Transformer源代码

Transformer是一种基于自注意力机制的神经网络模型，用于序列到序列的任务，如机器翻译、文本摘要等。以下是Transformer的PyTorch源代码：

import torch
import torch.nn as nn
import torch.nn.functional as F

class Transformer(nn.Module):
    def __init__(self, input_dim, output_dim, hidden_dim, num_layers, num_heads, dropout):
        super().__init__()
        self.embedding = nn.Embedding(input_dim, hidden_dim)
        self.pos_encoding = PositionalEncoding(hidden_dim, dropout)
        self.layers = nn.ModuleList([TransformerLayer(hidden_dim, num_heads, dropout) for _ in range(num_layers)])
        self.fc_out = nn.Linear(hidden_dim, output_dim)
        
    def forward(self, src):
        # src: [batch_size, src_len]
        embedded = self.embedding(src) * math.sqrt(self.hidden_dim)
        embedded = self.pos_encoding(embedded)
        for layer in self.layers:
            embedded = layer(embedded)
        output = self.fc_out(embedded[:, 0])
        return output

class TransformerLayer(nn.Module):
    def __init__(self, hidden_dim, num_heads, dropout):
        super().__init__()
        self.self_attn = MultiHeadAttention(hidden_dim, num_heads)
        self.norm1 = nn.LayerNorm(hidden_dim)
        self.ff = FeedForward(hidden_dim)
        self.norm2 = nn.LayerNorm(hidden_dim)
        self.dropout = nn.Dropout(dropout)
        
    def forward(self, src):
        # src: [batch_size, src_len, hidden_dim]
        attn_output, _ = self.self_attn(src, src, src)
        src = self.norm1(src + self.dropout(attn_output))
        ff_output = self.ff(src)
        src = self.norm2(src + self.dropout(ff_output))
        return src

class MultiHeadAttention(nn.Module):
    def __init__(self, hidden_dim, num_heads):
        super().__init__()
        self.hidden_dim = hidden_dim
        self.num_heads = num_heads
        self.head_dim = hidden_dim // num_heads
        self.fc_q = nn.Linear(hidden_dim, hidden_dim)
        self.fc_k = nn.Linear(hidden_dim, hidden_dim)
        self.fc_v = nn.Linear(hidden_dim, hidden_dim)
        self.fc_o = nn.Linear(hidden_dim, hidden_dim)
        
    def forward(self, query, key, value):
        # query: [batch_size, query_len, hidden_dim]
        # key: [batch_size, key_len, hidden_dim]
        # value: [batch_size, value_len, hidden_dim]
        batch_size = query.shape[0]
        Q = self.fc_q(query).view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2) # [batch_size, num_heads, query_len, head_dim]
        K = self.fc_k(key).view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2) # [batch_size, num_heads, key_len, head_dim]
        V = self.fc_v(value).view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2) # [batch_size, num_heads, value_len, head_dim]
        attn_scores = torch.matmul(Q, K.transpose(-2, -1)) / math.sqrt(self.head_dim) # [batch_size, num_heads, query_len, key_len]
        attn_probs = F.softmax(attn_scores, dim=-1)
        attn_output = torch.matmul(self.dropout(attn_probs), V) # [batch_size, num_heads, query_len, head_dim]
        attn_output = attn_output.transpose(1, 2).contiguous().view(batch_size, -1, self.hidden_dim) # [batch_size, query_len, hidden_dim]
        output = self.fc_o(attn_output)
        return output, attn_probs

class FeedForward(nn.Module):
    def __init__(self, hidden_dim):
        super().__init__()
        self.fc1 = nn.Linear(hidden_dim, hidden_dim * 4)
        self.fc2 = nn.Linear(hidden_dim * 4, hidden_dim)
        
    def forward(self, src):
        # src: [batch_size, src_len, hidden_dim]
        output = F.relu(self.fc1(src))
        output = self.fc2(output)
        return output

class PositionalEncoding(nn.Module):
    def __init__(self, hidden_dim, dropout, max_len=5000):
        super().__init__()
        self.dropout = nn.Dropout(dropout)
        pe = torch.zeros(max_len, hidden_dim)
        position = torch.arange(0, max_len, dtype=torch.float).unsqueeze(1)
        div_term = torch.exp(torch.arange(0, hidden_dim, 2).float() * (-math.log(10000.0) / hidden_dim))
        pe[:, 0::2] = torch.sin(position * div_term)
        pe[:, 1::2] = torch.cos(position * div_term)
        pe = pe.unsqueeze(0).transpose(0, 1)
        self.register_buffer('pe', pe)
        
    def forward(self, x):
        # x: [batch_size, seq_len, hidden_dim]
        x = x + self.pe[:x.size(1), :]
        return self.dropout(x)