电商C端导购的模型解释：AI大语言模型的可解释性

    import lime
    from lime.lime_text import LimeTextExplainer
    from transformers import BertTokenizer, BertForSequenceClassification
    import torch
    
    # 加载模型和分词器
    model = BertForSequenceClassification.from_pretrained("bert-base-uncased")
    tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
    
    # 定义预测函数
    def predict(texts):
    inputs = tokenizer(texts, return_tensors="pt", padding=True, truncation=True)
    outputs = model(**inputs)
    probabilities = torch.softmax(outputs.logits, dim=-1)
    return probabilities.detach().numpy()
    
    # 创建LIME解释器
    explainer = LimeTextExplainer(class_names=["negative", "positive"])
    
    # 分析特征重要性
    text = "This product is amazing!"
    explanation = explainer.explain_instance(text, predict, num_features=10)
    
    # 打印结果
    explanation.show_in_notebook()
    
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
    
    代码解读

    from transformers import BertTokenizer, BertModel
    import torch
    import matplotlib.pyplot as plt
    
    # 加载模型和分词器
    model = BertModel.from_pretrained("bert-base-uncased", output_attentions=True)
    tokenizer = BertTokenizer.from_pretrained("bert-base-uncased")
    
    # 获取自注意力权重
    text = "This product is amazing!"
    inputs = tokenizer(text, return_tensors="pt")
    outputs = model(**inputs)
    attentions = outputs.attentions
    
    # 可视化权重
    def plot_attention(attentions, layer, head):
    attention = attentions[layer][0, head].detach().numpy()
    plt.imshow(attention, cmap="viridis")
    plt.colorbar()
    plt.show()
    
    plot_attention(attentions, layer=0, head=0)
    
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
    
    代码解读

    from transformers import GPT2Tokenizer, GPT2LMHeadModel
    import torch
    from torch.utils.data import DataLoader, Dataset
    from torch.optim import Adam
    
    # 加载模型和分词器
    model = GPT2LMHeadModel.from_pretrained("gpt2")
    tokenizer = GPT2Tokenizer.from_pretrained("gpt2")
    
    # 定义数据集
    class ExplanationDataset(Dataset):
    def __init__(self, data):
        self.data = data
    
    def __len__(self):
        return len(self.data)
    
    def __getitem__(self, idx):
        return self.data[idx]
    
    # 创建数据加载器
    data = [("This product is amazing!", "The model thinks the product is of high quality.")]
    dataset = ExplanationDataset(data)
    dataloader = DataLoader(dataset, batch_size=1, shuffle=True)
    
    # 训练解释生成器
    optimizer = Adam(model.parameters(), lr=1e-4)
    loss_function = torch.nn.CrossEntropyLoss()
    
    for epoch in range(10):
    for text, explanation in dataloader:
        inputs = tokenizer(text, return_tensors="pt")
        targets = tokenizer(explanation, return_tensors="pt")["input_ids"]
    
        optimizer.zero_grad()
        outputs = model(**inputs)
        logits = outputs.logits[:, :-1]
        loss = loss_function(logits, targets)
        loss.backward()
        optimizer.step()
    
        print(f"Epoch {epoch}, Loss: {loss.item()}")
    
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
      
    
    代码解读