app.py

from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
import torch
import streamlit as st

# ✅ Step 1: Emoji 翻译模型（你自己训练的模型）
emoji_model_id = "JenniferHJF/qwen1.5-emoji-finetuned"
emoji_tokenizer = AutoTokenizer.from_pretrained(emoji_model_id, trust_remote_code=True)
emoji_model = AutoModelForCausalLM.from_pretrained(
    emoji_model_id,
    trust_remote_code=True,
    torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32
).to("cuda" if torch.cuda.is_available() else "cpu")
emoji_model.eval()

# ✅ Step 2: 可选择的冒犯性文本识别模型
model_options = {
    "Toxic-BERT": "unitary/toxic-bert",
    "Roberta Offensive": "cardiffnlp/twitter-roberta-base-offensive",
    "BERT Emotion": "bhadresh-savani/bert-base-go-emotion"
}

# Streamlit 侧边栏模型选择
selected_model = st.sidebar.selectbox("Choose classification model", list(model_options.keys()))
selected_model_id = model_options[selected_model]

classifier = pipeline("text-classification", model=selected_model_id, device=0 if torch.cuda.is_available() else -1)

def classify_emoji_text(text: str):
    """
    Step 1: 翻译文本中的 emoji
    Step 2: 使用分类器判断是否冒犯
    """
    prompt = f"输入：{text}\n输出："
    input_ids = emoji_tokenizer(prompt, return_tensors="pt").to(emoji_model.device)
    with torch.no_grad():
        output_ids = emoji_model.generate(**input_ids, max_new_tokens=64, do_sample=False)
    decoded = emoji_tokenizer.decode(output_ids[0], skip_special_tokens=True)

    # 保留真正输出部分（移除 prompt）
    if "输出：" in decoded:
        translated_text = decoded.split("输出：")[-1].strip()
    else:
        translated_text = decoded.strip()

    result = classifier(translated_text)[0]
    label = result["label"]
    score = result["score"]

    return translated_text, label, score