lt-tensor 0.0.1a37__tar.gz → 0.0.1a38__tar.gz

{lt_tensor-0.0.1a37 → lt_tensor-0.0.1a38}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lt-tensor
-Version: 0.0.1a37
+Version: 0.0.1a38
 Summary: General utilities for PyTorch and others. Built for general use.
 Home-page: https://github.com/gr1336/lt-tensor/
 Author: gr1336

{lt_tensor-0.0.1a37 → lt_tensor-0.0.1a38}/lt_tensor/__init__.py

@@ -1,4 +1,4 @@
-__version__ = "0.0.1a37"
+__version__ = "0.0.1a38"
 
 from . import (
     lr_schedulers,

{lt_tensor-0.0.1a37 → lt_tensor-0.0.1a38}/lt_tensor/losses.py

@@ -133,7 +133,7 @@ class MultiMelScaleLoss(Model):
         loss_mel_fn: Callable[[Tensor, Tensor], Tensor] = nn.L1Loss(),
         loss_pitch_fn: Callable[[Tensor, Tensor], Tensor] = nn.L1Loss(),
         loss_rms_fn: Callable[[Tensor, Tensor], Tensor] = nn.L1Loss(),
-        center: bool = True,
+        center: bool = False,
         power: float = 1.0,
         normalized: bool = False,
         pad_mode: str = "reflect",
@@ -149,6 +149,7 @@ class MultiMelScaleLoss(Model):
         lambda_rms: float = 1.0,
         lambda_pitch: float = 1.0,
         weight: float = 1.0,
+        mel: Literal["librosa", "torch"] = "torch",
     ):
         super().__init__()
         assert (
@@ -188,6 +189,7 @@ class MultiMelScaleLoss(Model):
             onesided,
             std,
             mean,
+            mel,
         )
 
     def _setup_mels(
@@ -206,6 +208,7 @@ class MultiMelScaleLoss(Model):
         onesided: Optional[bool],
         std: int,
         mean: int,
+        mel: str,
     ):
         assert (
             len(n_mels)
@@ -224,6 +227,7 @@ class MultiMelScaleLoss(Model):
             pad_mode=pad_mode,
             std=std,
             mean=mean,
+            mel_default=mel,
         )
         self.mel_spectrograms: List[AudioProcessor] = nn.ModuleList(
             [
@@ -247,12 +251,14 @@ class MultiMelScaleLoss(Model):
     def forward(
         self, input_wave: torch.Tensor, target_wave: torch.Tensor
     ) -> torch.Tensor:
-        assert self.use_istft_norm or input_wave.shape[-1] == target_wave.shape[-1]
+        assert self.use_istft_norm or input_wave.shape[-1] == target_wave.shape[-1], (
+            f"Size mismatch! input_wave {input_wave.shape[-1]} must match target_wave: {target_wave.shape[-1]}. "
+            "Alternatively 'use_istft_norm' can be set to Trie with will automatically force the audio to that size."
+        )
         target_wave = target_wave.to(input_wave.device)
         losses = 0.0
         for M in self.mel_spectrograms:
-            # Apply normalization if requested
-            if self.use_istft_norm:
+            if self.use_istft_norm and input_proc.shape[-1] != target_proc.shape[-1]:
                 input_proc = M.istft_norm(input_wave, length=target_wave.shape[-1])
                 target_proc = M.istft_norm(target_wave, length=target_wave.shape[-1])
             else:

{lt_tensor-0.0.1a37 → lt_tensor-0.0.1a38}/lt_tensor/model_zoo/audio_models/diffwave/__init__.py

@@ -1,14 +1,15 @@
-__all__ = ["DiffWave", "DiffWaveConfig", "SpectrogramUpsample", "DiffusionEmbedding"]
+__all__ = ["DiffWave", "DiffWaveConfig", "SpectrogramUpsampler", "DiffusionEmbedding"]
 
 import numpy as np
 from lt_tensor.torch_commons import *
 from torch.nn import functional as F
 from lt_tensor.config_templates import ModelConfig
 from lt_tensor.torch_commons import *
-from lt_tensor.model_zoo.convs import ConvNets, Conv1dEXT
+from lt_tensor.model_zoo.convs import ConvNets, ConvEXT
 from lt_tensor.model_base import Model
 from math import sqrt
 from lt_utils.common import *
+from lt_tensor.misc_utils import log_tensor
 
