app.py

#this was in part generated with gemini llm
import shutil
import zipfile
import pathlib
import tempfile

import gradio as gr
import pandas as pd
import PIL.Image

import huggingface_hub
import autogluon.multimodal

# --- Model and Path Configuration ---
MODEL_REPO_ID = "FaiyazAzam/24679-image-autolguon-predictor"
ZIP_FILENAME = "autogluon_image_predictor_dir.zip"
CACHE_DIR = pathlib.Path("hf_assets")
EXTRACT_DIR = CACHE_DIR / "predictor_native"

# --- Model Loading ---
def _prepare_predictor_dir() -> str:
    """Downloads, extracts, and prepares the AutoGluon model directory."""
    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 = autogluon.multimodal.MultiModalPredictor.load(PREDICTOR_DIR)

# --- Class Labels & Prediction Logic ---
CLASS_LABELS = {0: "Pen", 1: "Toy"}

def do_predict(pil_img: PIL.Image.Image):
    """
    Performs prediction on the input image and returns class probabilities.
    """
    if pil_img is None:
        return {}

    with tempfile.TemporaryDirectory() as tmpdir:
        img_path = pathlib.Path(tmpdir) / "input.png"
        pil_img.save(img_path)
        df = pd.DataFrame({"image": [str(img_path)]})
        proba_df = PREDICTOR.predict_proba(df)
        row_series = proba_df.iloc[0]
        pretty_dict = {
            CLASS_LABELS[class_idx]: prob
            for class_idx, prob in row_series.items()
            if class_idx in CLASS_LABELS
        }
        return pretty_dict

# --- Gradio User Interface ---
EXAMPLES = [
    ["https://www.penboutique.com/cdn/shop/articles/IMG_6759.jpg?v=1701974210&width=1920"],
    ["https://media.officedepot.com/images/f_auto,q_auto,e_sharpen,h_450/products/790761/790761_p_pilot_g_2_retractable_gel_pens/790761"],
    ["https://i5.walmartimages.com/seo/Disney-Pixar-Toy-Story-True-Talkers-Woody-Figure-with-15-Phrases_8c8c4a17-fb26-4f97-a284-1315c48c18ca.c35d5f2d8b932a490db9bb3f40977220.jpeg"]
]

with gr.Blocks() as demo:
    gr.Markdown("# Pen or Toy?")
    gr.Markdown(
        """
        This is a simple app that demonstrates how to use an AutoGluon Multimodal
        predictor in a Gradio Space to classify images. To use,
        just upload a photo or use one of the examples below. The result will be 
        generated automatically.
        """
    )
    with gr.Row():
        with gr.Column():
            image_in = gr.Image(type="pil", label="Input image", sources=["upload", "webcam"])
        with gr.Column():
            proba_pretty = gr.Label(num_top_classes=2, label="Class Probabilities")

    image_in.change(fn=do_predict, inputs=[image_in], outputs=[proba_pretty])

    # FIX: Added `fn` and `outputs` to allow example caching to work correctly.
    gr.Examples(
        examples=EXAMPLES,
        inputs=[image_in],
        outputs=[proba_pretty],
        fn=do_predict,
        label="Representative examples",
        examples_per_page=8,
        cache_examples=True,
    )

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