app.py

# -*- coding: utf-8 -*-
"""
Hugging Face Spaces / Gradio app
步驟導引：資料 → 切分/訓練 → 評估(Accuracy+混淆矩陣) → 測試介面
資料欄位需為: text, label (label 為 0/1)
"""

import io
import numpy as np
import pandas as pd
import matplotlib.pyplot as plt
import re

import gradio as gr
from sklearn.model_selection import train_test_split
from sklearn.pipeline import Pipeline
from sklearn.feature_extraction.text import CountVectorizer, TfidfTransformer
from sklearn.linear_model import LogisticRegression
from sklearn.metrics import accuracy_score, classification_report, confusion_matrix


# -------------------------
# 內建示範資料（30 筆）
# -------------------------
SAMPLE_DATA = [
    ("這家餐廳超好吃，我下次還要來！", 1),
    ("服務人員很親切，體驗很棒", 1),
    ("環境乾淨、氣氛舒服，推！", 1),
    ("口味不錯，份量也剛好", 1),
    ("真的超難吃，後悔來了", 0),
    ("太鹹又冷掉，失望", 0),
    ("等很久，服務態度也不好", 0),
    ("分量很少，完全不值這個價錢", 0),
    ("便宜又好吃，CP值很高", 1),
    ("味道一般般，下次可能不會再來", 0),
    ("甜點驚艷！好吃到想哭", 1),
    ("餐點很油膩，吃完不舒服", 0),
    ("飲料清爽解膩，搭配主餐剛好", 1),
    ("今天的餐點都冷掉了…", 0),
    ("出餐速度快，餐點擺盤也很漂亮", 1),
    ("座位太擁擠，聊天很吵", 0),
    ("牛肉很嫩、湯頭很香", 1),
    ("海鮮新鮮沒有腥味，值得再訪", 1),
    ("價格偏高，內容卻普通", 0),
    ("外帶包裝用心，回家吃也很好吃", 1),
    ("收銀動線很亂，結帳排超久", 0),
    ("主廚特餐驚喜連連，口味層次豐富", 1),
    ("烤物有點焦味，失望", 0),
    ("店內油煙味太重", 0),
    ("小菜與主餐搭配出色，份量足", 1),
    ("白飯硬到像沒熟", 0),
    ("服務貼心會主動加水與關心口味", 1),
    ("臨時加點又等了二十分鐘", 0),
    ("週年限定套餐超值，朋友都喜歡", 1),
    ("衛生紙與餐具不足，需要一直跟店員拿", 0),
]
SAMPLE_DF = pd.DataFrame(SAMPLE_DATA, columns=["text", "label"])


# -------------------------
# 工具函式（仍保留，雖目前 UI 不再使用上傳/貼上）
# -------------------------
def read_csv_file(file_obj) -> pd.DataFrame:
    return pd.read_csv(file_obj, encoding_errors="ignore")

def read_csv_text(text_block: str) -> pd.DataFrame:
    return pd.read_csv(io.StringIO(text_block))

def check_df(df: pd.DataFrame) -> str:
    cols = set(df.columns.str.lower())
    need = {"text", "label"}
    if not need.issubset(cols):
        return "❌ CSV 需要兩個欄位：text, label（0/1）。"
    if df.empty:
        return "❌ 資料為空。"
    return ""


def normalize(text: str) -> str:
    t = text
    t = re.sub(r"(太|超|非常)?好吃(死了|到爆|極了)?", " POS_GOOD ", t)
    t = re.sub(r"(太|超|非常)?難吃(死了|到爆|極了)?", " NEG_BAD ", t)
    t = re.sub(r"(不|沒|沒有)", " NOT ", t)
    return t


def build_pipeline() -> Pipeline:
    """字 n-gram + TF-IDF + Logistic Regression"""
    return Pipeline([
        ("vect", CountVectorizer(preprocessor=normalize, analyzer="char", ngram_range=(2, 5))),
        ("tfidf", TfidfTransformer()),
        ("clf", LogisticRegression(max_iter=200, class_weight="balanced", C=0.8, solver="liblinear")),
    ])


def plot_confusion_matrix(y_true, y_pred, labels=("負面(0)", "正面(1)")):
    cm = confusion_matrix(y_true, y_pred)
    fig, ax = plt.subplots(figsize=(4, 3))
    im = ax.imshow(cm, cmap="Blues")
    ax.set_title("Confusion Matrix - Logistic Regression")
    ax.set_xlabel("Predicted")
    ax.set_ylabel("True")
    ax.set_xticks([0, 1]); ax.set_yticks([0, 1])
    ax.set_xticklabels(labels); ax.set_yticklabels(labels)
    for i in range(cm.shape[0]):
        for j in range(cm.shape[1]):
            ax.text(j, i, cm[i, j], ha="center", va="center", color="black")
    fig.colorbar(im, ax=ax, fraction=0.046, pad=0.04)
    fig.tight_layout()
    return fig


