做照片书网站好自媒体平台源码

1--概述

        1. 自回归 TransFormer 规定Token只能看到自身及前面的Token，因此需生成一个符合规定的Attention Mask；（代码提供了两种方式自回归Attention Mask的定义方式）；

        2. 使用Cross Attention实现条件模态和输入模态之间的模态融合，输入模态作为Query，条件模态作为Key和Value；

2--代码

import torch
import torch.nn as nnclass CrossAttention(nn.Module):def __init__(self, embed_dim: int, num_heads: int):super().__init__()self.cross_attn = nn.MultiheadAttention(embed_dim, num_heads)def forward(self, input_x: torch.Tensor, condition: torch.Tensor, attn_mask: torch.Tensor = None):'''query: input_xkey: conditionval: condition'''input_x = self.cross_attn(input_x, condition, condition, attn_mask=attn_mask)[0]return input_xclass Cond_Autoregressive_layer(nn.Module):def __init__(self, input_dim: int, condtion_dim: int, embed_dim: int, num_heads: int):super(Cond_Autoregressive_layer, self).__init__()self.linear1 = nn.Linear(input_dim, embed_dim)self.linear2 = nn.Linear(condtion_dim, embed_dim)self.cond_multihead_attn = CrossAttention(embed_dim = embed_dim, num_heads = num_heads)def forward(self, input_x: torch.Tensor, conditon: torch.Tensor, attention_mask1: torch.Tensor, attention_mask2: torch.Tensor):# q, k, v, attention mask, here we set key and value are both condtion y1 = self.cond_multihead_attn(self.linear1(input_x), self.linear2(conditon), attn_mask = attention_mask1)y2 = self.cond_multihead_attn(self.linear1(input_x), self.linear2(conditon), attn_mask = attention_mask2)return y1, y2if __name__ == "__main__":# set sequence len, embedding dim, multi attention headseq_length = 10input_dim = 32condtion_dim = 128embed_dim = 64num_heads = 8# init input sequence and condtioninput_x = torch.randn(seq_length, 1, input_dim)condtion = torch.randn(seq_length, 1, condtion_dim)# create two attention mask (actually they have the same function)attention_mask1 = torch.triu((torch.ones((seq_length, seq_length)) == 1), diagonal=1) # bool typeattention_mask2 = attention_mask1.float() # True->1 False->0attention_mask2 = attention_mask2.masked_fill(attention_mask2 == 1, float("-inf"))  # Convert ones to -inf# init modelAG_layer = Cond_Autoregressive_layer(input_dim, condtion_dim, embed_dim, num_heads)# forwardy1, y2 = AG_layer(input_x, condtion, attention_mask1, attention_mask2)# here we demonstrate the attention_mask1 and attention_mask2 have the same functionassert(y1[0].equal(y2[0]))