bert ranking listwise demo

下面是用bert 训练listwise rank 的 demo 

import torch
from torch.utils.data import DataLoader, Dataset
from transformers import BertModel, BertTokenizer
from sklearn.metrics import pairwise_distances_argmin_min

class ListwiseRankingDataset(Dataset):
    def __init__(self, queries, documents, labels, tokenizer, max_length):
        self.input_ids = []
        self.attention_masks = []
        self.labels = []
        
        for query, doc_list, label_list in zip(queries, documents, labels):
            for doc, label in zip(doc_list, label_list):
                encoded_pair = tokenizer(query, doc, padding='max_length', truncation=True, max_length=max_length, return_tensors='pt')
                self.input_ids.append(encoded_pair['input_ids'])
                self.attention_masks.append(encoded_pair['attention_mask'])
                self.labels.append(label)
        
        self.input_ids = torch.cat(self.input_ids, dim=0)
        self.attention_masks = torch.cat(self.attention_masks, dim=0)
        self.labels = torch.tensor(self.labels)
        
    def __len__(self):
        return len(self.input_ids)
    
    def __getitem__(self, idx):
        input_id = self.input_ids[idx]
        attention_mask = self.attention_masks[idx]
        label = self.labels[idx]
        return input_id, attention_mask, label

class BERTListwiseRankingModel(torch.nn.Module):
    def __init__(self, bert_model_name):
        super(BERTListwiseRankingModel, self).__init__()
        self.bert = BertModel.from_pretrained(bert_model_name)
        self.dropout = torch.nn.Dropout(0.1)
        self.fc = torch.nn.Linear(self.bert.config.hidden_size, 1)
        
    def forward(self, input_ids, attention_mask):
        outputs = self.bert(input_ids=input_ids, attention_mask=attention_mask)
        pooled_output = self.dropout(outputs[1])
        logits = self.fc(pooled_output)
        return logits.squeeze()

# 初始化BERT模型和分词器
bert_model_name = 'bert-base-uncased'
tokenizer = BertTokenizer.from_pretrained(bert_model_name)

# 示例输入数据
queries = ['I like cats', 'The sun is shining']
documents = [['I like dogs', 'Dogs are cute'], ['It is raining', 'Rainy weather is gloomy']]
labels = [[1, 0], [0, 1]]

# 超参数
batch_size = 8
max_length = 128
learning_rate = 1e-5
num_epochs = 5

# 创建数据集和数据加载器
dataset = ListwiseRankingDataset(queries, documents, labels, tokenizer, max_length)
dataloader = DataLoader(dataset, batch_size=batch_size, shuffle=True)

# 初始化模型并加载预训练权重
model = BERTListwiseRankingModel(bert_model_name)
optimizer = torch.optim.AdamW(model.parameters(), lr=learning_rate)

# 训练模型
model.train()

for epoch in range(num_epochs):
    total_loss = 0
    
    for input_ids, attention_masks, labels in dataloader:
        optimizer.zero_grad()
        
        logits = model(input_ids, attention_masks)
        
        # 计算损失函数(使用交叉熵损失函数)
        loss = torch.nn.functional.binary_cross_entropy_with_logits(logits, labels.float())
        
        total_loss += loss.item()
        
        loss.backward()
        optimizer.step()
    
    print(f"Epoch {epoch+1}/{num_epochs} - Loss: {total_loss:.4f}")

# 推断模型
model.eval()

with torch.no_grad():
    embeddings = model.bert.embeddings.word_embeddings(dataset.input_ids)
    pairwise_distances = pairwise_distances_argmin_min(embeddings.numpy())

# 输出结果
for i, query in enumerate(queries):
    print(f"Query: {query}")
    print("Documents:")
    
    for j, doc in enumerate(documents[i]):
        doc_idx = pairwise_distances[0][i * len(documents[i]) + j]
        doc_dist = pairwise_distances[1][i * len(documents[i]) + j]
        
        print(f"Document index: {doc_idx}, Distance: {doc_dist:.4f}")
        print(f"Document: {doc}")
        print("")

    print("---------")

本图文内容来源于网友网络收集整理提供,作为学习参考使用,版权属于原作者。
THE END
分享
二维码
< <上一篇
下一篇>>