app.py

from transformers import pipeline, AutoTokenizer, AutoModelForCausalLM
import torch
import streamlit as st
from PIL import Image
import pytesseract
import pandas as pd
import plotly.express as px

# ✅ 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"
}

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

# ✅ 侧边栏：模型选择
with st.sidebar:
    st.header("🧠 Settings")
    selected_model = st.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)

# 初始化会话历史
if "history" not in st.session_state:
    st.session_state.history = []


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"]
    reasoning = (
        f"The sentence was flagged as '{label}' due to potentially offensive phrases. "
        "Consider replacing emotionally charged, ambiguous, or abusive terms."
    )

    st.session_state.history.append({
        "text": text,
        "translated": translated_text,
        "label": label,
        "score": score,
        "reason": reasoning
    })
    return translated_text, label, score, reasoning

# 主页面布局
st.title("🚨 Emoji Offensive Text Detector & Analysis")
st.markdown("---")

# 输入与分析
st.header("✍️ Input & Moderation")
def text_moderation_section():
    st.markdown("Enter text with emojis or upload an image with text.")
    text = st.text_area("Sentence (or OCR text will appear here):", height=120)

    uploaded_file = st.file_uploader("Or upload an image for OCR:", type=["jpg", "jpeg", "png"])
    if uploaded_file:
        image = Image.open(uploaded_file)
        st.image(image, caption="Uploaded Image", use_column_width=True)
        with st.spinner("Extracting text via OCR..."):
            ocr_text = pytesseract.image_to_string(image, lang="chi_sim+eng").strip()
        st.text_area("Extracted Text:", value=ocr_text, height=120)
        text = ocr_text

    if st.button("🚦 Analyze Text") and text:
        with st.spinner("Processing..."):
            try:
                translated, label, score, reason = classify_emoji_text(text)
                st.subheader("🔄 Translated Text")
                st.code(translated)
                st.subheader(f"🎯 Prediction: {label}")
                st.write(f"Confidence: {score:.2%}")
                st.subheader("🧠 Explanation")
                st.info(reason)
            except Exception as e:
                st.error(f"Error during processing: {e}")

# 分析仪表板
st.markdown("---")
st.header("📊 Violation Analysis")
def analysis_dashboard():
    if not st.session_state.history:
        st.info("No data to display. Please analyze some text first.")
        return

    df = pd.DataFrame(st.session_state.history)

    # 建议列表
    st.subheader("📝 Offensive Terms & Suggestions")
    for item in st.session_state.history:
        st.markdown(f"- **Input:** {item['text']}")
        st.markdown(f"  - Translated: {item['translated']}")
        st.markdown(f"  - Label: {item['label']} ({item['score']:.2%})")
        st.markdown(f"  - Suggestion: {item['reason']}")

    # 雷达图
    radar_df = pd.DataFrame({
        "Category": ["Insult", "Abuse", "Discrimination", "Hate Speech", "Vulgarity"],
        "Score": [0.7, 0.4, 0.3, 0.5, 0.6]
    })
    radar_fig = px.line_polar(
        radar_df,
        r='Score',
        theta='Category',
        line_close=True,
        title="⚠️ Risk Radar by Category"
    )
    radar_fig.update_traces(line_color='black')
    st.plotly_chart(radar_fig)

# 渲染各部分
text_moderation_section()
analysis_dashboard()