构建一个完整的BERT模型并进行训练是一个复杂且耗时的任务。BERT模型由多个组件组成,包括嵌入层、Transformer编码器和分类器等。编写这些组件的完整代码超出了文本的范围。然而,一个基本的BERT模型框架以便了解其结构和主要组件的设置。
import torch
import torch.nn as nn
# BERT Model
class BERTModel(nn.Module):
def __init__(self, vocab_size, embedding_dim, hidden_dim, num_layers, num_heads, max_seq_length, num_classes):
super(BERTModel, self).__init__()
self.embedding = nn.Embedding(vocab_size, embedding_dim)
self.position_embedding = nn.Embedding(max_seq_length, embedding_dim)
self.transformer_blocks = nn.ModuleList([
TransformerBlock(embedding_dim, hidden_dim, num_heads)
for _ in range(num_layers)
])
self.classifier = nn.Linear(embedding_dim, num_classes)
self.dropout = nn.Dropout(p=0.1)
def forward(self, input_ids, attention_mask):
embedded = self.embedding(input_ids) # [batch_size, seq_length, embedding_dim]
positions = torch.arange(0, input_ids.size(1), device=input_ids.device).unsqueeze(0).expand_as(input_ids)
position_embedded = self.position_embedding(positions) # [batch_size, seq_length, embedding_dim]
encoded = self.dropout(embedded + position_embedded) # [batch_size, seq_length, embedding_dim]
for transformer_block in self.transformer_blocks:
encoded = transformer_block(encoded, attention_mask)
pooled_output = encoded[:, 0, :] # [batch_size, embedding_dim]
logits = self.classifier(pooled_output) # [batch_size, num_classes]
return logits
# Transformer Block
class TransformerBlock(nn.Module):
def __init__(self, embedding_dim, hidden_dim, num_heads):
super(TransformerBlock, self).__init__()
self.attention = MultiHeadAttention(embedding_dim, num_heads)
self.feed_forward = FeedForward(hidden_dim, embedding_dim)
self.layer_norm1 = nn.LayerNorm(embedding_dim)
self.layer_norm2 = nn.LayerNorm(embedding_dim)
def forward(self, x, attention_mask):
attended = self.attention(x, x, x, attention_mask) # [batch_size, seq_length, embedding_dim]
residual1 = x + attended
normalized1 = self.layer_norm1(residual1) # [batch_size, seq_length, embedding_dim]
fed_forward = self.feed_forward(normalized1) # [batch_size, seq_length, embedding_dim]
residual2 = normalized1 + fed_forward
normalized2 = self.layer_norm2(residual2) # [batch_size, seq_length, embedding_dim]
return normalized2
# Multi-Head Attention
class MultiHeadAttention(nn.Module):
def __init__(self, embedding_dim, num_heads):
super(MultiHeadAttention, self).__init__()
self.num_heads = num_heads
self.head_dim = embedding_dim // num_heads
self.q_linear = nn.Linear(embedding_dim, embedding_dim)
self.k_linear = nn.Linear(embedding_dim, embedding_dim)
self.v_linear = nn.Linear(embedding_dim, embedding_dim)
self.out_linear = nn.Linear(embedding_dim, embedding_dim)
def forward(self, query, key, value, mask=None):
batch_size = query.size(0)
query = self.q_linear(query) # [batch_size, seq_length, embedding_dim]
key = self.k_linear(key) # [batch_size, seq_length, embedding_dim]
value = self.v_linear(value) # [batch_size, seq_length, embedding_dim]
query = self._split_heads(query) # [batch_size, num_heads, seq_length, head_dim]
key = self._split_heads(key) # [batch_size, num_heads, seq_length, head_dim]
value = self._split_heads(value) # [batch_size, num_heads, seq_length, head_dim]
scores = torch.matmul(query, key.transpose(-1, -2)) # [batch_size, num_heads, seq_length, seq_length]
scores = scores / torch.sqrt(torch.tensor(self.head_dim, dtype=torch.float32, device=scores.device))
if mask is not None:
scores = scores.masked_fill(mask.unsqueeze(1).unsqueeze(2), -1e9)
attention_outputs = torch.softmax(scores, dim=-1) # [batch_size, num_heads, seq_length, seq_length]
attention_outputs = self.dropout(attention_outputs)
attended = torch.matmul(attention_outputs, value) # [batch_size, num_heads, seq_length, head_dim]
attended = attended.transpose(1, 2).contiguous() # [batch_size, seq_length, num_heads, head_dim]
attended = attended.view(batch_size, -1, self.embedding_dim) # [batch_size, seq_length, embedding_dim]
attended = self.out_linear(attended) # [batch_size, seq_length, embedding_dim]
return attended
def _split_heads(self, x):
batch_size, seq_length, embedding_dim = x.size()
x = x.view(batch_size, seq_length, self.num_heads, self.head_dim)
x = x.transpose(1, 2).contiguous()
return x
# Feed Forward
class FeedForward(nn.Module):
def __init__(self, hidden_dim, embedding_dim):
super(FeedForward, self).__init__()
self.linear1 = nn.Linear(embedding_dim, hidden_dim)
self.activation = nn.ReLU()
self.dropout = nn.Dropout(p=0.1)
self.linear2 = nn.Linear(hidden_dim, embedding_dim)
def forward(self, x):
x = self.linear1(x) # [batch_size, seq_length, hidden_dim]
x = self.activation(x)
x = self.dropout(x)
x = self.linear2(x) # [batch_size, seq_length, embedding_dim]
return x
# Example usage
vocab_size = 10000
embedding_dim = 300
hidden_dim = 768
num_layers = 12
num_heads = 12
max_seq_length = 512
num_classes = 2
model = BERTModel(vocab_size, embedding_dim, hidden_dim, num_layers, num_heads, max_seq_length, num_classes)
input_ids = torch.tensor([[1, 2, 3, 4, 5]]).long()
attention_mask = torch.tensor([[1, 1, 1, 1, 1]]).long()
logits = model(input_ids, attention_mask)
print(logits.shape) # [1, num_classes]
这段代码给出了一个基本的BERT模型结构,并包含了Transformer块、注意力机制和前馈神经网络等组件。您需要根据自己的需求和数据集来调整参数和模型结构。
请注意,这只是一个简化的版本,真实的BERT模型还包括Masked Language Modeling(MLM)和Next Sentence Prediction(NSP)等预训练任务。此外,还需要进行数据预处理、损失函数的定义和训练循环等。在实际环境中,强烈建议使用已经经过大规模预训练的BERT模型,如Hugging Face的transformers库中的预训练模型,以获得更好的性能效果。