File size: 1,723 Bytes
589ece6 e7c8685 589ece6 e7c8685 589ece6 e7c8685 589ece6 e7c8685 589ece6 e7c8685 589ece6 e7c8685 ab58365 589ece6 e7c8685 589ece6 e7c8685 589ece6 e7c8685 589ece6 e7c8685 589ece6 e7c8685 589ece6 | 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 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 | import torch
from torch import nn
class GatingUnit(nn.Module):
def __init__(self, dim):
super().__init__()
self.proj_1 = nn.Linear(dim, dim)
self.proj_2 = nn.Linear(dim, dim)
self.proj_3 = nn.Linear(dim, dim)
self.silu = nn.SiLU()
def forward(self, x):
u, v = x, x
u = self.proj_1(u)
u = self.silu(u)
v = self.proj_2(v)
g = u * v
g = self.proj_3(g)
out = g
return out
class NormalizerBlock(nn.Module):
def __init__(self, d_model, num_tokens):
super().__init__()
self.norm_global = nn.LayerNorm(d_model * num_tokens, elementwise_affine=False)
self.norm_local = nn.LayerNorm(d_model, elementwise_affine=False)
self.gating = GatingUnit(d_model)
def forward(self, x):
residual = x
dim0 = x.shape[0]
dim1 = x.shape[1]
dim2 = x.shape[2]
x = x.reshape([dim0,dim1*dim2])
x = self.norm_global(x)
x = x.reshape([dim0,dim1,dim2])
x = x + residual
residual = x
x = self.norm_local(x)
x = self.gating(x)
out = x + residual
return out
class Normalizer(nn.Module):
def __init__(self, d_model, num_tokens, num_layers):
super().__init__()
self.model = nn.Sequential(
*[NormalizerBlock(d_model, num_tokens) for _ in range(num_layers)]
)
def forward(self, x):
return self.model(x)
|