app.py

import os
import shutil
import zipfile
import pathlib
import pandas as pd
import gradio as gr
import huggingface_hub
import autogluon.tabular as agt
import shutil

# --- Model Loading ---
def _prepare_predictor_dir() -> str:
    """Downloads and extracts the model files from Hugging Face."""
    # Add a check to clear the cache before download
    if CACHE_DIR.exists():
        print(f"Clearing cache directory: {CACHE_DIR}")
        shutil.rmtree(CACHE_DIR)
        CACHE_DIR.mkdir(parents=True, exist_ok=True)




# --- Settings and Metadata ---
MODEL_REPO_ID = "bcueva/2024-24679-tabular-autolguon-predictor"
ZIP_FILENAME = "autogluon_predictor_dir.zip"
CACHE_DIR = pathlib.Path("hf_assets")
EXTRACT_DIR = CACHE_DIR / "predictor_native"

# Features for the model
FEATURE_COLS = [
    "Capacity_ml",
    "Height_cm",
    "Diameter_cm",
    "Weight_g",
    "Material",
]

# The target variable to be predicted
TARGET_COL = "Use_Type"

# Define the possible values for the 'Material' feature
MATERIAL_LABELS = [
    "Ceramic",
    "Glass",
    "Plastic",
    "Stainless Steel",

]

# Mapping the integer labels back to human-readable labels for the 'Use_Type' prediction
# These labels are based on the model's training data.
OUTCOME_LABELS = {
    0: "Hot",
    1: "Cold",
}

# --- Model Loading ---
def _prepare_predictor_dir() -> str:
    """Downloads and extracts the model files from Hugging Face."""
    CACHE_DIR.mkdir(parents=True, exist_ok=True)
    local_zip = huggingface_hub.hf_hub_download(
        repo_id=MODEL_REPO_ID,
        filename=ZIP_FILENAME,
        repo_type="model",
        local_dir=str(CACHE_DIR),
        local_dir_use_symlinks=False,
    )
    if EXTRACT_DIR.exists():
        shutil.rmtree(EXTRACT_DIR)
    EXTRACT_DIR.mkdir(parents=True, exist_ok=True)
    with zipfile.ZipFile(local_zip, "r") as zf:
        zf.extractall(str(EXTRACT_DIR))
    contents = list(EXTRACT_DIR.iterdir())
    predictor_root = contents[0] if (len(contents) == 1 and contents[0].is_dir()) else EXTRACT_DIR
    return str(predictor_root)

PREDICTOR_DIR = _prepare_predictor_dir()
PREDICTOR = agt.TabularPredictor.load(PREDICTOR_DIR, require_py_version_match=False)

# --- Prediction Function ---
def _human_label(c):
    """Maps the model's numerical output to a human-readable label."""
    try:
        ci = int(c)
        if ci in OUTCOME_LABELS:
            return OUTCOME_LABELS[ci]
    except (ValueError, TypeError):
        pass
    if c in OUTCOME_LABELS:
        return OUTCOME_LABELS[c]
    return str(c)

def do_predict(capacity_ml, height_cm, diameter_cm, weight_g, material):
    """
    Takes user input, formats it for the model, and returns a prediction.
    'Cup_ID' is excluded as per the user request.
    """
    # Create a DataFrame for a single prediction
    # 'Cup_ID' is included as a placeholder with a dummy value (e.g., 0)
    # because the model's training data included it.
    row = {
        "Cup_ID": 0,  # Dummy value
        "Capacity_ml": capacity_ml,
        "Height_cm": height_cm,
        "Diameter_cm": diameter_cm,
        "Weight_g": weight_g,
        "Material": material,
    }
    X = pd.DataFrame([row])

    # Get the raw prediction and its label
    pred_series = PREDICTOR.predict(X)
    raw_pred = pred_series.iloc[0]
    pred_label = _human_label(raw_pred)

    # Get prediction probabilities
    try:
        proba = PREDICTOR.predict_proba(X)
        if isinstance(proba, pd.Series):
            proba = proba.to_frame().T

        # Format probabilities into a dictionary for Gradio
        proba_dict = {
            _human_label(cls): float(val) for cls, val in proba.iloc[0].items()
        }
        proba_dict = dict(sorted(proba_dict.items(), key=lambda kv: kv[1], reverse=True))

    except Exception as e:
        print(f"Could not get probabilities: {e}")
        proba_dict = None

    return proba_dict

# --- Gradio UI ---
with gr.Blocks(fill_height=True) as demo:
    gr.Markdown("# Cup Use Predictor ☕️")
    gr.Markdown("""
    Enter the physical properties of a cup to predict its intended use type (Hot or Cold).
    This app uses a pre-trained **AutoGluon** model to classify the cup's purpose.
    """)

    with gr.Column():
        material = gr.Radio(choices=MATERIAL_LABELS, value="Ceramic", label="Material")

    with gr.Row():
      with gr.Column():
        capacity_ml = gr.Number(value=350, label="Capacity (ml)")
        height_cm = gr.Number(value=10.0, label="Height (cm)")

      with gr.Column():
        diameter_cm = gr.Number(value=8.0, label="Diameter (cm)")
        weight_g = gr.Number(value=250, label="Weight (g)")

    predict_btn = gr.Button("Predict Use Type")
    output_label = gr.Label(num_top_classes=2, label="Prediction")

    inputs = [capacity_ml, height_cm, diameter_cm, weight_g, material]

    predict_btn.click(fn=do_predict, inputs=inputs, outputs=output_label)

    gr.Examples(
        examples=[
            [478, 8, 7.7, 315, "Ceramic"], # Example for a coffee mug (likely 'Hot')
            [442, 13.8, 6.4, 155, "Glass"], # Example for a tall drinking glass (likely 'Cold')
            [392, 18, 5.7, 61, "Plastic"], # Example for a small tea cup (likely 'Hot')
            [302, 17.5, 5.5, 783, "Stainless Steel"], # Example for a disposable soda cup (likely 'Cold')
        ],
        inputs=inputs,
        label="Representative Examples",
        examples_per_page=5,
    )

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