File size: 955 Bytes
8aa674c |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 |
import torch.nn as nn
from einops.layers.torch import Rearrange
class NormModule(nn.Module):
def __init__(self, norm_type, norm_dim, **kwargs): # ["BN,"GN","LN"]
super().__init__()
assert norm_type in ["BN", "GN", "LN"]
if norm_type == "BN":
self.norm = nn.BatchNorm1d(num_features=norm_dim, **kwargs)
elif norm_type == "GN":
self.norm = nn.GroupNorm(num_channels=norm_dim, **kwargs)
elif norm_type == "LN":
self.norm = nn.Sequential(
Rearrange("b c t -> b t c"),
nn.LayerNorm(norm_dim),
Rearrange("b t c-> b c t"),
)
def forward(self, x):
return self.norm(x)
class LayerNorm(nn.LayerNorm):
def __init__(self, *args, **kwarg):
super().__init__(*args, **kwarg)
def forward(self, x):
assert x.dim() == 3
return super().forward(x.permute(0, 2, 1)).permute(0, 2, 1)
|