app.py

import gradio as gr
import torch
import torch.nn as nn
from transformers import BertTokenizer, BertModel

# device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
device = "cpu"
names = ['负向', '正向']

# 分词器
tokenizer = BertTokenizer.from_pretrained("bert-base-chinese")

class Model(nn.Module):
    def __init__(self, bert_model):
        super().__init__()
        self.bert = bert_model
        # 全连接，模型输入为768，分类为2
        self.fc = nn.Linear(768, 2)

    #
    def forward(self, input_ids, attention_mask, token_type_ids):
        # 使用预训练模型提取特征, 上游任务不参与训练，锁定权重
        with torch.no_grad():
            # Correctly call the BertModel instance stored in self.bert
            output = self.bert(input_ids, attention_mask, token_type_ids)
        # 下游参与训练，二分类任务，获取最新后的状态
        output = self.fc(output.last_hidden_state[:, 0])
        # softmax激活函数，NV结构，获取特征值dim维度为1
        output = output.softmax(dim=1)
        return output

# 加载预训练模型
bert_model = BertModel.from_pretrained("ckiplab/bert-base-chinese").to(device)
model = Model(bert_model).to(device)
model.load_state_dict(torch.load("params/1bert.pt", map_location=torch.device(device)))
# 切换到eval模式
model.eval()

def collate_fn(data):
    sentes = []
    sentes.append(data)

    #编码
    data = tokenizer.batch_encode_plus(
        batch_text_or_text_pairs=sentes,
        truncation=True,
        padding="max_length",
        max_length=350,
        return_tensors="pt",
        return_length=True
    )
    input_ids = data["input_ids"]
    attention_mask = data["attention_mask"]
    token_type_ids = data["token_type_ids"]
    return input_ids, attention_mask, token_type_ids


def analyze_sentiment(text):
    input_ids, attention_mask, token_type_ids = collate_fn(text)
    input_ids = input_ids.to(device)
    attention_mask = attention_mask.to(device)
    token_type_ids = token_type_ids.to(device)
    # 上游不参与训练
    with torch.no_grad():
        out = model(input_ids, attention_mask, token_type_ids)
    # 找到每个样本在指定维度上的最大值的索引
    out = out.argmax(dim=1)
    return f"{names[out]}评价", names[out]

def create_interface():
    """
    创建Gradio界面
    """
    with gr.Blocks(title="情感分析应用", theme=gr.themes.Soft()) as demo:
        gr.Markdown("# 🎭 情感分析应用")
        gr.Markdown("输入文本，AI将分析其情感倾向")
        
        with gr.Row():
            with gr.Column(scale=2):
                # 输入区域
                text_input = gr.Textbox(
                    label="输入要分析的文本",
                    placeholder="请输入您想要分析情感的文本...",
                    lines=4,
                    max_lines=10
                )
                
                # 按钮
                with gr.Row():
                    analyze_btn = gr.Button("🔍 分析情感", variant="primary")
                    clear_btn = gr.Button("🗑️ 清空", variant="secondary")
            
            with gr.Column(scale=2):
                # 输出区域
                result_summary = gr.Textbox(
                    label="分析结果",
                    lines=3,
                    interactive=False
                )
                
                # 情感标签显示
                sentiment_label = gr.Label(
                    label="情感分类",
                )
        
        # 示例文本
        gr.Markdown("### 📝 示例文本")
        examples = gr.Examples(
            examples=[
                ["这个产品真的很棒，我非常满意！"],
                ["服务态度太差了，完全不推荐"],
                ["还可以吧，没什么特别的感觉"],
                ["质量很好，物流也很快，五星好评！"],
                ["价格太贵了，性价比不高"]
            ],
            inputs=text_input,
            outputs=[result_summary, sentiment_label],
            fn=analyze_sentiment,
            cache_examples=False
        )
        
        # 绑定事件
        analyze_btn.click(
            fn=analyze_sentiment,
            inputs=text_input,
            outputs=[result_summary, sentiment_label]
        )
        
        clear_btn.click(
            fn=lambda: ("", "", ""),
            outputs=[text_input, result_summary, sentiment_label]
        )
        
        # 回车键触发分析
        text_input.submit(
            fn=analyze_sentiment,
            inputs=text_input,
            outputs=[result_summary, sentiment_label]
        )
    
    return demo

if __name__ == "__main__":
    demo = create_interface()
    demo.launch(debug=True)