 
 class DiffWaveConfig(ModelConfig):
@@ -21,12 +22,8 @@ class DiffWaveConfig(ModelConfig):
     unconditional = False
     apply_norm: Optional[Literal["weight", "spectral"]] = None
     apply_norm_resblock: Optional[Literal["weight", "spectral"]] = None
-    noise_schedule: list[int] = np.linspace(1e-4, 0.05, 50).tolist()
+    noise_schedule: list[int] = np.linspace(1e-4, 0.05, 25).tolist()
     # settings for auto-fixes
-    interpolate = False
-    interpolation_mode: Literal[
-        "nearest", "linear", "bilinear", "bicubic", "trilinear", "area", "nearest-exact"
-    ] = "nearest"
 
     def __init__(
         self,
@@ -37,16 +34,6 @@ class DiffWaveConfig(ModelConfig):
         dilation_cycle_length=10,
         unconditional=False,
         noise_schedule: list[int] = np.linspace(1e-4, 0.05, 50).tolist(),
-        interpolate_cond=False,
-        interpolation_mode: Literal[
-            "nearest",
-            "linear",
-            "bilinear",
-            "bicubic",
-            "trilinear",
-            "area",
-            "nearest-exact",
-        ] = "nearest",
         apply_norm: Optional[Literal["weight", "spectral"]] = None,
         apply_norm_resblock: Optional[Literal["weight", "spectral"]] = None,
     ):
@@ -58,8 +45,6 @@ class DiffWaveConfig(ModelConfig):
             "residual_channels": residual_channels,
             "unconditional": unconditional,
             "noise_schedule": noise_schedule,
-            "interpolate": interpolate_cond,
-            "interpolation_mode": interpolation_mode,
             "apply_norm": apply_norm,
             "apply_norm_resblock": apply_norm_resblock,
         }
@@ -102,19 +87,34 @@ class DiffusionEmbedding(Model):
         return table
 
 
-class SpectrogramUpsample(Model):
+class SpectrogramUpsampler(Model):
     def __init__(self):
         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])
-        self.activation = nn.LeakyReLU(0.4)
+        self.conv_net = nn.Sequential(
+            ConvEXT(
+                1,
+                1,
+                [3, 32],
+                stride=[1, 16],
+                padding=[1, 8],
+                module_type="2d",
+                transpose=True,
+            ),
+            nn.LeakyReLU(0.1),
+            ConvEXT(
+                1,
+                1,
+                [3, 32],
+                stride=[1, 16],
+                padding=[1, 8],
+                module_type="2d",
+                transpose=True,
+            ),
+            nn.LeakyReLU(0.1),
+        )
 
-    def forward(self, x):
-        x = torch.unsqueeze(x, 1)
-        x = self.activation(self.conv1(x))
-        x = self.activation(self.conv2(x))
-        x = torch.squeeze(x, 1)
-        return x
+    def forward(self, x: Tensor):
+        return self.conv_net(x.unsqueeze(0)).squeeze(1)
 
 
 class ResidualBlock(Model):
@@ -133,7 +133,7 @@ class ResidualBlock(Model):
         :param uncond: disable spectrogram conditional
         """
         super().__init__()
-        self.dilated_conv = Conv1dEXT(
+        self.dilated_conv = ConvEXT(
             residual_channels,
             2 * residual_channels,
             3,
@@ -142,18 +142,18 @@ class ResidualBlock(Model):
             apply_norm=apply_norm,
         )
         self.diffusion_projection = nn.Linear(512, residual_channels)
-        if not uncond:  # conditional model
-            self.conditioner_projection = Conv1dEXT(
+        self.uncoditional = uncond
+        self.conditioner_projection = None
+        if not uncond:
+            self.conditioner_projection = ConvEXT(
                 n_mels,
                 2 * residual_channels,
                 1,
                 apply_norm=apply_norm,
             )
-        else:  # unconditional model
-            self.conditioner_projection = None
 
-        self.output_projection = Conv1dEXT(
-            residual_channels, 2 * residual_channels, 1, apply_norm == apply_norm
+        self.output_projection = ConvEXT(
+            residual_channels, 2 * residual_channels, 1, apply_norm=apply_norm
         )
 
     def forward(
@@ -164,20 +164,15 @@ class ResidualBlock(Model):
     ):
 
         diffusion_step = self.diffusion_projection(diffusion_step).unsqueeze(-1)
-        y = x + diffusion_step
-        if (
-            conditioner is None or 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 = (x + diffusion_step).squeeze(1)
+        y = self.dilated_conv(y)
+        if not self.uncoditional and conditioner is not None:
+            y = y + self.conditioner_projection(conditioner)
 
