app.py

import os
os.environ["HF_HOME"] = "/tmp/.cache/huggingface"
os.environ["MPLCONFIGDIR"] = "/tmp/.config/matplotlib"

import streamlit as st
import joblib
import pandas as pd
import shap
import matplotlib.pyplot as plt
import platform
from huggingface_hub import hf_hub_download
import matplotlib.font_manager as fm

# 自訂函數：嘗試找系統中可用的中文字型路徑
def find_chinese_font():
    for font_path in fm.findSystemFonts(fontext='ttf'):
        # 這裡可依你系統字型名關鍵字調整
        if ("NotoSans" in font_path and ("TC" in font_path or "TraditionalChinese" in font_path)) \
           or "STHeiti" in font_path or "Heiti" in font_path or "LiHei" in font_path:
            return font_path
    return None

chinese_font_path = find_chinese_font()
if chinese_font_path:
    chinese_font_prop = fm.FontProperties(fname=chinese_font_path)
    print(f"使用中文字型檔: {chinese_font_path}")
else:
    chinese_font_prop = None
    print("找不到適合的中文字型檔，請確認系統已安裝中文字型")

# matplotlib 全局字型設定（Fallback用）
if platform.system() == 'Windows':
    plt.rcParams['font.family'] = 'Microsoft JhengHei'
elif platform.system() == 'Darwin':  # macOS
    plt.rcParams['font.family'] = 'AppleGothic'
else:
    # Linux 預設用找到的字型名稱，沒找到就用 DejaVu Sans
    if chinese_font_prop:
        plt.rcParams['font.family'] = chinese_font_prop.get_name()
    else:
        plt.rcParams['font.family'] = 'DejaVu Sans'

plt.rcParams['axes.unicode_minus'] = False  # 負號用 ASCII 減號

@st.cache_resource(show_spinner=True)
def load_model_and_explainer():
    model_path = hf_hub_download(
        repo_id="jung-ming/Ocean-Meets-Forest",
        filename="rf_model_with_encoder.pkl",
        repo_type="model"
    )
    bundle = joblib.load(model_path)
    model = bundle["model"]
    le = bundle["label_encoder"]
    explainer = shap.TreeExplainer(model, feature_perturbation="interventional")
    return model, le, explainer

model, le, explainer = load_model_and_explainer()

ship_type_to_code = dict(zip(le.classes_, le.transform(le.classes_)))

st.title("🚢 台中港艘次預測系統")
st.markdown("請輸入以下資訊，模型將預測該月艘次數")

port_count = st.selectbox("航線組合數", list(range(1, 100)))
year = st.selectbox("年", [2020, 2021, 2022, 2023, 2024, 2025])
month = st.selectbox("月", list(range(1, 13)))
ship_type = st.selectbox("船舶種類", list(ship_type_to_code.keys()))

if st.button("🔮 開始預測"):
    ship_type_encoded = ship_type_to_code[ship_type]
    # input_df 用中文欄位名，符合模型訓練時格式
    input_df = pd.DataFrame({
        "航線組合數": [port_count],
        "年": [year],
        "月": [month],
        "船舶種類_編碼": [ship_type_encoded]
    })
    pred = model.predict(input_df)[0]
    st.success(f"預測結果：🚢 約為 {pred:.2f} 艘次")

    st.subheader("🧠 模型決策解釋圖（SHAP Waterfall plot）")

    shap_values = explainer(input_df)

    # 把SHAP特徵名稱轉成英文，但不改input_df欄位名稱
    feature_name_map = {
        "航線組合數": "Route Count",
        "年": "Year",
        "月": "Month",
        "船舶種類_編碼": "Ship Type Code"
    }
    shap_values.feature_names = [feature_name_map.get(f, f) for f in shap_values.feature_names]

    ax = shap.plots.waterfall(shap_values[0], show=False)

    # 負號替換（避免顯示問題）
    for text in ax.texts:
        if text.get_text().startswith('\u2212'):
            text.set_text(text.get_text().replace('\u2212', '-'))

    st.pyplot(ax.figure)
    plt.close(ax.figure)