SelfAttention和MultiHeadAttion实现demo

#encoding:utf-8

from math import sqrt

import torch

import torch.nn as nn

class Self_Attention(nn.Module):

def init(self, input_dim, dim_k, dim_v):

super(Self_Attention, self). init()

self.q = nn.Linear(input_dim, dim_k)

self.k = nn.Linear(input_dim, dim_k)

self.v = nn.Linear(input_dim, dim_v)

self.norm_fact = 1 / sqrt(dim_k)

def forward(self, x):

print("x.shape:", x.shape)

print("q.shape:", self.q.shape)

Q = self.q(x)

print("Q.shape:", Q.shape)

K = self.k(x)

print("K.shape:", K.shape)

V = self.v(x)

print("V.shape:", V.shape)

atten = nn.Softmax(dim=-1)(torch.bmm(Q,K.permute(0,2,1))) * self.norm_fact

output = torch.bmm(atten, V)

return output

print("\n")

print("self attention:")

x = torch.randn(4,3,1024)

print(x)

print("input size:", x.size())

self_attention = Self_Attention(1024,128,5)

res = self_attention(x)

print("\n")

print(res)

print("output size:", res.size())

print("\n")

class Self_Attention_Muti_Head(nn.Module):

def init(self, input_dim, dim_k, dim_v, nums_head):

super(Self_Attention_Muti_Head, self).init()

assert dim_k % nums_head == 0

assert dim_v % nums_head == 0

self.q = nn.Linear(input_dim, dim_k)

self.k = nn.Linear(input_dim, dim_k)

self.v = nn.Linear(input_dim, dim_v)

self.nums_head = nums_head

self.dim_k = dim_k

self.dim_v = dim_v

self._norm_fact = 1 / sqrt(dim_k)

def forward(self, x):

Q = self.q(x).reshape(-1, x.shape[0], x.shape[1], self.dim_k//self.nums_head)

K = self.k(x).reshape(-1, x.shape[0], x.shape[1], self.dim_k//self.nums_head)

V = self.v(x).reshape(-1, x.shape[0], x.shape[1], self.dim_v//self.nums_head)

print("x.shape:", x.shape)

print("Q.shape", Q.size())

atten = nn.Softmax(dim=-1)(torch.matmul(Q, K.permute(0,1,3,2)))

output = torch.matmul(atten, V).reshape(x.shape[0], x.shape[1], -1)

return output

print("\n")

print("multi head attention:")

x = torch.randn(4,3,1024)

print(x)

print(x.size())

self_attention = Self_Attention_Muti_Head(1024,128,6,2)

res = self_attention(x)

print("\n")

print(res)

print(res.size())


有个问题:

根据文献:https://arxiv.org/pdf/1911.02150.pdf,感觉这里说的Multi Head Attenion和 Group Query Attention意思是一样的:

这下面这张经典的图中的的Grouped-query意思是一样的:

哪里没理解到位?

相关推荐
倔强青铜三5 分钟前
苦练Python第23天:元组秘籍与妙用
人工智能·python·面试
Norvyn_729 分钟前
LeetCode|Day18|20. 有效的括号|Python刷题笔记
笔记·python·leetcode
AndrewHZ36 分钟前
【图像处理基石】如何入门色彩评估?
图像处理·人工智能·深度学习·色彩科学·hvs·色彩评估·颜色工程
chao_7891 小时前
更灵活方便的初始化、清除方法——fixture【pytest】
服务器·自动化测试·python·pytest
心情好的小球藻1 小时前
Python应用进阶DAY9--类型注解Type Hinting
开发语言·python
都叫我大帅哥1 小时前
LangChain加载HTML内容全攻略:从入门到精通
python·langchain
静心问道1 小时前
TrOCR: 基于Transformer的光学字符识别方法,使用预训练模型
人工智能·深度学习·transformer·多模态
亲持红叶1 小时前
GLU 变种:ReGLU 、 GEGLU 、 SwiGLU
人工智能·深度学习·神经网络·激活函数
惜.己1 小时前
使用python读取json数据,简单的处理成元组数组
开发语言·python·测试工具·json
都叫我大帅哥3 小时前
Python的Optional:让你的代码优雅处理“空值”危机
python