赞
踩
如果说当下最火的AI技术和话题是什么,恐怕很难绕开ChatGPT。各大厂商都在表示未来要跟进ChatGPT技术,开发在自然语言处理智能系统,可见其影响力。本篇博客追个热度,来简单的介绍下ChatGPT到底是一项什么技术,究竟如何完成复杂的语言处理任务的。
如果想要了解ChatGPT是一项什么技术,不如直接问问他:
可以看到ChatGPT是一项基于Transformer模型实现的NLP软件。关于Transformer模型,我们已经在之前文章《注意力机制原理概述》介绍过。这里提到的LM技术,其实就是Transformer。
那么,ChatGPT的Transformer架构是什么样?我让chatgpt打印了他自己的主干网络代码,如下:
完整的代码分为四个模块:GPT框架,GPT层,多头注意力模块代码和位置编码。
GPT框架:
- import torch
- import torch.nn as nn
- import torch.nn.functional as F
-
- class GPT(nn.Module):
-
- def init(self, vocab_size, embed_dim, n_layers, n_heads, dropout):
- super(GPT, self).init()
- self.vocab_size = vocab_size
- self.embed_dim = embed_dim
- self.n_layers = n_layers
- self.n_heads = n_heads
- self.dropout = dropout
- self.embedding = nn.Embedding(vocab_size, embed_dim)
- self.pos_encoding = nn.Embedding(vocab_size, embed_dim)
- self.layers = nn.ModuleList([GPTLayer(embed_dim, n_heads, dropout) for _ in range(n_layers)])
- self.fc = nn.Linear(embed_dim, vocab_size)
-
- def forward(self, x):
- # Embedding
- x = self.embedding(x)
- # Position Encoding
- x = self.pos_encoding(x)
- # GPT Layers
- for layer in self.layers:
- x = layer(x)
- # Output
- x = self.fc(x)
- return x
GPT层代码:
- class GPTLayer(nn.Module):
- def init(self, embed_dim, n_heads, dropout):
- super(GPTLayer, self).init()
- self.embed_dim = embed_dim
- self.n_heads = n_heads
- self.dropout = dropout
- self.attention = MultiHeadAttention(embed_dim, n_heads, dropout)
- self.feed_forward = PositionwiseFeedForward(embed_dim, dropout)
-
- def forward(self, x):
- # Attention
- x = self.attention(x)
- # Feed Forward
- x = self.feed_forward(x)
- return x
多头注意力模块代码:
- class MultiHeadAttention(nn.Module):
- def init(self, embed_dim, n_heads, dropout):
- super(MultiHeadAttention, self).init()
- self.embed_dim = embed_dim
- self.n_heads = n_heads
- self.dropout = dropout
-
- self.query_linear = nn.Linear(embed_dim, embed_dim)
- self.key_linear = nn.Linear(embed_dim, embed_dim)
- self.value_linear = nn.Linear(embed_dim, embed_dim)
- self.output_linear = nn.Linear(embed_dim, embed_dim)
-
- def forward(self, x):
- # Query, Key, Value
- query = self.query_linear(x)
- key = self.key_linear(x)
- value = self.value_linear(x)
- # Attention
- scores = torch.matmul(query, key.transpose(-2, -1)) / math.sqrt(query.size(-1))
- scores = F.softmax(scores, dim=-1)
- scores = self.dropout(scores)
- # Output
- output = torch.matmul(scores, value)
- output = self.output_linear(output)
- return output
位置编码:
- class PositionwiseFeedForward(nn.Module):
- def init(self, embed_dim, dropout):
- super(PositionwiseFeedForward, self).init()
- self.embed_dim = embed_dim
- self.dropout = dropout
- self.fc1 = nn.Linear(embed_dim, embed_dim)
- self.fc2 = nn.Linear(embed_dim, embed_dim)
-
- def forward(self, x):
- # Feed Forward
- x = self.fc1(x)
- x = F.relu(x)
- x = self.dropout(x)
- x = self.fc2(x)
- x = self.dropout(x)
- return x
-
-
GPT主体框架主要由GPT层叠加构成。GPT层也比较容易理解,基本就是由多头注意力处理模块构建的。按照注意力机制原理,多头注意力处理首先将输入按照查询、键、值做对应的线性变换,之后输入一个多分枝的注意力结构,建立一个具有关联关系的评分结果。按照评分结果,实现对查询的值预测,实现NLP任务。这个模型基本就是2017年Vaswani工作 [1] 的复现。这让我们真正理解了多头注意力的强大之处。
chatgpt是一个有趣且有用的AI工具,对于泛NLP任务,具有目前最优秀的处理分析能力。我认为其在客户问答,信息查询,文字编辑等任务中,将产生深远的影响。其背后基于多头注意力机制的Transformer模型,已被验证在逻辑关系学习领域,具有惊人的技术优势。相信在该技术路线上,未来还会有更加优秀的工作被不断提出。
[1] A. Vaswani, N. Shazeer, N. Parmar, et al. Attention is all you need. Advances in neural information processing systems, 2017,5998‒6008.
Copyright © 2003-2013 www.wpsshop.cn 版权所有,并保留所有权利。