inference.py

# inference.py （完整手动加载版本，绕过所有 Auto 机制）

import torch
from modeling_tinytransformer import TinyTransformerModel   # 直接导入你的模型类
from tokenization_tinytransformer import TinyTokenizer      # 你的 tokenizer
from configuration_tinytransformer import TinyTransformerConfig  # 你的 config 类

model_path = "./tiny-sentiment-model"
device = "cuda" if torch.cuda.is_available() else "cpu"

print("加载 tokenizer（手动）...")
tokenizer = TinyTokenizer()

print("加载 config（手动）...")
config = TinyTransformerConfig.from_pretrained(model_path)  # 用你的 config 类加载

print("加载模型（手动）...")
model = TinyTransformerModel.from_pretrained(model_path, config=config)
model.to(device)
model.eval()

print("模型加载完成！设备：", device)

# ----------------------- 预测函数（保持不变） -----------------------
def predict(text: str):
    inputs = tokenizer(
        text,
        padding=True,
        truncation=True,
        max_length=64,
        return_tensors="pt"
    )
    inputs = {k: v.to(device) for k, v in inputs.items()}
    
    with torch.no_grad():
        outputs = model(**inputs)
    
    logits = outputs["logits"] if isinstance(outputs, dict) else outputs.logits
    probs = torch.softmax(logits, dim=-1)[0]
    
    pred_label = torch.argmax(probs).item()
    confidence = probs[pred_label].item()
    
    label_map = {0: "负面", 1: "正面"}
    
    print(f"文本: {text}")
    print(f"预测: {label_map[pred_label]} (置信度: {confidence:.4f})")
    print(f"负面概率: {probs[0]:.4f} | 正面概率: {probs[1]:.4f}")
    print("-"*50)

# 测试
if __name__ == "__main__":
    test_texts = [
        "质量超级好，强烈推荐",
        "服务差劲，东西还贵",
        "一般般，没什么惊喜",
        "老板人好，下次还来"
    ]
    for text in test_texts:
        predict(text)