app.py

import gradio as gr
from transformers import AutoConfig, AutoTokenizer, AutoModelForSequenceClassification
import torch

# 1. 设置模型路径
model_dir = "."

# 2. 加载模型
# 增加容错机制：如果加载训练好的模型失败，自动加载官方小模型作为兜底，防止网页打不开
try:
    config = AutoConfig.from_pretrained(model_dir)
    tokenizer = AutoTokenizer.from_pretrained(model_dir)
    model = AutoModelForSequenceClassification.from_pretrained(model_dir, config=config)
except Exception as e:
    print(f"Error loading model: {e}")
    print("Loading fallback model (bert-tiny) for debugging...")
    model_dir = "prajjwal1/bert-tiny" 
    config = AutoConfig.from_pretrained(model_dir)
    tokenizer = AutoTokenizer.from_pretrained(model_dir)
    model = AutoModelForSequenceClassification.from_pretrained(model_dir, config=config)

# 3. 定义推理函数
def inference(input_text):
    if not input_text:
        return "Please input some text."
        
    inputs = tokenizer.batch_encode_plus(
        [input_text],
        max_length=512,
        truncation=True,
        padding="max_length", 
        return_tensors="pt",
    )

    with torch.no_grad():
        logits = model(**inputs).logits

    predicted_class_id = logits.argmax().item()
    
    # 获取标签映射
    if hasattr(model.config, "id2label") and model.config.id2label:
        output = model.config.id2label[predicted_class_id]
    else:
        # 手动映射：根据你的训练数据，通常 0 是正常，1 是灾难
        labels = {0: "Normal (正常)", 1: "Disaster (灾难)"}
        output = labels.get(predicted_class_id, f"Class {predicted_class_id}")
        
    return output

# 4. 构建界面 (移除了 css 参数)
with gr.Blocks() as demo:
    gr.Markdown("### 灾难推文检测器 (Disaster Tweet Detector)")
    
    with gr.Row():
        with gr.Column():
            input_text = gr.Textbox(placeholder="Insert your prompt here:", label="Input Text")
            generate_bt = gr.Button("Generate", variant="primary")
        
        with gr.Column():
            answer = gr.Textbox(label="Prediction Result")

    # 绑定点击事件
    generate_bt.click(
        fn=inference, 
        inputs=[input_text], 
        outputs=[answer]
    )
    
    # 示例数据
    gr.Examples(
        examples=[
            ["The sky is blue and I am happy."],
            ["Huge fire in the downtown building! We need help!"]
        ],
        inputs=input_text
    )

if __name__ == "__main__":
    demo.queue()
    demo.launch()