Delete src/miocodec/pupu/modules/vocoder_blocks/norm2d.py with huggingface_hub
Browse files
src/miocodec/pupu/modules/vocoder_blocks/norm2d.py
DELETED
|
@@ -1,84 +0,0 @@
|
|
| 1 |
-
import torch.nn as nn
|
| 2 |
-
import einops
|
| 3 |
-
from torch.nn.utils import spectral_norm, weight_norm
|
| 4 |
-
|
| 5 |
-
CONV_NORMALIZATIONS = frozenset(
|
| 6 |
-
[
|
| 7 |
-
"none",
|
| 8 |
-
"weight_norm",
|
| 9 |
-
"spectral_norm",
|
| 10 |
-
"time_layer_norm",
|
| 11 |
-
"layer_norm",
|
| 12 |
-
"time_group_norm",
|
| 13 |
-
]
|
| 14 |
-
)
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
class ConvLayerNorm(nn.LayerNorm):
|
| 18 |
-
"""
|
| 19 |
-
Convolution-friendly LayerNorm that moves channels to last dimensions
|
| 20 |
-
before running the normalization and moves them back to original position right after.
|
| 21 |
-
"""
|
| 22 |
-
|
| 23 |
-
def __init__(self, normalized_shape, **kwargs):
|
| 24 |
-
super().__init__(normalized_shape, **kwargs)
|
| 25 |
-
|
| 26 |
-
def forward(self, x):
|
| 27 |
-
x = einops.rearrange(x, "b ... t -> b t ...")
|
| 28 |
-
x = super().forward(x)
|
| 29 |
-
x = einops.rearrange(x, "b t ... -> b ... t")
|
| 30 |
-
return
|
| 31 |
-
|
| 32 |
-
|
| 33 |
-
def apply_parametrization_norm(module: nn.Module, norm: str = "none") -> nn.Module:
|
| 34 |
-
assert norm in CONV_NORMALIZATIONS
|
| 35 |
-
if norm == "weight_norm":
|
| 36 |
-
return weight_norm(module)
|
| 37 |
-
elif norm == "spectral_norm":
|
| 38 |
-
return spectral_norm(module)
|
| 39 |
-
else:
|
| 40 |
-
# We already check was in CONV_NORMALIZATION, so any other choice
|
| 41 |
-
# doesn't need reparametrization.
|
| 42 |
-
return module
|
| 43 |
-
|
| 44 |
-
|
| 45 |
-
def get_norm_module(
|
| 46 |
-
module: nn.Module, causal: bool = False, norm: str = "none", **norm_kwargs
|
| 47 |
-
) -> nn.Module:
|
| 48 |
-
"""Return the proper normalization module. If causal is True, this will ensure the returned
|
| 49 |
-
module is causal, or return an error if the normalization doesn't support causal evaluation.
|
| 50 |
-
"""
|
| 51 |
-
assert norm in CONV_NORMALIZATIONS
|
| 52 |
-
if norm == "layer_norm":
|
| 53 |
-
assert isinstance(module, nn.modules.conv._ConvNd)
|
| 54 |
-
return ConvLayerNorm(module.out_channels, **norm_kwargs)
|
| 55 |
-
elif norm == "time_group_norm":
|
| 56 |
-
if causal:
|
| 57 |
-
raise ValueError("GroupNorm doesn't support causal evaluation.")
|
| 58 |
-
assert isinstance(module, nn.modules.conv._ConvNd)
|
| 59 |
-
return nn.GroupNorm(1, module.out_channels, **norm_kwargs)
|
| 60 |
-
else:
|
| 61 |
-
return nn.Identity()
|
| 62 |
-
|
| 63 |
-
|
| 64 |
-
class NormConv2d(nn.Module):
|
| 65 |
-
"""Wrapper around Conv2d and normalization applied to this conv
|
| 66 |
-
to provide a uniform interface across normalization approaches.
|
| 67 |
-
"""
|
| 68 |
-
|
| 69 |
-
def __init__(
|
| 70 |
-
self,
|
| 71 |
-
*args,
|
| 72 |
-
norm: str = "none",
|
| 73 |
-
norm_kwargs={},
|
| 74 |
-
**kwargs,
|
| 75 |
-
):
|
| 76 |
-
super().__init__()
|
| 77 |
-
self.conv = apply_parametrization_norm(nn.Conv2d(*args, **kwargs), norm)
|
| 78 |
-
self.norm = get_norm_module(self.conv, causal=False, norm=norm, **norm_kwargs)
|
| 79 |
-
self.norm_type = norm
|
| 80 |
-
|
| 81 |
-
def forward(self, x):
|
| 82 |
-
x = self.conv(x)
|
| 83 |
-
x = self.norm(x)
|
| 84 |
-
return x
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|