app.py

import os, pathlib, zipfile, tempfile
import pandas as pd
import gradio as gr
from PIL import Image
from huggingface_hub import hf_hub_download, snapshot_download
import autogluon.multimodal as ag

MODEL_REPO_ID = "samder03/2025-24679-image-autogluon-predictor"

CLASS_LABELS = {
    0: "No Stop Sign",
    1: "Stop Sign",
    "class_0": "No Stop Sign",
    "class_1": "Stop Sign",
}

def _human_label(c):
    try:
        ci = int(c)
        return CLASS_LABELS.get(ci, str(c))
    except Exception:
        return CLASS_LABELS.get(str(c), str(c))

def _locate_predictor_dir_from_repo_folder(repo_dir: str) -> str:
    rd = pathlib.Path(repo_dir)
    for p in rd.rglob("predictor.pkl"):
        return str(p.parent)
    return ""

def _prepare_predictor_dir() -> str:
    repo_dir = snapshot_download(repo_id=MODEL_REPO_ID, repo_type="model")
    pred_dir = _locate_predictor_dir_from_repo_folder(repo_dir)
    if pred_dir:
        return pred_dir

    # Fallback: try to find a zip and extract
    zips = list(pathlib.Path(repo_dir).rglob("*.zip"))
    if not zips:
        raise FileNotFoundError("Could not find a predictor directory or .zip in the model repo.")
    zip_path = str(zips[0])

    workdir = tempfile.mkdtemp(prefix="ag_img_predictor_")
    with zipfile.ZipFile(zip_path, "r") as zf:
        zf.extractall(workdir)
    entries = list(pathlib.Path(workdir).iterdir())
    if len(entries) == 1 and entries[0].is_dir():
        return str(entries[0])
    return workdir

PREDICTOR_DIR = _prepare_predictor_dir()
PREDICTOR = ag.MultiModalPredictor.load(PREDICTOR_DIR)

def _ensure_rgb(img: Image.Image) -> Image.Image:
    return img.convert("RGB") if img.mode != "RGB" else img

def _resize_shorter(img: Image.Image, shorter: int) -> Image.Image:
    w, h = img.size
    if min(w, h) == shorter:
        return img
    if w < h:
        new_w = shorter
        new_h = int(h * (shorter / w))
    else:
        new_h = shorter
        new_w = int(w * (shorter / h))
    return img.resize((new_w, new_h), Image.BICUBIC)

def _center_crop(img: Image.Image, size: int) -> Image.Image:
    w, h = img.size
    side = min(w, h, size)
    left = (w - side) // 2
    top  = (h - side) // 2
    return img.crop((left, top, left + side, top + side)).resize((size, size), Image.BICUBIC)

def _validate_image(pil_img: Image.Image, max_pixels: int = 8_000_000):
    if pil_img is None:
        return False, "No image provided."
    if pil_img.width * pil_img.height > max_pixels:
        return False, f"Image too large (>{max_pixels:,} pixels). Please upload a smaller image."
    return True, ""

def preprocess(pil_img: Image.Image, resize_shorter: int, do_center_crop: bool, crop_size: int) -> Image.Image:
    img = _ensure_rgb(pil_img)
    img = _resize_shorter(img, resize_shorter)
    if do_center_crop:
        img = _center_crop(img, crop_size)
    return img

def do_predict(pil_img: Image.Image, resize_shorter: int, do_center_crop: bool, crop_size: int, top_k: int):
    ok, msg = _validate_image(pil_img)
    if not ok:
        return None, None, {"Error": 1.0}

    pre_img = preprocess(pil_img, resize_shorter, do_center_crop, crop_size)

    tmpdir = pathlib.Path(tempfile.mkdtemp(prefix="ag_img_run_"))
    orig_path = tmpdir / "original.png"
    pre_path  = tmpdir / "preprocessed.png"
    pil_img.save(orig_path)
    pre_img.save(pre_path)

    df = pd.DataFrame({"image": [str(pre_path)]})
    proba_df = PREDICTOR.predict_proba(df)
    proba_df = proba_df.rename(columns={c: _human_label(c) for c in proba_df.columns})
    row = proba_df.iloc[0]

    items = sorted(row.items(), key=lambda kv: float(kv[1]), reverse=True)[:max(1, int(top_k))]
    pretty = {k: float(v) for k, v in items}

    return Image.open(orig_path), Image.open(pre_path), pretty

with gr.Blocks(title="Stop Sign Classifier") as demo:
    gr.Markdown("# Stop Sign? — Image Classifier (Classmate Model)")
    gr.Markdown(
        "Upload a PNG/JPG (or use webcam). You’ll see the **original** and the **preprocessed** image, "
        "plus ranked class probabilities."
    )

    with gr.Row():
        image_in = gr.Image(type="pil", label="Input image (PNG/JPG)", sources=["upload", "webcam"])
        with gr.Column():
            with gr.Accordion("Inference Parameters", open=False):
                resize_shorter = gr.Slider(64, 1024, value=384, step=16, label="Resize (shorter side)")
                do_center_crop = gr.Checkbox(value=True, label="Center-crop to square")
                crop_size      = gr.Slider(64, 1024, value=384, step=16, label="Crop size (if center-crop)")
                top_k          = gr.Slider(1, 2, value=2, step=1, label="Top-K classes to display")

    with gr.Row():
        img_orig = gr.Image(label="Original")
        img_proc = gr.Image(label="Preprocessed")
    proba_pretty = gr.Label(num_top_classes=2, label="Class probabilities (Top-K)")

    image_in.change(
        fn=do_predict,
        inputs=[image_in, resize_shorter, do_center_crop, crop_size, top_k],
        outputs=[img_orig, img_proc, proba_pretty]
    )

    gr.Examples(
      examples=[["examples/example_0.png"], ["examples/example_11.png"], ["examples/example_23.png"]],
      inputs=[image_in],
      label="Example images",
      examples_per_page=6,
      cache_examples=False,
  )


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