AI 基础知识十三 Transformer注意力机制(Attention)

指掀涛澜天下惊2026-04-03 10:52

注意力机制

Transformer 的核心是自注意力 与多头注意力，让序列每个位置都能动态关注全局相关信息，并行捕捉长程依赖。

自注意力公式

多头注意力公式

计算步骤

参照论文说明

  1. Q、 K矩阵相乘

  2. 缩放处理

  3. 加掩码处理 是可选项

  4. Softmax 归一化指数函数

  5. 与V矩阵相乘

本着简单的原则，用一个实例来说明Q,K,V计算过程

实例

自注意力

实现代码

例子 "Welcome to Machine Learning Pad Pad" 经过 词嵌入 和**位置编码，**得到6X4矩阵,为了方便计算对这个矩阵手动设置特定的数据

cpp 复制代码 
/*

	{"Pad",      0},
	{"Welcome",  1},
	{"to",       2},
	{"Machine",  3},
	{"Learning", 4}

	1. Welcome to Machine Learning Pad Pad -- > [1,2,3,4,0,0]

	2. Embedding + PositionalEncoding        ->  x

	3. x: [6 ,4]  

*/

	auto x = torch::tensor({
				{{1.0, 0.0, 0.0, 0.0}, // Welcome
				 {2.0, 0.0, 0.0, 0.0}, // to
				 {3.0, 0.0, 0.0, 0.0}, // Machine
				 {4.0, 0.0, 0.0, 0.0}, // Learning
				 {0.0, 0.0, 0.0, 0.0}, // Pad
				 {0.0, 0.0, 0.0, 0.0}  // Pad
				} }, torch::kFloat);

Q,K,V是一组权重 它的词嵌入的 维度**，**为了方便计算都它们设定单位矩阵

cpp 复制代码 
class SelfAttention : public torch::nn::Module
{
public:
	SelfAttention()
	{


	}
	void InitQKV(int64_t dim)
	{
		auto linear = torch::nn::LinearOptions(dim, dim).bias(false);

		Q = register_module("q", torch::nn::Linear(linear));
		K = register_module("k", torch::nn::Linear(linear));
		V = register_module("v", torch::nn::Linear(linear));

		norm_fact = 1.0 / sqrt(dim); // 缩放

		auto onesw = torch::eye(dim);  //单位矩阵

		Q->weight.set_data(onesw);
		K->weight.set_data(onesw);
		V->weight.set_data(onesw);
	}

	
	torch::nn::Linear Q{ nullptr };
	torch::nn::Linear K{ nullptr };
	torch::nn::Linear V{ nullptr };
	double norm_fact = 0 ;
};

torch::nn::Transformer 要求输入张量形状**seq, batch, dim,** 这里简单化为**seq,dim**

参照论文实现计算步骤

cpp 复制代码 
	auto forward(torch::Tensor x,torch::Tensor mask = {})
	{
		torch::Tensor q ;
		torch::Tensor k ;
		torch::Tensor v;
		torch::Tensor kt;
		torch::Tensor out;

		auto dim = x.dim();
	
	
		// x: [seq, dim]
		InitQKV(x.size(1));
		
		/// 1.输入x 与 q k v 运算  q k v是 单位矩阵所以 q k v = x
		 q = Q->forward(x);
		 k = K->forward(x);
		 v = V->forward(x);

		 cout << "q k v \n" << q << endl;

		
		kt = k.transpose(0, 1);// kt 是 k 的置换矩阵  kt: [dim,seq]
		

		cout << "kt \n" << kt << endl;

		auto attn_score = torch::matmul(q, kt); //2.  q:[seq, dim] X kt: [dim,seq] -> [seq, seq]
		cout << "q X kt \n" << attn_score << endl;

		attn_score = attn_score * norm_fact;     //3. 矩阵缩放
		cout << "scale q.X.kt  \n" << attn_score << endl;

		if (mask.defined())
		{
			attn_score += mask;
		}


		attn_score = torch::softmax(attn_score, -1);//4. Softmax 归一化指数函数
		cout << "torch::softmax q.X.kt  \n" << attn_score << endl;

