lt-tensor 0.0.1a14__py3-none-any.whl → 0.0.1a16__py3-none-any.whl

lt_tensor/model_zoo/audio_models/diffwave/model.py

@@ -0,0 +1,201 @@
+__all__ = ["DiffWave", "SpectrogramUpsampler", "DiffusionEmbedding"]
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from math import sqrt
+
+
+class AttrDict(dict):
+    def __init__(self, *args, **kwargs):
+        super(AttrDict, self).__init__(*args, **kwargs)
+        self.__dict__ = self
+
+    def override(self, attrs):
+        if isinstance(attrs, dict):
+            self.__dict__.update(**attrs)
+        elif isinstance(attrs, (list, tuple, set)):
+            for attr in attrs:
+                self.override(attr)
+        elif attrs is not None:
+            raise NotImplementedError
+        return self
+
+
+params = AttrDict(
+    # Training params
+    batch_size=16,
+    learning_rate=2e-4,
+    max_grad_norm=None,
+    # Data params
+    sample_rate=22050,
+    n_mels=80,
+    n_fft=1024,
+    hop_samples=256,
+    crop_mel_frames=62,  # Probably an error in paper.
+    # Model params
+    residual_layers=30,
+    residual_channels=64,
+    dilation_cycle_length=10,
+    unconditional=False,
+    noise_schedule=np.linspace(1e-4, 0.05, 50).tolist(),
+    inference_noise_schedule=[0.0001, 0.001, 0.01, 0.05, 0.2, 0.5],
+    # unconditional sample len
+    audio_len=22050 * 5,  # unconditional_synthesis_samples
+)
+
+
+def Conv1d(*args, **kwargs):
+    layer = nn.Conv1d(*args, **kwargs)
+    nn.init.kaiming_normal_(layer.weight)
+    return layer
+
+
+class DiffusionEmbedding(nn.Module):
+    def __init__(self, max_steps):
+        super().__init__()
+        self.register_buffer(
+            "embedding", self._build_embedding(max_steps), persistent=False
+        )
+        self.projection1 = nn.Linear(128, 512)
+        self.projection2 = nn.Linear(512, 512)
+        self.activation = nn.SiLU()
+
+    def forward(self, diffusion_step):
+        if diffusion_step.dtype in [torch.int32, torch.int64]:
+            x = self.embedding[diffusion_step]
+        else:
+            x = self._lerp_embedding(diffusion_step)
+        x = self.projection1(x)
+        x = self.activation(x)
+        x = self.projection2(x)
+        x = self.activation(x)
+        return x
+
+    def _lerp_embedding(self, t):
+        low_idx = torch.floor(t).long()
+        high_idx = torch.ceil(t).long()
+        low = self.embedding[low_idx]
+        high = self.embedding[high_idx]
+        return low + (high - low) * (t - low_idx)
+
+    def _build_embedding(self, max_steps):
+        steps = torch.arange(max_steps).unsqueeze(1)  # [T,1]
+        dims = torch.arange(64).unsqueeze(0)  # [1,64]
+        table = steps * 10.0 ** (dims * 4.0 / 63.0)  # [T,64]
+        table = torch.cat([torch.sin(table), torch.cos(table)], dim=1)
+        return table
+
+
+class SpectrogramUpsampler(nn.Module):
+    def __init__(self, n_mels):
+        super().__init__()
+        self.conv1 = nn.ConvTranspose2d(1, 1, [3, 32], stride=[1, 16], padding=[1, 8])
+        self.conv2 = nn.ConvTranspose2d(1, 1, [3, 32], stride=[1, 16], padding=[1, 8])
+
+    def forward(self, x):
+        x = torch.unsqueeze(x, 1)
+        x = self.conv1(x)
+        x = F.leaky_relu(x, 0.4)
+        x = self.conv2(x)
+        x = F.leaky_relu(x, 0.4)
+        x = torch.squeeze(x, 1)
+        return x
+
+
+class ResidualBlock(nn.Module):
+    def __init__(self, n_mels, residual_channels, dilation, uncond=False):
+        """
+        :param n_mels: inplanes of conv1x1 for spectrogram conditional
+        :param residual_channels: audio conv
+        :param dilation: audio conv dilation
+        :param uncond: disable spectrogram conditional
+        """
+        super().__init__()
+        self.dilated_conv = Conv1d(
+            residual_channels,
+            2 * residual_channels,
+            3,
+            padding=dilation,
+            dilation=dilation,
+        )
+        self.diffusion_projection = nn.Linear(512, residual_channels)
+        if not uncond:  # conditional model
+            self.conditioner_projection = Conv1d(n_mels, 2 * residual_channels, 1)
+        else:  # unconditional model
+            self.conditioner_projection = None
+
+        self.output_projection = Conv1d(residual_channels, 2 * residual_channels, 1)
+
+    def forward(self, x, diffusion_step, conditioner=None):
+        assert (conditioner is None and self.conditioner_projection is None) or (
