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 = "jenniferhk008/roberta-hfl-emoji-aug3epoch"
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"
}

# ✅ 页面配置
st.set_page_config(page_title="Emoji Offensive Text Detector", page_icon="🚨", layout="wide")

# ✅ 页面标题
st.title("🧠 Emoji-based Offensive Language Classifier")

st.markdown("""
This application translates emojis in a sentence and classifies whether the final sentence is offensive or not using two AI models.
- The **first model** translates emoji or symbolic phrases into standard Chinese text.
- The **second model** performs offensive language detection.
""")

# 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)

# ✅ 输入区域
st.markdown("### ✍️ Input your sentence:")
default_text = "你是🐷"
text = st.text_area("Enter sentence with emojis:", value=default_text, height=150)

# ✅ 主逻辑封装函数
def classify_emoji_text(text: str):
    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)
    translated_text = decoded.split("输出：")[-1].strip() if "输出：" in decoded else decoded.strip()

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

    return translated_text, label, score

# ✅ 触发按钮
if st.button("🚦 Analyze"):
    with st.spinner("🔍 Processing..."):
        try:
            translated, label, score = classify_emoji_text(text)
            st.markdown("### 🔄 Translated sentence:")
            st.code(translated, language="text")

            st.markdown(f"### 🎯 Prediction: `{label}`")
            st.markdown(f"### 📊 Confidence Score: `{score:.2%}`")

        except Exception as e:
            st.error(f"❌ An error occurred during processing:\n\n{e}")
else:
    st.info("👈 Please input text and click the button to classify.")