		out = torch::matmul(attn_score, v); /// 5.与V矩阵相乘  [seq, seq] X v:[seq, dim] -> [seq, dim]

		cout << "torch::matmul V  \n" << out << endl;
	
		return out;
	}

重点解析

  1. = =

实际意义是

矩阵相乘结果

每个字符都能其他字符产生运算，也就是它能根据上下文来确定语意，字符序列长度N，Transformer时间复杂度为

  1. Softmax 归一化指数函数

数学公: 式 输入向量

=

第 i 个元素的 Softmax 输出为

每行内所有数据相加等于1, 原数据按一定比例缩小

  1. 与V矩阵相乘

qkv现在全部建立关系了

当要求输入张量形状**seq, batch, dim**时，其流程都一样，要变换处理张量

高维张量矩阵相乘

公式： a..,..., M,N * b...,...,N, K = ..,...,M, K 看到最后两维和两维矩阵相乘一样

整理之的代码，支持两三维输入张量

cpp 复制代码 
	auto forward(torch::Tensor x,torch::Tensor mask = {})
	{
		torch::Tensor q ;
		torch::Tensor k ;
		torch::Tensor v;
		torch::Tensor kt;
		torch::Tensor out;

		auto dim = x.dim();
	

		if (dim == 3)
		{
			//x:  [batch, seq, dim]  --->  [seq, batch, dim]
			x = x.permute({1,0,2});
			InitQKV(x.size(2));
		}
		else
		{
			// x: [seq, dim]
			InitQKV(x.size(1));
		}
		/// 1.输入x 与 q k v 运算  q k v是 单位矩阵所以 q k v = x
		 q = Q->forward(x);
		 k = K->forward(x);
		 v = V->forward(x);

		 cout << "q k v \n" << q << endl;

		if (dim == 3)
		{
			kt = k.permute({ 1,2,0 });
		 	v = v.permute({ 1,0,2 });
		}
		else
		{
			kt = k.transpose(0, 1);// kt 是 k 的置换矩阵  kt: [dim,seq]
		}

	
		cout << "kt \n" << kt << endl;

		auto attn_score = torch::matmul(q, kt); //2.  
		cout << "q X kt \n" << attn_score << endl;

		attn_score = attn_score * norm_fact;     //3. 矩阵缩放
		cout << "scale q.X.kt  \n" << attn_score << endl;

		if (mask.defined())
		{
			attn_score += mask;
		}


		attn_score = torch::softmax(attn_score, -1);//4. Softmax 归一化指数函数
		cout << "torch::softmax q.X.kt  \n" << attn_score << endl;

		out = torch::matmul(attn_score, v); /// 5. qKt * V

		cout << "torch::matmul V  \n" << out << endl;
	
		return out;
	}

多头注意力

在"自注意力"的基础上增加

  1. 维度被为多份 分别用于Q K V 计算

2.将多份重新拼接

3.最后加输出投影

输入两维张量时，写一个函数forward

  1. 输入张量x 形状**seq, dim， q、k、v** 形状 seq, dim

  2. 将q、k、v 拆分成 H, S, Dk ， seq简写S, H：头数量, Dk = dim/ H

cpp 复制代码 
q = q.view({ seq,H,Dk }); //q: [seq, dim] ->   [S, H, Dk] 
k = k.view({ seq,H,Dk });
v = v.view({ seq,H,Dk });

q = q.permute({ 1,0,2 }); //[S, H, Dk] --->[H, S, Dk]
k = k.permute({ 1,0,2 });
v = v.permute({ 1,0,2 });
  1. 调 形状 H, Dk, S
cpp 复制代码 
auto kt = k.permute({ 0,2,1 }); //kt:  [H, S, Dk] --> [H, Dk, S]
  1. 与V矩阵相乘之后 输出形状 H, S, Dk， 要转换成 S, H, Dk
cpp 复制代码 
auto out = torch::matmul(attn_score, v); // [H, S, S] * [H, S, Dk]  ->  out: [H, S, Dk]
  1. S, H, Dk 拼接成 **seq, dim，**最后输出投影
cpp 复制代码 
out = out.transpose(1, 0).contiguous().view({ seq, dim }); //  [H, S, Dk] --> [S, H, Dk] -> [seq, dim]

cout << "torch::matmul QK * V  \n" << out.squeeze() << endl;

out = Wo->forward(out);