+            conditioner is not None and self.conditioner_projection is not None
+        )
+
+        diffusion_step = self.diffusion_projection(diffusion_step).unsqueeze(-1)
+        y = x + diffusion_step
+        if self.conditioner_projection is None:  # using a unconditional model
+            y = self.dilated_conv(y)
+        else:
+            conditioner = self.conditioner_projection(conditioner)
+            y = self.dilated_conv(y) + conditioner
+
+        gate, filter = torch.chunk(y, 2, dim=1)
+        y = torch.sigmoid(gate) * torch.tanh(filter)
+
+        y = self.output_projection(y)
+        residual, skip = torch.chunk(y, 2, dim=1)
+        return (x + residual) / sqrt(2.0), skip
+
+
+class DiffWave(nn.Module):
+    def __init__(self, params):
+        super().__init__()
+        self.params = params
+        self.input_projection = Conv1d(1, params.residual_channels, 1)
+        self.diffusion_embedding = DiffusionEmbedding(len(params.noise_schedule))
+        if self.params.unconditional:  # use unconditional model
+            self.spectrogram_upsampler = None
+        else:
+            self.spectrogram_upsampler = SpectrogramUpsampler(params.n_mels)
+
+        self.residual_layers = nn.ModuleList(
+            [
+                ResidualBlock(
+                    params.n_mels,
+                    params.residual_channels,
+                    2 ** (i % params.dilation_cycle_length),
+                    uncond=params.unconditional,
+                )
+                for i in range(params.residual_layers)
+            ]
+        )
+        self.skip_projection = Conv1d(
+            params.residual_channels, params.residual_channels, 1
+        )
+        self.output_projection = Conv1d(params.residual_channels, 1, 1)
+        nn.init.zeros_(self.output_projection.weight)
+
+    def forward(self, audio, diffusion_step, spectrogram=None):
+        assert (spectrogram is None and self.spectrogram_upsampler is None) or (
+            spectrogram is not None and self.spectrogram_upsampler is not None
+        )
+        x = audio.unsqueeze(1)
+        x = self.input_projection(x)
+        x = F.relu(x)
+
+        diffusion_step = self.diffusion_embedding(diffusion_step)
+        if self.spectrogram_upsampler:  # use conditional model
+            spectrogram = self.spectrogram_upsampler(spectrogram)
+
+        skip = None
+        for layer in self.residual_layers:
+            x, skip_connection = layer(x, diffusion_step, spectrogram)
+            skip = skip_connection if skip is None else skip_connection + skip
+
+        x = skip / sqrt(len(self.residual_layers))
+        x = self.skip_projection(x)
+        x = F.relu(x)
+        x = self.output_projection(x)
+        return x

lt_tensor/model_zoo/audio_models/hifigan/__init__.py

@@ -0,0 +1,393 @@
+__all__ = ["HifiganGenerator"]
+from lt_utils.common import *
+from lt_tensor.torch_commons import *
+from lt_tensor.model_zoo.residual import ConvNets
+from torch.nn import functional as F
+
+import torch
+import torch.nn.functional as F
+import torch.nn as nn
+
+
+def get_padding(kernel_size, dilation=1):
+    return int((kernel_size * dilation - dilation) / 2)
+
+
+class ResBlock1(ConvNets):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3, 5)):
+        super().__init__()
+
+        self.convs1 = nn.ModuleList(
+            [
+                weight_norm(
+                    nn.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    nn.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+                weight_norm(
+                    nn.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[2],
+                        padding=get_padding(kernel_size, dilation[2]),
+                    )
+                ),
+            ]
+        )
+        self.convs1.apply(self.init_weights)
+
+        self.convs2 = nn.ModuleList(
+            [
+                weight_norm(
+                    nn.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    nn.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+                weight_norm(
+                    nn.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=1,
+                        padding=get_padding(kernel_size, 1),
+                    )
+                ),
+            ]
+        )
+        self.convs2.apply(self.init_weights)
+        self.activation = nn.LeakyReLU(0.1)
