feat: option to norm inputs with mu-law

Author: flavioschneider

audio_diffusion_pytorch/modules.py

@@ -8,7 +8,7 @@
 from einops_exts import rearrange_many
 from torch import Tensor, einsum
 
-from .utils import default, exists, prod
+from .utils import default, exists, prod, wave_norm, wave_unnorm
 
 """
 Utils
@@ -809,6 +809,7 @@ def __init__(
         use_nearest_upsample: bool,
         use_skip_scale: bool,
         use_context_time: bool,
+        norm: float = 0.0,
         out_channels: Optional[int] = None,
         context_features: Optional[int] = None,
         context_channels: Optional[Sequence[int]] = None,
@@ -822,6 +823,8 @@ def __init__(
         use_context_channels = len(context_channels) > 0
         context_mapping_features = None
 
+        self.norm = norm
+        self.use_norm = norm > 0.0
         self.num_layers = num_layers
         self.use_context_time = use_context_time
         self.use_context_features = use_context_features
@@ -997,9 +1000,11 @@ def forward(
         # Concat context channels at layer 0 if provided
         channels = self.get_channels(channels_list, layer=0)
         x = torch.cat([x, channels], dim=1) if exists(channels) else x
-
         mapping = self.get_mapping(time, features)
 
+        if self.use_norm:
+            x = wave_norm(x, peak=self.norm)
+
         x = self.to_in(x, mapping)
         skips_list = [x]
 
@@ -1019,6 +1024,9 @@ def forward(
         x += skips_list.pop()
         x = self.to_out(x, mapping)
 
+        if self.use_norm:
+            x = wave_unnorm(x, peak=self.norm)
+
         return x
 
 
@@ -1120,11 +1128,14 @@ def __init__(
         num_blocks: Sequence[int],
         use_noisy: bool = False,
         bottleneck: Optional[Bottleneck] = None,
+        norm: float = 0.0,
     ):
         super().__init__()
         num_layers = len(multipliers) - 1
         self.bottleneck = bottleneck
         self.use_noisy = use_noisy
+        self.use_norm = norm > 0.0
+        self.norm = norm
 
         assert len(factors) >= num_layers and len(num_blocks) >= num_layers
 
@@ -1174,6 +1185,9 @@ def __init__(
     def encode(
         self, x: Tensor, with_info: bool = False
     ) -> Union[Tensor, Tuple[Tensor, Any]]:
+        if self.use_norm:
+            x = wave_norm(x, peak=self.norm)
+
         x = self.to_in(x)
         for downsample in self.downsamples:
             x = downsample(x)
@@ -1190,7 +1204,12 @@ def decode(self, x: Tensor) -> Tensor:
             x = upsample(x)
         if self.use_noisy:
             x = torch.cat([x, torch.randn_like(x)], dim=1)
-        return self.to_out(x)
+        x = self.to_out(x)
+
+        if self.use_norm:
+            x = wave_unnorm(x, peak=self.norm)
+
+        return x
 
 
 class MultiEncoder1d(nn.Module):

audio_diffusion_pytorch/utils.py

@@ -83,3 +83,16 @@ def downsample(waveforms: Tensor, factor: int, **kwargs) -> Tensor:
 
 def upsample(waveforms: Tensor, factor: int, **kwargs) -> Tensor:
     return resample(waveforms, factor_in=1, factor_out=factor, **kwargs)
+
+
+def wave_norm(x: Tensor, bits: int = 24, peak: float = 0.5) -> Tensor:
+    mu = 2 ** bits
+    x = torch.sign(x) * torch.log1p(mu * torch.abs(x)) / math.log1p(mu)
+    return x * peak
+
+
+def wave_unnorm(x: Tensor, bits: int = 24, peak: float = 0.5) -> Tensor:
+    x = x / peak
+    mu = 2 ** bits
+    x = torch.sign(x) * (torch.exp(torch.abs(x) * math.log1p(mu)) - 1) / mu
+    return x

setup.py

@@ -3,7 +3,7 @@
 setup(
     name="audio-diffusion-pytorch",
     packages=find_packages(exclude=[]),
-    version="0.0.52",
+    version="0.0.53",
     license="MIT",
     description="Audio Diffusion - PyTorch",
     long_description_content_type="text/markdown",