任务
之前写了一个相似度任务的版本：bert 相似度任务训练简单版本,faiss 寻找相似 topk-CSDN博客
相似度用的是 0，1，相当于分类任务，现在我们相似度有评分，不再是 0,1 了，分数为 0-5，数字越大代表两个句子越相似，这一次的比较完整，评估，验证集，相似度模型都有了。
数据集
链接：https://pan.baidu.com/s/1B1-PKAKNoT_JwMYJx_zT1g
提取码：er1z
原始数据好几千条，我训练数据用了部分 2500 条，验证，测试 300 左右，使用 cpu 也用了好几个小时
train.py
import torch
import os
import time
from torch.utils.data import Dataset, DataLoader
from transformers import BertTokenizer, BertModel, AdamW, get_cosine_schedule_with_warmup
from sklearn.metrics.pairwise import cosine_similarity
import numpy as np
# 设备选择
# device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
device = 'cpu'
# 定义文本相似度数据集类
class TextSimilarityDataset(Dataset):
def __init__(self, file_path, tokenizer, max_len=128):
self.data = []
with open(file_path, 'r', encoding='utf-8') as f:
for line in f.readlines():
text1, text2, similarity_score = line.strip().split('\t')
inputs1 = tokenizer(text1, padding='max_length', truncation=True, max_length=max_len)
inputs2 = tokenizer(text2, padding='max_length', truncation=True, max_length=max_len)
self.data.append({
'input_ids1': inputs1['input_ids'],
'attention_mask1': inputs1['attention_mask'],
'input_ids2': inputs2['input_ids'],
'attention_mask2': inputs2['attention_mask'],
'similarity_score': float(similarity_score),
})
def __len__(self):
return len(self.data)
def __getitem__(self, idx):
return self.data[idx]
def cosine_similarity_torch(vec1, vec2, eps=1e-8):
dot_product = torch.mm(vec1, vec2.t())
norm1 = torch.norm(vec1, 2, dim=1, keepdim=True)
norm2 = torch.norm(vec2, 2, dim=1, keepdim=True)
similarity_scores = dot_product / (norm1 * norm2.t()).clamp(min=eps)
return similarity_scores
# 定义模型，这里我们不仅计算两段文本的[CLS] token的点积，而是整个句向量的余弦相似度
class BertSimilarityModel(torch.nn.Module):
def __init__(self, pretrained_model):
super(BertSimilarityModel, self).__init__()
self.bert = BertModel.from_pretrained(pretrained_model)
self.dropout = torch.nn.Dropout(p=0.1) # 引入Dropout层以防止过拟合
def forward(self, input_ids1, attention_mask1, input_ids2, attention_mask2):
embeddings1 = self.dropout(self.bert(input_ids=input_ids1, attention_mask=attention_mask1)['last_hidden_state'])
embeddings2 = self.dropout(self.bert(input_ids=input_ids2, attention_mask=attention_mask2)['last_hidden_state'])
# 计算两个文本向量的余弦相似度
embeddings1 = torch.mean(embeddings1, dim=1)
embeddings2 = torch.mean(embeddings2, dim=1)
similarity_scores = cosine_similarity_torch(embeddings1, embeddings2)
# 映射到[0, 5]评分范围
normalized_similarities = (similarity_scores + 1) * 2.5
return normalized_similarities.unsqueeze(1)
# 自定义损失函数，使用Smooth L1 Loss，更适合处理回归问题
class SmoothL1Loss(torch.nn.Module):
def __init__(self):
super(SmoothL1Loss, self).__init__()
def forward(self, predictions, targets):
diff = predictions - targets
abs_diff = torch.abs(diff)
quadratic = torch.where(abs_diff < 1, 0.5 * diff ** 2, abs_diff - 0.5)
return torch.mean(quadratic)
def train_model(model, train_loader, val_loader, epochs=3, model_save_path='../output/bert_similarity_model.pth'):
model.to(device)
criterion = SmoothL1Loss() # 使用自定义的Smooth L1 Loss
optimizer = AdamW(model.parameters(), lr=5e-5) # 调整初始学习率为5e-5
num_training_steps = len(train_loader) * epochs
scheduler = get_cosine_schedule_with_warmup(optimizer, num_warmup_steps=0.1*num_training_steps, num_training_steps=num_training_steps) # 使用带有warmup的余弦退火学习率调度
best_val_loss = float('inf')
for epoch in range(epochs):
model.train()
for batch in train_loader:
input_ids1 = batch['input_ids1'].to(device)
attention_mask1 = batch['attention_mask1'].to(device)
input_ids2 = batch['input_ids2'].to(device)
attention_mask2 = batch['attention_mask2'].to(device)
similarity_scores = batch['similarity_score'].to(device)
optimizer.zero_grad()
outputs = model(input_ids1, attention_mask1, input_ids2, attention_mask2)
loss = criterion(outputs, similarity_scores.unsqueeze(1))
loss.backward()
optimizer.step()
scheduler.step()
# 验证阶段
model.eval()
with torch.no_grad():
val_loss = 0
total_val_samples = 0
for batch in val_loader:
input_ids1 = batch['input_ids1'].to(device)
attention_mask1 = batch['attention_mask1'].to(device)
input_ids2 = batch['input_ids2'].to(device)
attention_mask2 = batch['attention_mask2'].to(device)
similarity_scores = batch['similarity_score'].to(device)