输入三维张量时，写一个函数forward2去实现，除了张量形状调整不同外其他都一样，实现细节只能看代码

cpp 复制代码 
	auto forward(torch::Tensor x, int64_t head = 2, torch::Tensor mask = {})
	{
		x.squeeze_(); //x: [batch, seq ,dim]  -->   [seq, dim]
		assert(x.dim() == 2);
		//x: [seq, dim]
		  
		InitQKV(x.size(1), head);

		auto seq = x.size(0);
		auto dim = x.size(1);

		auto q = Q->forward(x);
		auto k = K->forward(x);
		auto v = V->forward(x);
		q = q.view({ seq,H,Dk }); //q: [seq, dim] ->   [S, H, Dk] 
		k = k.view({ seq,H,Dk });
		v = v.view({ seq,H,Dk });

		q = q.permute({ 1,0,2 }); //[S, H, Dk] --->[H, S, Dk]
		k = k.permute({ 1,0,2 });
		v = v.permute({ 1,0,2 });

		cout << "q k v \n" << q << endl;

		auto kt = k.permute({ 0,2,1 }); //kt:  [H, S, Dk] --> [H, Dk, S]

		cout << "kt \n" << kt << endl;

		auto attn_score = torch::matmul(q, kt);  // [H, S, Dk] *  [H, Dk, S]
		cout << "q X kt \n" << attn_score << endl;


		attn_score = attn_score * norm_fact;
		cout << "scale q.X.kt  \n" << attn_score << endl;

		if (mask.defined())
		{
			attn_score += mask;
		}

		attn_score = torch::softmax(attn_score, -1); /// attn_score: [H, S, S]
		cout << "torch::softmax q.X.kt  \n" << attn_score.squeeze() << endl;

		auto out = torch::matmul(attn_score, v); // [H, S, S] * [H, S, Dk]  ->  out: [H, S, Dk]
		out = out.transpose(1, 0).contiguous().view({ seq, dim }); //  [H, S, Dk] --> [S, H, Dk] -> [seq, dim]

		cout << "torch::matmul QK * V  \n" << out.squeeze() << endl;
		
		out = Wo->forward(out);

		return out;
	}


	auto forward2(torch::Tensor x, int64_t head = 2,torch::Tensor mask = {})
	{
		assert(x.dim() == 3);
	  
		x = x.permute({ 1,0,2 });  //   x: x: [batch, seq, dim]-->  [seq, batch, dim]

		InitQKV(x.size(2), head);

		auto seq = x.size(0);
		auto batch = x.size(1);
		auto dim = x.size(2);

		auto q = Q->forward(x);
		auto k = K->forward(x);
		auto v = V->forward(x);
		q = q.view({ seq,batch,H,Dk}); //q: [seq, batch, dim] ->   [S, B, H, Dk] 
		k = k.view({ seq,batch,H,Dk });
		v = v.view({ seq,batch,H,Dk });
		
		q = q.permute({1,2,0,3}); //[S, B, H, Dk] --->[B, H, S, Dk]
		k = k.permute({ 1,2,0,3 });
		v = v.permute({ 1,2,0,3 });
		
		cout << "q k v \n" << q << endl;

		auto kt = k.permute({ 0,1,3,2}); //kt:  [B, H, S, Dk] --> [B, H, Dk, S]

		cout << "kt \n" << kt.squeeze() << endl;

		auto attn_score = torch::matmul(q, kt);
		cout << "q X kt \n" << attn_score << endl;


		attn_score = attn_score * norm_fact;
		cout << "scale q.X.kt  \n" << attn_score << endl;

		if (mask.defined())
		{
			attn_score += mask;
		}

		attn_score = torch::softmax(attn_score, -1); /// attn_score: [B, H, S, S]
		cout << "torch::softmax q.X.kt  \n" << attn_score << endl;

