|
|
import torch.nn as nn |
|
|
from einops.layers.torch import Rearrange |
|
|
|
|
|
|
|
|
class NormModule(nn.Module): |
|
|
def __init__(self, norm_type, norm_dim, **kwargs): |
|
|
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) |
|
|
|
