs = min(
+                (self.filters_scale ** (i + 1)) * self.filters, self.max_filters
+            )
+            self.convs.append(
+                weight_norm(
+                    nn.Conv2d(
+                        in_chs,
+                        out_chs,
+                        kernel_size=self.kernel_size,
+                        stride=self.stride,
+                        dilation=(dilation, 1),
+                        padding=self.get_2d_padding(self.kernel_size, (dilation, 1)),
+                    )
+                )
+            )
+            in_chs = out_chs
+        out_chs = min(
+            (self.filters_scale ** (len(self.dilations) + 1)) * self.filters,
+            self.max_filters,
+        )
+        self.convs.append(
+            weight_norm(
+                nn.Conv2d(
+                    in_chs,
+                    out_chs,
+                    kernel_size=(self.kernel_size[0], self.kernel_size[0]),
+                    padding=self.get_2d_padding(
+                        (self.kernel_size[0], self.kernel_size[0])
+                    ),
+                )
+            )
+        )
+
+        self.conv_post = weight_norm(
+            nn.Conv2d(
+                out_chs,
+                self.out_channels,
+                kernel_size=(self.kernel_size[0], self.kernel_size[0]),
+                padding=self.get_2d_padding((self.kernel_size[0], self.kernel_size[0])),
+            )
+        )
+
+        self.activation = torch.nn.LeakyReLU(negative_slope=0.1)
+        self.resample = T.Resample(orig_freq=self.fs, new_freq=self.fs * 2)
+
+        self.cqtd_normalize_volume = cqtd_normalize_volume
+        if self.cqtd_normalize_volume:
+            print(
+                f"[INFO] cqtd_normalize_volume set to True. Will apply DC offset removal & peak volume normalization in CQTD!"
+            )
+
+    def get_2d_padding(
+        self,
+        kernel_size: Tuple[int, int],
+        dilation: Tuple[int, int] = (1, 1),
+    ):
+        return (
+            ((kernel_size[0] - 1) * dilation[0]) // 2,
+            ((kernel_size[1] - 1) * dilation[1]) // 2,
+        )
+
+    def forward(self, x: torch.tensor) -> Tuple[torch.Tensor, List[torch.Tensor]]:
+        fmap = []
+
+        if self.cqtd_normalize_volume:
+            # Remove DC offset
+            x = x - x.mean(dim=-1, keepdims=True)
+            # Peak normalize the volume of input audio
+            x = 0.8 * x / (x.abs().max(dim=-1, keepdim=True)[0] + 1e-9)
+
+        x = self.resample(x)
+
+        z = self.cqt_transform(x)
+
+        z_amplitude = z[:, :, :, 0].unsqueeze(1)
+        z_phase = z[:, :, :, 1].unsqueeze(1)
+
+        z = torch.cat([z_amplitude, z_phase], dim=1)
+        z = torch.permute(z, (0, 1, 3, 2))  # [B, C, W, T] -> [B, C, T, W]
+
+        latent_z = []
+        for i in range(self.n_octaves):
+            latent_z.append(
+                self.conv_pres[i](
+                    z[
+                        :,
+                        :,
+                        :,
+                        i * self.bins_per_octave : (i + 1) * self.bins_per_octave,
+                    ]
+                )
+            )
+        latent_z = torch.cat(latent_z, dim=-1)
+
+        for i, l in enumerate(self.convs):
+            latent_z = l(latent_z)
+
+            latent_z = self.activation(latent_z)
+            fmap.append(latent_z)
+
+        latent_z = self.conv_post(latent_z)
+
+        return latent_z, fmap
+
+
+class MultiScaleSubbandCQTDiscriminator(_MultiDiscriminatorT):
+    def __init__(
+        self,
+        sampling_rate: int,
+        cqtd_filters: int = 128,
+        cqtd_max_filters: int = 1024,
+        cqtd_filters_scale: Number = 1,
+        cqtd_dilations: list = [1, 2, 4],
+        cqtd_hop_lengths: list = [512, 256, 256],
+        cqtd_n_octaves: list = [9, 9, 9],
+        cqtd_bins_per_octaves: list = [24, 36, 48],
+        cqtd_in_channels: int = 1,
+        cqtd_out_channels: int = 1,
+        cqtd_normalize_volume: bool = False,
+    ):
+        super().__init__()
+
+        self.discriminators = nn.ModuleList(
+            [
+                DiscriminatorCQT(
+                    hop_length=cqtd_hop_lengths[i],
+                    n_octaves=cqtd_n_octaves[i],
+                    bins_per_octave=cqtd_bins_per_octaves[i],
+                    sampling_rate=sampling_rate,
+                    cqtd_filters=cqtd_filters,
+                    cqtd_max_filters=cqtd_max_filters,
+                    cqtd_filters_scale=cqtd_filters_scale,
+                    cqtd_dilations=cqtd_dilations,
+                    cqtd_in_channels=cqtd_in_channels,
+                    cqtd_out_channels=cqtd_out_channels,
+                    cqtd_normalize_volume=cqtd_normalize_volume,
+                )
+                for i in range(len(cqtd_hop_lengths))
+            ]
+        )
+
+    def forward(self, y: torch.Tensor, y_hat: torch.Tensor) -> Tuple[
+        List[torch.Tensor],
+        List[torch.Tensor],
+        List[List[torch.Tensor]],
+        List[List[torch.Tensor]],
+    ]:
+
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+
+        for disc in self.discriminators:
+            y_d_r, fmap_r = disc(y)
+            y_d_g, fmap_g = disc(y_hat)
+            y_d_rs.append(y_d_r)
+            fmap_rs.append(fmap_r)
+            y_d_gs.append(y_d_g)
+            fmap_gs.append(fmap_g)
+
+        return y_d_rs, y_d_gs, fmap_rs, fmap_gs