		auto out = torch::matmul(attn_score, v); // [B, H, S, S] * [B, H, S, Dk]  ->  out: [B, H, S, Dk]
		out = out.transpose(1, 2).contiguous().view({ seq,batch, dim }); //  [B, H, S, Dk] --> [B, S, H, Dk] -> [seq,batch, dim]
		
		cout << "torch::matmul QK * V  \n" << out << endl;
		
		out = Wo->forward(out);

		return out;
	}

完整代码

cpp 复制代码 
#include <torch/torch.h>
#include <iostream>
#include <torch/serialize.h>
#include <regex>
//#include <iostream>
#include <fstream>
using namespace std;


class FeedForwardNet : public torch::nn::Module
{
	//Q = register_module("q", torch::nn::Linear(linear));
};



class SelfAttention : public torch::nn::Module
{
public:
	SelfAttention()
	{


	}
	void InitQKV(int64_t dim)
	{
		auto linear = torch::nn::LinearOptions(dim, dim).bias(false);

		Q = register_module("q", torch::nn::Linear(linear));
		K = register_module("k", torch::nn::Linear(linear));
		V = register_module("v", torch::nn::Linear(linear));

		norm_fact = 1.0 / sqrt(dim); // 缩放

		auto onesw = torch::eye(dim); //单位矩阵

		Q->weight.set_data(onesw);
		K->weight.set_data(onesw);
		V->weight.set_data(onesw);
	}

	
	auto forward(torch::Tensor x,torch::Tensor mask = {})
	{
		torch::Tensor q ;
		torch::Tensor k ;
		torch::Tensor v;
		torch::Tensor kt;
		torch::Tensor out;

		auto dim = x.dim();
	

		if (dim == 3)
		{
			//x:  [batch, seq, dim]  --->  [seq, batch, dim]
			x = x.permute({1,0,2});
			InitQKV(x.size(2));
		}
		else
		{
			// x: [seq, dim]
			InitQKV(x.size(1));
		}
		/// 1.输入x 与 q k v 运算  q k v是 单位矩阵所以 q k v = x
		 q = Q->forward(x);
		 k = K->forward(x);
		 v = V->forward(x);

		 cout << "q k v \n" << q << endl;

		if (dim == 3)
		{
			kt = k.permute({ 1,2,0 });
		 	v = v.permute({ 1,0,2 });
		}
		else
		{
			kt = k.transpose(0, 1);// kt 是 k 的置换矩阵  kt: [dim,seq]
		}

	
		cout << "kt \n" << kt << endl;

		auto attn_score = torch::matmul(q, kt); //2.  
		cout << "q X kt \n" << attn_score << endl;

		attn_score = attn_score * norm_fact;     //3. 矩阵缩放
		cout << "scale q.X.kt  \n" << attn_score << endl;

		if (mask.defined())
		{
			attn_score += mask;
		}


		attn_score = torch::softmax(attn_score, -1);//4. Softmax 归一化指数函数
		cout << "torch::softmax q.X.kt  \n" << attn_score << endl;

		out = torch::matmul(attn_score, v); /// 5. qKt * V

		cout << "torch::matmul V  \n" << out << endl;
	
		return out;
	}

	torch::nn::Linear Q{ nullptr };
	torch::nn::Linear K{ nullptr };
	torch::nn::Linear V{ nullptr };
	double norm_fact = 0 ;
};

class MultiHeadAttention: public torch::nn::Module
{
public:
	void InitQKV(int64_t dim, int64_t head=2)
	{
		assert(dim % head == 0);

		auto linear = torch::nn::LinearOptions(dim, dim).bias(false);

		Q = register_module("q", torch::nn::Linear(linear));
		K = register_module("k", torch::nn::Linear(linear));
		V = register_module("v", torch::nn::Linear(linear));
		Wo = register_module("Wo", torch::nn::Linear(linear)); // 输出投影

		norm_fact = 1.0 / sqrt(dim);
		
		Dk = dim / head;
		H = head;

		auto onesw = torch::eye(dim);   
		Q->weight.set_data(onesw);
		K->weight.set_data(onesw);
		V->weight.set_data(onesw);
		Wo->weight.set_data(onesw);
	}

	auto forward(torch::Tensor x, int64_t head = 2, torch::Tensor mask = {})
	{
		x.squeeze_(); //x: [batch, seq ,dim]  -->   [seq, dim]
		assert(x.dim() == 2);
		//x: [seq, dim]
		  
