app.py

import os
import re
import pandas as pd
import gradio as gr

from sklearn.model_selection import train_test_split
from sklearn.feature_extraction.text import CountVectorizer
from sklearn.tree import DecisionTreeClassifier
from sklearn.metrics import accuracy_score, classification_report


# -----------------------------
# Helpers
# -----------------------------
def _guess_column(df: pd.DataFrame, candidates):
    """Find the first matching column name (case-insensitive) from candidates."""
    cols_lower = {c.lower(): c for c in df.columns}
    for cand in candidates:
        if cand.lower() in cols_lower:
            return cols_lower[cand.lower()]
    return None

def _clean_address_series(s: pd.Series) -> pd.Series:
    """Light cleaning: ensure string, strip, collapse whitespace."""
    s = s.astype(str).fillna("")
    s = s.str.replace(r"\s+", " ", regex=True).str.strip()
    return s

def _require(cond: bool, msg: str):
    if not cond:
        raise ValueError(msg)


# -----------------------------
# Core: Train / Predict
# -----------------------------
def train_from_csv(
    file_obj,
    address_col_name,
    label_col_name,
    test_size,
    max_features,
    openai_key,
    state,
):
    """
    Train model from uploaded CSV.
    - file_obj: gr.File
    - address_col_name/label_col_name: optional user overrides
    """
    _require(file_obj is not None, "Please upload a CSV file first.")

    # Store OpenAI key in-session (optional; not used unless you extend the app)
    # Do NOT print it. Keep it in state.
    if openai_key:
        state["openai_key"] = openai_key

    path = file_obj.name
    df = pd.read_csv(path)

    # Auto-detect columns if not provided
    address_col = address_col_name.strip() if address_col_name.strip() else _guess_column(
        df, ["address", "Address", "full_address", "Full_Address", "addr", "ADDR"]
    )
    label_col = label_col_name.strip() if label_col_name.strip() else _guess_column(
        df, ["label", "Label", "category", "Category", "class", "Class", "y", "Y"]
    )

    _require(address_col is not None, "Could not find an address column. Provide it in 'Address column name'.")
    _require(label_col is not None, "Could not find a label column. Provide it in 'Label column name'.")

    _require(address_col in df.columns, f"Address column '{address_col}' not found in CSV.")
    _require(label_col in df.columns, f"Label column '{label_col}' not found in CSV.")

    # Clean + drop bad rows
    df = df[[address_col, label_col]].copy()
    df[address_col] = _clean_address_series(df[address_col])
    df[label_col] = df[label_col].astype(str).fillna("").str.strip()

    df = df[(df[address_col] != "") & (df[label_col] != "")]
    _require(len(df) >= 50, f"Not enough usable rows after cleaning: {len(df)}. Need at least ~50.")

    # Vectorize
    vectorizer = CountVectorizer(max_features=int(max_features))
    X = vectorizer.fit_transform(df[address_col])
    y = df[label_col]

    # Split
    X_train, X_val, y_train, y_val = train_test_split(
        X, y, test_size=float(test_size), random_state=42, stratify=y if y.nunique() > 1 else None
    )

    # Train model
    model = DecisionTreeClassifier(random_state=42)
    model.fit(X_train, y_train)

    # Validate
    y_pred = model.predict(X_val)
    acc = accuracy_score(y_val, y_pred)
    report = classification_report(y_val, y_pred, zero_division=0)

    # Save to state for prediction
    state["model"] = model
    state["vectorizer"] = vectorizer
    state["address_col"] = address_col
    state["label_col"] = label_col
    state["trained"] = True

    summary = (
        f"✅ Trained DecisionTreeClassifier\n"
        f"- Rows used: {len(df)}\n"
        f"- Address col: {address_col}\n"
        f"- Label col: {label_col}\n"
        f"- Validation accuracy: {acc:.4f}\n\n"
        f"Classification report:\n{report}"
    )

    return summary, state


def predict_address(address_text, state):
    _require(state.get("trained"), "Model not trained yet. Upload CSV and click Train first.")
    _require(address_text is not None and address_text.strip() != "", "Enter an address to classify.")

    model = state["model"]
    vectorizer = state["vectorizer"]

    addr = re.sub(r"\s+", " ", address_text.strip())
    X = vectorizer.transform([addr])
    pred = model.predict(X)[0]

    # Optional confidence if the model supports predict_proba
    conf_str = ""
    if hasattr(model, "predict_proba"):
        probs = model.predict_proba(X)[0]
        classes = model.classes_
        p = float(probs[list(classes).index(pred)])
        conf_str = f" (confidence ~ {p:.3f})"

    return f"{pred}{conf_str}"


def clear_model(state):
    state.clear()
    state.update({"trained": False})
    return "Cleared trained model from session.", state


# -----------------------------
# UI
# -----------------------------
with gr.Blocks(title="Address Classifier Trainer") as demo:
    state = gr.State({"trained": False})

    gr.Markdown(
        """
# Address Classification Trainer (CSV → Train → Predict)

**Workflow**
1) Drag & drop your labeled CSV (15k or any size)  
2) Click **Train**  
3) Enter an address and click **Predict**  

**Required CSV columns**
- Address column (e.g., `address`)
- Label column (e.g., `label` or `category`)
"""
    )

    with gr.Row():
        file_in = gr.File(label="Upload labeled CSV (drag & drop)", file_types=[".csv"])
        openai_key_in = gr.Textbox(
            label="OpenAI API Key (optional; stored only in this session)",
            type="password",
            placeholder="sk-...",
        )

    with gr.Row():
        address_col_in = gr.Textbox(label="Address column name (optional override)", placeholder="address")
        label_col_in = gr.Textbox(label="Label column name (optional override)", placeholder="label")

    with gr.Row():
        test_size_in = gr.Slider(0.05, 0.5, value=0.2, step=0.05, label="Validation split size")
        max_features_in = gr.Slider(1000, 50000, value=20000, step=1000, label="Max vocabulary size (CountVectorizer)")

    with gr.Row():
        train_btn = gr.Button("Train", variant="primary")
        clear_btn = gr.Button("Clear model")

    train_out = gr.Textbox(label="Training output", lines=18)

    gr.Markdown("## Test a single address")
    with gr.Row():
        address_in = gr.Textbox(label="Address input", placeholder="123 Main St, Baltimore, MD 21201", lines=2)
        predict_btn = gr.Button("Predict", variant="primary")

    pred_out = gr.Textbox(label="Prediction", lines=2)

    # Wire actions
    train_btn.click(
        fn=train_from_csv,
        inputs=[file_in, address_col_in, label_col_in, test_size_in, max_features_in, openai_key_in, state],
        outputs=[train_out, state],
    )

    predict_btn.click(
        fn=predict_address,
        inputs=[address_in, state],
        outputs=[pred_out],
    )

    clear_btn.click(
        fn=clear_model,
        inputs=[state],
        outputs=[train_out, state],
    )

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