Upload folder using huggingface_hub

README.md

@@ -1,12 +1,31 @@
 ---
 title: Sign Image Classifier
-emoji: 🏆
-colorFrom: indigo
-colorTo: pink
+emoji: 🖼️
+colorFrom: blue
+colorTo: green
 sdk: gradio
-sdk_version: 5.47.2
+sdk_version: "4.31.5"
 app_file: app.py
 pinned: false
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# Sign Image Classifier — Classmate Model
+
+This Space wraps a classmate-trained AutoGluon `MultiModalPredictor` for **sign identification**.
+
+## Dataset and Model
+- Model: [its-zion-18/sign-image-autogluon-predictor](https://huggingface.co/its-zion-18/sign-image-autogluon-predictor)
+- Dataset: [ecopus/sign_identification](https://huggingface.co/datasets/ecopus/sign_identification)
+
+## How it works
+- Upload an image (PNG/JPG) or use webcam.
+- App shows both the **original** and the **preprocessed** image that goes into the model.
+- Adjustable preprocessing parameters:
+  - Resize side
+  - Keep aspect ratio
+
+## Output
+- Sorted class probabilities.
+
+## Limitations
+- The model achieved 100% accuracy on the original split, but since the augmented split was derived from the original images (rotations, flips, etc.), there may be data leakage. Generalization performance is uncertain.

app.py

@@ -0,0 +1,108 @@
+import os
+import shutil
+import zipfile
+import pathlib
+import tempfile
+import pandas as pd
+from PIL import Image
+import gradio
+import huggingface_hub
+import autogluon.multimodal as ag_mm
+from datasets import load_dataset
+
+MODEL_REPO_ID = "its-zion-18/sign-image-autogluon-predictor"
+ZIP_FILENAME  = "autogluon_image_predictor_dir.zip"
+HF_TOKEN = os.environ.get("HUGGINGFACEHUB_API_TOKEN", None)
+
+CACHE_DIR   = pathlib.Path("hf_assets")
+EXTRACT_DIR = CACHE_DIR / "predictor_native"
+
+ALLOWED_EXTS = {".png", ".jpg", ".jpeg"}
+MAX_FILE_MB = 5
+
+def _prepare_predictor_dir() -> str:
+    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",
+        token=HF_TOKEN,
+        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 = ag_mm.MultiModalPredictor.load(PREDICTOR_DIR)
+
+def _validate_image_file(tmp_path: pathlib.Path) -> None:
+    size_mb = tmp_path.stat().st_size / (1024 * 1024)
+    if size_mb > MAX_FILE_MB:
+        raise ValueError(f"File too large: {size_mb:.2f} MB > {MAX_FILE_MB} MB")
+    if tmp_path.suffix.lower() not in ALLOWED_EXTS:
+        raise ValueError(f"Unsupported file type: {tmp_path.suffix}. Allowed: {sorted(ALLOWED_EXTS)}")
+
+def _preprocess_for_model(pil_img: Image.Image, resize_side: int, keep_aspect: bool) -> Image.Image:
+    img = pil_img.convert("RGB")
+    if keep_aspect:
+        img.thumbnail((resize_side, resize_side), Image.BILINEAR)
+    else:
+        img = img.resize((resize_side, resize_side), Image.BILINEAR)
+    return img
+
+def do_predict(pil_img: Image.Image, resize_side: int = 224, keep_aspect: bool = True):
+    if pil_img is None:
+        return None, None, {"Error": "No image provided."}
+
+    tmpdir = pathlib.Path(tempfile.mkdtemp())
+    orig_path = tmpdir / "original.png"
+    pil_img.save(orig_path, format="PNG")
+    _validate_image_file(orig_path)
+
+    pre_img = _preprocess_for_model(pil_img, resize_side=resize_side, keep_aspect=keep_aspect)
+    pre_path = tmpdir / "preprocessed.png"
+    pre_img.save(pre_path, format="PNG")
+
+    df = pd.DataFrame({"image": [str(pre_path)]})
+    proba_df = PREDICTOR.predict_proba(df)
+
+    row = proba_df.iloc[0]
+    pretty = {str(k): float(v) for k, v in row.items()}
+    pretty = dict(sorted(pretty.items(), key=lambda kv: kv[1], reverse=True))
+
+    return pil_img, pre_img, pretty
+
+# Prepare 3 example images from dataset
+ds = load_dataset("ecopus/sign_identification", split="original")
+EXAMPLES = [[ds[i]["image"]] for i in range(3)]
+
+with gradio.Blocks() as demo:
+    gradio.Markdown("# Sign Image Classifier — Classmate Model")
+    gradio.Markdown("Upload an image (PNG/JPG) or use webcam. The app shows both the original and preprocessed (resized) version.")
+
+    with gradio.Row():
+        image_in = gradio.Image(type="pil", label="Upload Image", sources=["upload", "webcam"])
+        with gradio.Column():
+            resize_side = gradio.Slider(64, 512, value=224, step=16, label="Resize side (pixels)")
+            keep_aspect = gradio.Checkbox(value=True, label="Keep aspect ratio")
+
+    with gradio.Row():
+        orig_out = gradio.Image(type="pil", label="Original")
+        prep_out = gradio.Image(type="pil", label="Preprocessed")
+
+    proba_out = gradio.Label(num_top_classes=5, label="Class Probabilities")
+
+    for comp in [image_in, resize_side, keep_aspect]:
+        comp.change(fn=do_predict, inputs=[image_in, resize_side, keep_aspect],
+                    outputs=[orig_out, prep_out, proba_out])
+
+    gradio.Examples(examples=EXAMPLES, inputs=[image_in], label="Example Images", cache_examples=False)
+
+demo.launch()

requirements.txt

@@ -0,0 +1,7 @@
+gradio
+huggingface_hub
+pandas
+pillow
+torch
+autogluon.multimodal
+datasets