|
import torch |
|
from torch import nn |
|
|
|
|
|
|
|
|
|
|
|
class MemoryUnit(nn.Module): |
|
def __init__(self,dim): |
|
super().__init__() |
|
|
|
|
|
self.norm_token = nn.LayerNorm(dim) |
|
self.proj_1 = nn.Linear(dim,dim) |
|
self.proj_2 = nn.Linear(dim,dim) |
|
self.proj_3 = nn.Linear(dim,dim) |
|
|
|
def forward(self, x): |
|
|
|
x = self.norm_token(x) |
|
u, v = x, x |
|
u = self.proj_1(u) |
|
u = self.norm_token(u) |
|
v = self.proj_2(v) |
|
g = u * v |
|
x = self.proj_3(g) |
|
x = self.norm_token(x) |
|
|
|
return x |
|
|
|
class InteractionUnit(nn.Module): |
|
def __init__(self,dim,score_dim): |
|
super().__init__() |
|
|
|
|
|
self.norm_token = nn.LayerNorm(dim) |
|
self.norm_score = nn.LayerNorm(score_dim) |
|
|
|
def forward(self, x): |
|
|
|
x = self.norm_token(x) |
|
q,k,v = x,x,x |
|
score = torch.matmul(q, k.transpose(-1, -2)) |
|
interaction = self.norm_score(score) |
|
x = torch.matmul(interaction,v) |
|
x = self.norm_token(x) |
|
|
|
return x |
|
|
|
class InteractorBlock(nn.Module): |
|
def __init__(self, d_model, num_tokens): |
|
super().__init__() |
|
|
|
|
|
self.memory = MemoryUnit(d_model) |
|
self.interaction = InteractionUnit(d_model,num_tokens) |
|
|
|
def forward(self, x): |
|
|
|
residual = x |
|
|
|
x = self.interaction(x) |
|
|
|
x = x + residual |
|
|
|
residual = x |
|
|
|
x = self.memory(x) |
|
|
|
|
|
out = x + residual |
|
|
|
|
|
return out |
|
|
|
|
|
|
|
class Interactor(nn.Module): |
|
def __init__(self, d_model,num_tokens, num_layers): |
|
super().__init__() |
|
|
|
self.model = nn.Sequential( |
|
*[InteractorBlock(d_model,num_tokens) for _ in range(num_layers)] |
|
) |
|
|
|
def forward(self, x): |
|
|
|
return self.model(x) |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|