		InitQKV(x.size(1), head);

		auto seq = x.size(0);
		auto dim = x.size(1);

		auto q = Q->forward(x);
		auto k = K->forward(x);
		auto v = V->forward(x);
		q = q.view({ seq,H,Dk }); //q: [seq, dim] ->   [S, H, Dk] 
		k = k.view({ seq,H,Dk });
		v = v.view({ seq,H,Dk });

		q = q.permute({ 1,0,2 }); //[S, H, Dk] --->[H, S, Dk]
		k = k.permute({ 1,0,2 });
		v = v.permute({ 1,0,2 });

		cout << "q k v \n" << q << endl;

		auto kt = k.permute({ 0,2,1 }); //kt:  [H, S, Dk] --> [H, Dk, S]

		cout << "kt \n" << kt << endl;

		auto attn_score = torch::matmul(q, kt);  // [H, S, Dk] *  [H, Dk, S]
		cout << "q X kt \n" << attn_score << endl;


		attn_score = attn_score * norm_fact;
		cout << "scale q.X.kt  \n" << attn_score << endl;

		if (mask.defined())
		{
			attn_score += mask;
		}

		attn_score = torch::softmax(attn_score, -1); /// attn_score: [H, S, S]
		cout << "torch::softmax q.X.kt  \n" << attn_score.squeeze() << endl;

		auto out = torch::matmul(attn_score, v); // [H, S, S] * [H, S, Dk]  ->  out: [H, S, Dk]
		out = out.transpose(1, 0).contiguous().view({ seq, dim }); //  [H, S, Dk] --> [S, H, Dk] -> [seq, dim]

		cout << "torch::matmul QK * V  \n" << out.squeeze() << endl;
		
		out = Wo->forward(out);

		return out;
	}


	auto forward2(torch::Tensor x, int64_t head = 2,torch::Tensor mask = {})
	{
		assert(x.dim() == 3);
	  
		x = x.permute({ 1,0,2 });  //   x: x: [batch, seq, dim]-->  [seq, batch, dim]

		InitQKV(x.size(2), head);

		auto seq = x.size(0);
		auto batch = x.size(1);
		auto dim = x.size(2);

		auto q = Q->forward(x);
		auto k = K->forward(x);
		auto v = V->forward(x);
		q = q.view({ seq,batch,H,Dk}); //q: [seq, batch, dim] ->   [S, B, H, Dk] 
		k = k.view({ seq,batch,H,Dk });
		v = v.view({ seq,batch,H,Dk });
		
		q = q.permute({1,2,0,3}); //[S, B, H, Dk] --->[B, H, S, Dk]
		k = k.permute({ 1,2,0,3 });
		v = v.permute({ 1,2,0,3 });
		
		cout << "q k v \n" << q << endl;

		auto kt = k.permute({ 0,1,3,2}); //kt:  [B, H, S, Dk] --> [B, H, Dk, S]

		cout << "kt \n" << kt.squeeze() << endl;

		auto attn_score = torch::matmul(q, kt);
		cout << "q X kt \n" << attn_score << endl;


		attn_score = attn_score * norm_fact;
		cout << "scale q.X.kt  \n" << attn_score << endl;

		if (mask.defined())
		{
			attn_score += mask;
		}

		attn_score = torch::softmax(attn_score, -1); /// attn_score: [B, H, S, S]
		cout << "torch::softmax q.X.kt  \n" << attn_score << endl;

		auto out = torch::matmul(attn_score, v); // [B, H, S, S] * [B, H, S, Dk]  ->  out: [B, H, S, Dk]
		out = out.transpose(1, 2).contiguous().view({ seq,batch, dim }); //  [B, H, S, Dk] --> [B, S, H, Dk] -> [seq,batch, dim]
		
		cout << "torch::matmul QK * V  \n" << out << endl;
		
		out = Wo->forward(out);