# -------------------------
# Step1：載入／追加
# -------------------------
def use_sample_dataset():
    df = SAMPLE_DF.copy()
    msg = f"✅ 已載入內建示範資料，共 {len(df)} 筆。"
    return df, df.iloc[::-1], msg 

def append_example(df, new_text, new_label):
    """把一筆 (text, label) 追加到現有資料集"""
    if df is None or not isinstance(df, pd.DataFrame) or df.empty:
        df = SAMPLE_DF.copy()  # 保險：尚未載入時以內建資料起始
    if not new_text or not str(new_text).strip():
        return df, df.iloc[::-1], "❌ 請先輸入文字。"
    try:
        y = int(new_label)
        if y not in (0, 1):
            raise ValueError
    except Exception:
        return df, df.iloc[::-1], "❌ 標籤須為 0 或 1。"

    new_row = pd.DataFrame([[str(new_text).strip(), y]], columns=["text", "label"])
    df2 = pd.concat([df, new_row], ignore_index=True)
    msg = f"✅ 已加入 1 筆，目前共 {len(df2)} 筆。"
    return df2, df2.iloc[::-1],msg


# -------------------------
# Step2：訓練與評估
# -------------------------
def train_and_eval(df, test_ratio):
    if df is None:
        return None, "❌ 請先載入資料。", None, None

    df = df.dropna(subset=["text", "label"]).copy()
    df["label"] = df["label"].astype(int)

    X_train, X_test, y_train, y_test = train_test_split(
        df["text"], df["label"],
        test_size=test_ratio, random_state=42, stratify=df["label"]
    )

    model = build_pipeline()
    model.fit(X_train, y_train)

    y_pred = model.predict(X_test)
    acc = accuracy_score(y_test, y_pred)
    rpt = classification_report(y_test, y_pred, digits=3)
    rpt_text = f"Accuracy: {acc:.3f}\n\n{rpt}"

    fig = plot_confusion_matrix(y_test, y_pred)

    dist = df["label"].value_counts().sort_index().to_dict()
    tip = f"資料分布：0→{dist.get(0,0)} , 1→{dist.get(1,0)}"

    return model, rpt_text + "\n" + tip, fig, "✅ 訓練完成，可以到下一步做測試。"


# -------------------------
# Step3：單句推論
# -------------------------
def predict_one(text, model):
    if model is None:
        return "❌ 尚未訓練模型。"
    if not text or not text.strip():
        return "請先輸入文字。"
    proba = model.predict_proba([text])[0][1]
    label = "正面(1)" if proba >= 0.5 else "負面(0)"
    return f"{label}（信心 {proba:.2f}）"


# -------------------------
# 介面
# -------------------------
with gr.Blocks(title="情感分類小幫手（LogReg + 字 n-gram）", theme=gr.themes.Soft()) as demo:
    gr.Markdown("## 情感分類小幫手（LogReg + 字 n-gram）\n一步一步完成：資料 → 訓練 → 評估 → 測試")

    # 讓學生進入頁面就有 30 筆資料
    state_df = gr.State(SAMPLE_DF.copy())
    state_model = gr.State()

    with gr.Tab("Step 1｜建立/擴充資料集"):
        gr.Markdown(
            "本活動不需上傳檔案，**直接用內建 30 筆資料**，並可在下方 **新增一筆 (text,label)** 進行擴充。\n"
            "- label：0=負面、1=正面\n"
        )

        # 操作列：載入預設資料 + 新增一筆
        with gr.Row():
            btn_use_sample = gr.Button("重新載入內建示範資料（30筆）")
        with gr.Row():
            add_text = gr.Textbox(label="新增一筆文字", placeholder="例：上菜超快，服務又好！", lines=2)
            add_label = gr.Radio(choices=[0, 1], value=1, label="標籤 (0=負面, 1=正面)")
            btn_add = gr.Button("加入到資料集")

        msg1 = gr.Markdown()
        df_preview = gr.Dataframe(
            headers=["text", "label"],
            label="資料預覽（全部資料）",
            wrap=True,
            interactive=False,
            value=SAMPLE_DF.iloc[::-1]  
        )

        # 綁定事件
        btn_use_sample.click(fn=use_sample_dataset, outputs=[state_df, df_preview, msg1])
        btn_add.click(fn=append_example, inputs=[state_df, add_text, add_label], outputs=[state_df, df_preview, msg1])

    with gr.Tab("Step 2｜切分與訓練"):
        gr.Markdown("設定測試集比例，按下 **開始訓練**。")
        split = gr.Slider(0.1, 0.5, value=0.3, step=0.05, label="測試集比例 test_size")
        btn_train = gr.Button("開始訓練")
        train_msg = gr.Markdown()
        report_box = gr.Textbox(label="評估報告（Accuracy / Precision / Recall / F1）", lines=12)
        cm_plot = gr.Plot(label="混淆矩陣")

        btn_train.click(
            fn=train_and_eval,
            inputs=[state_df, split],
            outputs=[state_model, report_box, cm_plot, train_msg]
        )

    with gr.Tab("Step 3｜測試/推論"):
        gr.Markdown("單句測試")
        with gr.Row():
            test_text = gr.Textbox(label="單句輸入", placeholder="例：這家好吃到爆！", lines=2)
            btn_pred = gr.Button("預測")
        pred_out = gr.Textbox(label="結果", interactive=False)

        btn_pred.click(fn=predict_one, inputs=[test_text, state_model], outputs=[pred_out])


if __name__ == "__main__":
    demo.launch()