+        gate, _filter = y.chunk(2, dim=1)
+        y = gate.sigmoid() * _filter.tanh()
         y = self.output_projection(y)
-        residual, skip = torch.chunk(y, 2, dim=1)
+        residual, skip = y.chunk(2, dim=1)
         return (x + residual) / sqrt(2.0), skip
 
 
@@ -186,19 +181,17 @@ class DiffWave(Model):
         super().__init__()
         self.params = params
         self.n_hop = self.params.hop_samples
-        self.interpolate = self.params.interpolate
-        self.interpolate_mode = self.params.interpolation_mode
-        self.input_projection = Conv1dEXT(
+        self.input_projection = ConvEXT(
             in_channels=1,
             out_channels=params.residual_channels,
             kernel_size=1,
             apply_norm=self.params.apply_norm,
+            activation_out=nn.LeakyReLU(0.1),
         )
         self.diffusion_embedding = DiffusionEmbedding(len(params.noise_schedule))
-        if self.params.unconditional:  # use unconditional model
-            self.spectrogram_upsample = None
-        else:
-            self.spectrogram_upsample = SpectrogramUpsample()
+        self.spectrogram_upsampler = (
+            SpectrogramUpsampler() if not self.params.unconditional else None
+        )
 
         self.residual_layers = nn.ModuleList(
             [
@@ -212,18 +205,18 @@ class DiffWave(Model):
                 for i in range(params.residual_layers)
             ]
         )
-        self.skip_projection = Conv1dEXT(
+        self.skip_projection = ConvEXT(
             in_channels=params.residual_channels,
             out_channels=params.residual_channels,
             kernel_size=1,
             apply_norm=self.params.apply_norm,
+            activation_out=nn.LeakyReLU(0.1),
         )
-        self.output_projection = Conv1dEXT(
-            params.residual_channels, 1, 1, apply_norm=self.params.apply_norm
+        self.output_projection = ConvEXT(
+            params.residual_channels, 1, 1, apply_norm=self.params.apply_norm, init_weights=True,
         )
         self.activation = nn.LeakyReLU(0.1)
-        self.r_sqrt = sqrt(len(self.residual_layers))
-        nn.init.zeros_(self.output_projection.weight)
+        self._res_d = sqrt(len(self.residual_layers))
 
     def forward(
         self,
@@ -231,31 +224,25 @@ class DiffWave(Model):
         diffusion_step: Tensor,
         spectrogram: Optional[Tensor] = None,
     ):
-        T = x.shape[-1]
-        if x.ndim == 2:
-            x = audio.unsqueeze(1)
-        x = self.activation(self.input_projection(x))
+        if not self.params.unconditional:
+            assert spectrogram is not None
+        if audio.ndim < 3:
+            if audio.ndim == 2:
+                audio = audio.unsqueeze(1)
+            else:
+                audio = audio.unsqueeze(0).unsqueeze(0)
 
+        x = self.input_projection(audio)
         diffusion_step = self.diffusion_embedding(diffusion_step)
-        if spectrogram is not None and self.spectrogram_upsample is not None:
-            if self.auto_interpolate:
-                # a little heavy, but helps a lot to fix mismatched shapes,
-                # not always recommended due to data loss
-                spectrogram = F.interpolate(
-                    input=spectrogram,
-                    size=int(T * self.n_hop),
-                    mode=self.interpolate_mode,
-                )
-            spectrogram = self.spectrogram_upsample(spectrogram)
+        if not self.params.unconditional:  # use conditional model
+            spectrogram = self.spectrogram_upsampler(spectrogram)
 
-        skip = None
+        skip = torch.zeros_like(x, device=x.device)
         for i, layer in enumerate(self.residual_layers):
             x, skip_connection = layer(x, diffusion_step, spectrogram)
-            if i == 0:
-                skip = skip_connection
-            else:
-                skip = skip_connection + skip
-        x = skip / self.r_sqrt
-        x = self.activation(self.skip_projection(x))
+            skip += skip_connection
+
+        x = skip / self._res_d
+        x = self.skip_projection(x)
         x = self.output_projection(x)
         return x