val_outputs = model(input_ids1, attention_mask1, input_ids2, attention_mask2)
val_loss += criterion(val_outputs, similarity_scores.unsqueeze(1)).item()
total_val_samples += len(similarity_scores)
val_loss /= len(val_loader)
print(f'Epoch {epoch + 1}, Validation Loss: {val_loss:.4f}')
if val_loss < best_val_loss:
best_val_loss = val_loss
torch.save(model.state_dict(), model_save_path)
def collate_to_tensors(batch):
'''把数据处理为模型可用的数据，不同任务可能需要修改一下，'''
input_ids1 = torch.tensor([example['input_ids1'] for example in batch])
attention_mask1 = torch.tensor([example['attention_mask1'] for example in batch])
input_ids2 = torch.tensor([example['input_ids2'] for example in batch])
attention_mask2 = torch.tensor([example['attention_mask2'] for example in batch])
similarity_score = torch.tensor([example['similarity_score'] for example in batch])
return {'input_ids1': input_ids1, 'attention_mask1': attention_mask1, 'input_ids2': input_ids2,
'attention_mask2': attention_mask2, 'similarity_score': similarity_score}
# 加载数据集和预训练模型
tokenizer = BertTokenizer.from_pretrained('../bert-base-chinese')
model = BertSimilarityModel('../bert-base-chinese')
# 加载数据并创建
train_data = TextSimilarityDataset('../data/STS-B/STS-B.train - 副本.data', tokenizer)
val_data = TextSimilarityDataset('../data/STS-B/STS-B.valid - 副本.data', tokenizer)
test_data = TextSimilarityDataset('../data/STS-B/STS-B.test - 副本.data', tokenizer)
train_loader = DataLoader(train_data, batch_size=32, shuffle=True, collate_fn=collate_to_tensors)
val_loader = DataLoader(val_data, batch_size=32, collate_fn=collate_to_tensors)
test_loader = DataLoader(test_data, batch_size=32, collate_fn=collate_to_tensors)
optimizer = AdamW(model.parameters(), lr=2e-5)
# 开始训练
train_model(model, train_loader, val_loader)
# 加载最佳模型进行测试
model.load_state_dict(torch.load('../output/bert_similarity_model.pth'))
test_loss = 0
total_test_samples = 0
with torch.no_grad():
for batch in test_loader:
input_ids1 = batch['input_ids1'].to(device)
attention_mask1 = batch['attention_mask1'].to(device)
input_ids2 = batch['input_ids2'].to(device)
attention_mask2 = batch['attention_mask2'].to(device)
similarity_scores = batch['similarity_score'].to(device)
test_outputs = model(input_ids1, attention_mask1, input_ids2, attention_mask2)
test_loss += torch.nn.functional.mse_loss(test_outputs, similarity_scores.unsqueeze(1)).item()
total_test_samples += len(similarity_scores)
test_loss /= len(test_loader)
print(f'Test Loss: {test_loss:.4f}')
predit.py
这个脚本是用来看看效果的，直接传入两个文本，使用训练好的模型来计算相似度的
import torch
from transformers import BertTokenizer, BertModel
def cosine_similarity_torch(vec1, vec2, eps=1e-8):
dot_product = torch.mm(vec1, vec2.t())
norm1 = torch.norm(vec1, 2, dim=1, keepdim=True)
norm2 = torch.norm(vec2, 2, dim=1, keepdim=True)
similarity_scores = dot_product / (norm1 * norm2.t()).clamp(min=eps)
return similarity_scores
# 定义模型，这里我们不仅计算两段文本的[CLS] token的点积，而是整个句向量的余弦相似度
class BertSimilarityModel(torch.nn.Module):
def __init__(self, pretrained_model):
super(BertSimilarityModel, self).__init__()
self.bert = BertModel.from_pretrained(pretrained_model)
self.dropout = torch.nn.Dropout(p=0.1) # 引入Dropout层以防止过拟合
def forward(self, input_ids1, attention_mask1, input_ids2, attention_mask2):
'''如果是用来预测，forward 会被禁用'''
pass
# 加载预训练模型和分词器
tokenizer = BertTokenizer.from_pretrained('../bert-base-chinese')
model = BertSimilarityModel('../bert-base-chinese')
model.load_state_dict(torch.load('../output/bert_similarity_model.pth')) # 请确保路径正确
model.eval() # 设置模型为评估模式
def calculate_similarity(text1, text2):
# 对输入文本进行编码
inputs1 = tokenizer(text1, padding='max_length', truncation=True, max_length=128, return_tensors='pt')
inputs2 = tokenizer(text2, padding='max_length', truncation=True, max_length=128, return_tensors='pt')
# 计算相似度
with torch.no_grad():
embeddings1 = model.bert(**inputs1.to('cpu'))['last_hidden_state'][:, 0]
embeddings2 = model.bert(**inputs2.to('cpu'))['last_hidden_state'][:, 0]
similarity_score = cosine_similarity_torch(embeddings1, embeddings2).item()
# 映射到[0, 5]评分范围（假设训练时有此步骤）
normalized_similarity = (similarity_score + 1) * 2.5
return normalized_similarity
# 示例
text1 = "瑞典驻利比亚班加西领事馆发生汽车炸弹袭击，无人员伤亡"
text2 = "汽车炸弹击中瑞典驻班加西领事馆，无人受伤。"
similarity = calculate_similarity(text1, text2)
print(f"两个句子的相似度为：{similarity:.2f}")