+
+    def forward(self, x):
+        for c1, c2 in zip(self.convs1, self.convs2):
+            xt = c1(self.activation(x))
+            xt = c2(self.activation(xt))
+            x = xt + x
+        return x
+
+    def remove_weight_norm(self):
+        for l in self.convs1:
+            remove_weight_norm(l)
+        for l in self.convs2:
+            remove_weight_norm(l)
+
+
+class ResBlock2(ConvNets):
+    def __init__(self, channels, kernel_size=3, dilation=(1, 3)):
+        super().__init__()
+        self.convs = nn.ModuleList(
+            [
+                weight_norm(
+                    nn.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[0],
+                        padding=get_padding(kernel_size, dilation[0]),
+                    )
+                ),
+                weight_norm(
+                    nn.Conv1d(
+                        channels,
+                        channels,
+                        kernel_size,
+                        1,
+                        dilation=dilation[1],
+                        padding=get_padding(kernel_size, dilation[1]),
+                    )
+                ),
+            ]
+        )
+        self.convs.apply(self.init_weights)
+        self.activation = nn.LeakyReLU(0.1)
+
+    def forward(self, x):
+        for c in self.convs:
+            xt = c(self.activation(x))
+            x = xt + x
+        return x
+
+
+class HifiganGenerator(ConvNets):
+    def __init__(self, h):
+        super().__init__()
+        self.h = h
+        self.num_kernels = len(h.resblock_kernel_sizes)
+        self.num_upsamples = len(h.upsample_rates)
+        self.conv_pre = weight_norm(
+            nn.Conv1d(80, h.upsample_initial_channel, 7, 1, padding=3)
+        )
+        resblock = ResBlock1 if h.resblock == "1" else ResBlock2
+        self.activation = nn.LeakyReLU(0.1)
+        self.ups = nn.ModuleList()
+        for i, (u, k) in enumerate(zip(h.upsample_rates, h.upsample_kernel_sizes)):
+            self.ups.append(
+                weight_norm(
+                    nn.ConvTranspose1d(
+                        h.upsample_initial_channel // (2**i),
+                        h.upsample_initial_channel // (2 ** (i + 1)),
+                        k,
+                        u,
+                        padding=(k - u) // 2,
+                    )
+                )
+            )
+
+        self.resblocks = nn.ModuleList()
+        for i in range(len(self.ups)):
+            ch = h.upsample_initial_channel // (2 ** (i + 1))
+            for j, (k, d) in enumerate(
+                zip(h.resblock_kernel_sizes, h.resblock_dilation_sizes)
+            ):
+                self.resblocks.append(resblock(h, ch, k, d))
+
+        self.conv_post = weight_norm(nn.Conv1d(ch, 1, 7, 1, padding=3))
+        self.ups.apply(self.init_weights)
+        self.conv_post.apply(self.init_weights)
+
+    def forward(self, x):
+        x = self.conv_pre(x)
+        for i in range(self.num_upsamples):
+            x = self.ups[i](self.activation(x))
+            xs = None
+            for j in range(self.num_kernels):
+                if xs is None:
+                    xs = self.resblocks[i * self.num_kernels + j](x)
+                else:
+                    xs += self.resblocks[i * self.num_kernels + j](x)
+            x = xs / self.num_kernels
+        x = self.conv_post(self.activation(x))
+        x = torch.tanh(x)
+
+        return x
+
+
+class DiscriminatorP(ConvNets):
+    def __init__(self, period, kernel_size=5, stride=3, use_spectral_norm=False):
+        super(DiscriminatorP, self).__init__()
+        self.period = period
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(
+                    nn.Conv2d(
+                        1,
+                        32,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(5, 1), 0),
+                    )
+                ),
+                norm_f(
+                    nn.Conv2d(
+                        32,
+                        128,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(5, 1), 0),
+                    )
+                ),
+                norm_f(
+                    nn.Conv2d(
+                        128,
+                        512,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(5, 1), 0),
+                    )
+                ),
+                norm_f(
+                    nn.Conv2d(