		return out;
	}


	torch::nn::Linear Q{ nullptr };
	torch::nn::Linear K{ nullptr };
	torch::nn::Linear V{ nullptr };
	torch::nn::Linear Wo{ nullptr };

	double norm_fact = 0;
	int64_t Dk;
	int64_t H;
};


void TransformerAttentionMain()
{


	auto x = torch::tensor({
				{{1.0, 0.0, 0.0, 0.0}, // Welcome
				 {2.0, 0.0, 0.0, 0.0}, // to
				 {3.0, 0.0, 0.0, 0.0}, // Machine
				 {4.0, 0.0, 0.0, 0.0}, // Learning
				 {0.0, 0.0, 0.0, 0.0}, // Pad
				 {0.0, 0.0, 0.0, 0.0}  // Pad
				} }, torch::kFloat);



	
	auto w = torch::tensor({
			{
				{0.0, 1.0, 0.0, 0.0},
				{0.0, 2.0, 0.0, 0.0},
				{0.0, 3.0, 0.0, 0.0},
				{0.0, 4.0, 0.0, 0.0}
			}
		}, torch::kFloat);

	cout << "input\n" << x << endl;
	
	cout << "-------------SelfAttention--------------------\n"  << endl;
	auto x1 = x.squeeze();
	auto atten= SelfAttention();
	auto y = atten.forward(x1);
	cout << "-------------SelfAttention--------------------\n"  << endl;
	

	cout << "\n\n-------------MultiHeadAttention--------------------\n"  << endl;
	auto multiAtten = MultiHeadAttention();
	multiAtten.forward2(x,1);
	cout << "-------------MultiHeadAttention--------------------\n"  << endl;

}

感谢大家的支持，如要问题欢迎提问指正。

相关推荐
天天代码码天天5 分钟前
让普通照片“动”起来：用 C# + Vulkan 做一个照片深度与景深小工具
人工智能
QYR-分析8 分钟前
后置行车记录仪：8.6% 高增，车辆事故取证与全天候安防的后摄影像终端
人工智能
通信大模型17 分钟前
面向AI原生6G网络的可解释任务导向型Token通信
人工智能·计算机视觉
Magician_liu26 分钟前
机器学习基本概念
人工智能·机器学习
Industio_触觉智能27 分钟前
瑞芯微AI算力NPU路线图与协处理器｜触觉智能瑞芯微RK方案商出品
人工智能·嵌入式硬件·硬件架构·智能硬件·npu·瑞芯微·端侧ai
EchoMind-Henry32 分钟前
EviScene：识别出AI图，不等于查明真假
人工智能
打工仔折腾 AI35 分钟前
在线PDF工具隐私焦虑？把BentoPDF部署到极空间NAS本地处理
人工智能·后端·python·pdf·ai agent 实战
AI视频工具指南40 分钟前
AI短剧怎么做？一个矛盾怎样接着讲三集
人工智能·aigc·音视频
单向箔1 小时前
10｜Skill：把重复 prompt 固化成自己的 AI 工作流
人工智能·prompt·ai编程
没有梦想的咸鱼185-1037-16631 小时前
基金本子被批“科学问题不凝练”？试试这个“三问法”
人工智能·chatgpt·ai写作
热门推荐
01「全语录+省流版」邵艾伦 × 孙宇晨：聊风口、聊个人IP、聊年轻人如何抓住 AI 时代的机会02国内可直接用、免费额度/永久免费的大模型API清单（含 SiliconFlow、火山、阿里、智谱、百度、Kimi、DeepSeek、DMXAPI 等）03历年考研数学一、数学二、数学三真题试卷及答案PDF04GitHub 镜像站点05几个好用的ip纯净度检测网站06蜘蛛磁力 搜索引擎大全，如何使用蜘蛛磁力查找磁力链接07Windows 10/11 如何退出微软账户并改用本地账户登录（含找不到选项的解决方法）08Telegram APK 下载安装完整指南 — 2026年最新09微信历史版本含下载地址（ Windows PC | 安卓 | MAC ）及设置微信不更新10轻量化AI渗透武器库｜ARTEX自主攻防系统、资产梳理、漏洞挖掘、链路分析全流程智能化落地