{lt_tensor-0.0.1a37 → lt_tensor-0.0.1a38}/lt_tensor/model_zoo/convs.py

@@ -1,4 +1,4 @@
-__all__ = ["ConvNets", "Conv1dEXT"]
+__all__ = ["ConvNets", "ConvEXT"]
 import math
 from lt_utils.common import *
 import torch.nn.functional as F
@@ -6,6 +6,7 @@ from lt_tensor.torch_commons import *
 from lt_tensor.model_base import Model
 from lt_tensor.misc_utils import log_tensor
 from lt_tensor.model_zoo.fusion import AdaFusion1D, AdaIN1D
+from lt_utils.misc_utils import default
 
 
 def spectral_norm_select(module: nn.Module, enabled: bool):
@@ -52,10 +53,7 @@ class ConvNets(Model):
             m.weight.data.normal_(mean, std)
 
 
-class Conv1dEXT(ConvNets):
-
-    # TODO: Use this module to replace all that are using normalizations, mostly those in `audio_models`
-
+class ConvEXT(ConvNets):
     def __init__(
         self,
         in_channels: int,
@@ -72,6 +70,10 @@ class Conv1dEXT(ConvNets):
         apply_norm: Optional[Literal["weight", "spectral"]] = None,
         activation_in: nn.Module = nn.Identity(),
         activation_out: nn.Module = nn.Identity(),
+        module_type: Literal["1d", "2d", "3d"] = "1d",
+        transpose: bool = False,
+        weight_init: Optional[Callable[[nn.Module], None]] = None,
+        init_weights: bool = True,
         *args,
         **kwargs,
     ):
@@ -91,23 +93,30 @@ class Conv1dEXT(ConvNets):
             device=device,
             dtype=dtype,
         )
+        match module_type.lower():
+            case "1d":
+                md = nn.Conv1d if not transpose else nn.ConvTranspose1d
+            case "2d":
+                md = nn.Conv2d if not transpose else nn.ConvTranspose2d
+            case "3d":
+                md = nn.Conv3d if not transpose else nn.ConvTranspose3d
+            case _:
+                raise ValueError(
+                    f"module_type {module_type} is not a valid module type! use '1d', '2d' or '3d'"
+                )
+
         if apply_norm is None:
-            self.cnn = nn.Conv1d(**cnn_kwargs)
-            self.has_wn = False
+            self.cnn = md(**cnn_kwargs)
         else:
-            self.has_wn = True
             if apply_norm == "spectral":
-                self.cnn = spectral_norm(nn.Conv1d(**cnn_kwargs))
+                self.cnn = spectral_norm(md(**cnn_kwargs))
             else:
-                self.cnn = weight_norm(nn.Conv1d(**cnn_kwargs))
+                self.cnn = weight_norm(md(**cnn_kwargs))
         self.actv_in = activation_in
         self.actv_out = activation_out
-        self.cnn.apply(self.init_weights)
+        if init_weights:
+            weight_init = default(weight_init, self.init_weights)
+            self.cnn.apply(weight_init)
 
     def forward(self, input: Tensor):
         return self.actv_out(self.cnn(self.actv_in(input)))
-
-    def remove_norms(self, name="weight"):
-        if self.has_wn:
-            remove_norm(self.cnn, name)
-            self.has_wn = False

{lt_tensor-0.0.1a37 → lt_tensor-0.0.1a38}/lt_tensor/processors/audio.py

@@ -202,6 +202,8 @@ class AudioProcessor(Model):
         *args,
         **kwargs,
     ):
+        if wave.ndim == 1:
+            wave = wave.unsqueeze(0)
         wave = torch.nn.functional.pad(
             wave.unsqueeze(1),
             (self.mel_lib_padding, self.mel_lib_padding),

{lt_tensor-0.0.1a37 → lt_tensor-0.0.1a38}/lt_tensor.egg-info/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lt-tensor
-Version: 0.0.1a37
+Version: 0.0.1a38
 Summary: General utilities for PyTorch and others. Built for general use.
 Home-page: https://github.com/gr1336/lt-tensor/
 Author: gr1336

{lt_tensor-0.0.1a37 → lt_tensor-0.0.1a38}/setup.py

@@ -4,7 +4,7 @@ with open("README.md", "r", encoding="utf-8") as f:
     long_description = f.read()
 
 setup(
-    version="0.0.1a37",
+    version="0.0.1a38",
     name="lt-tensor",
     description="General utilities for PyTorch and others. Built for general use.",
     long_description=long_description,