+                        512,
+                        1024,
+                        (kernel_size, 1),
+                        (stride, 1),
+                        padding=(get_padding(5, 1), 0),
+                    )
+                ),
+                norm_f(nn.Conv2d(1024, 1024, (kernel_size, 1), 1, padding=(2, 0))),
+            ]
+        )
+        self.conv_post = norm_f(nn.Conv2d(1024, 1, (3, 1), 1, padding=(1, 0)))
+        self.activation = nn.LeakyReLU(0.1)
+
+    def forward(self, x):
+        fmap = []
+
+        # 1d to 2d
+        b, c, t = x.shape
+        if t % self.period != 0:  # pad first
+            n_pad = self.period - (t % self.period)
+            x = F.pad(x, (0, n_pad), "reflect")
+            t = t + n_pad
+        x = x.view(b, c, t // self.period, self.period)
+
+        for l in self.convs:
+            x = l(x)
+            x = self.activation(x)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class MultiPeriodDiscriminator(ConvNets):
+    def __init__(self):
+        super(MultiPeriodDiscriminator, self).__init__()
+        self.discriminators = nn.ModuleList(
+            [
+                DiscriminatorP(2),
+                DiscriminatorP(3),
+                DiscriminatorP(5),
+                DiscriminatorP(7),
+                DiscriminatorP(11),
+            ]
+        )
+
+    def forward(self, y, y_hat):
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(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
+
+
+class DiscriminatorS(ConvNets):
+    def __init__(self, use_spectral_norm=False):
+        super(DiscriminatorS, self).__init__()
+        norm_f = weight_norm if use_spectral_norm == False else spectral_norm
+        self.convs = nn.ModuleList(
+            [
+                norm_f(nn.Conv1d(1, 128, 15, 1, padding=7)),
+                norm_f(nn.Conv1d(128, 128, 41, 2, groups=4, padding=20)),
+                norm_f(nn.Conv1d(128, 256, 41, 2, groups=16, padding=20)),
+                norm_f(nn.Conv1d(256, 512, 41, 4, groups=16, padding=20)),
+                norm_f(nn.Conv1d(512, 1024, 41, 4, groups=16, padding=20)),
+                norm_f(nn.Conv1d(1024, 1024, 41, 1, groups=16, padding=20)),
+                norm_f(nn.Conv1d(1024, 1024, 5, 1, padding=2)),
+            ]
+        )
+        self.conv_post = norm_f(nn.Conv1d(1024, 1, 3, 1, padding=1))
+        self.activation = nn.LeakyReLU(0.1)
+
+    def forward(self, x):
+        fmap = []
+        for l in self.convs:
+            x = l(x)
+            x = self.activation(x)
+            fmap.append(x)
+        x = self.conv_post(x)
+        fmap.append(x)
+        x = torch.flatten(x, 1, -1)
+
+        return x, fmap
+
+
+class MultiScaleDiscriminator(ConvNets):
+    def __init__(self):
+        super(MultiScaleDiscriminator, self).__init__()
+        self.discriminators = nn.ModuleList(
+            [
+                DiscriminatorS(use_spectral_norm=True),
+                DiscriminatorS(),
+                DiscriminatorS(),
+            ]
+        )
+        self.meanpools = nn.ModuleList(
+            [nn.AvgPool1d(4, 2, padding=2), nn.AvgPool1d(4, 2, padding=2)]
+        )
+
+    def forward(self, y, y_hat):
+        y_d_rs = []
+        y_d_gs = []
+        fmap_rs = []
+        fmap_gs = []
+        for i, d in enumerate(self.discriminators):
+            if i != 0:
+                y = self.meanpools[i - 1](y)
+                y_hat = self.meanpools[i - 1](y_hat)
+            y_d_r, fmap_r = d(y)
+            y_d_g, fmap_g = d(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
+
+
+def feature_loss(fmap_r, fmap_g):
+    loss = 0
+    for dr, dg in zip(fmap_r, fmap_g):
+        for rl, gl in zip(dr, dg):
+            loss += torch.mean(torch.abs(rl - gl))
+
+    return loss * 2
+
+
+def discriminator_loss(disc_real_outputs, disc_generated_outputs):
+    loss = 0
+    r_losses = []
+    g_losses = []
+    for dr, dg in zip(disc_real_outputs, disc_generated_outputs):
+        r_loss = torch.mean((1 - dr) ** 2)
+        g_loss = torch.mean(dg**2)
+        loss += r_loss + g_loss
+        r_losses.append(r_loss.item())
+        g_losses.append(g_loss.item())
+
+    return loss, r_losses, g_losses
+
+
+def generator_loss(disc_outputs):
+    loss = 0
+    gen_losses = []
+    for dg in disc_outputs:
+        l = torch.mean((1 - dg) ** 2)
+        gen_losses.append(l)
+        loss += l
+
+    return loss, gen_losses