Upload app.py with huggingface_hub

app.py

@@ -1,12 +1,23 @@
-
-import os, io, zipfile, pathlib, shutil, traceback
+import os, shutil, zipfile, pathlib, traceback, math
 import numpy as np
 import pandas as pd
-from PIL import Image, ImageOps
+
 import gradio as gr
+from PIL import Image, ImageOps
+
+from datasets import load_dataset
 from huggingface_hub import hf_hub_download
+
+# AutoGluon (multimodal)
 from autogluon.multimodal import MultiModalPredictor
-from datasets import load_dataset
+
+# ---------------- Settings ----------------
+TITLE = "💄 Lipstick Detection (EfficientNet-B0 via AutoGluon)"
+DESC  = (
+    "- Upload a face image; the model predicts **Lipstick** vs **No Lipstick**.\n"
+    "- Left: original; Right: the **preprocessed** 224×224 view used by the model.\n"
+    "- Teaching demo only."
+)
 
 # =========================
 # CONFIG
@@ -48,20 +59,21 @@ def load_mm_predictor():
     if downloaded.endswith(".zip"):
         with zipfile.ZipFile(downloaded, "r") as zf:
             zf.extractall(MODEL_DIR)
-        # predictor saves as a directory; point to top-level folder
-        # If there is exactly one folder inside, use it
         kids = [p for p in MODEL_DIR.iterdir()]
         load_path = MODEL_DIR
         if len(kids) == 1 and kids[0].is_dir():
             load_path = kids[0]
     else:
-        # .pkl can be loaded from its file path’s parent
         load_path = downloaded
 
     predictor = MultiModalPredictor.load(str(load_path))
     return predictor
 
-PREDICTOR = load_mm_predictor()
+try:
+    PREDICTOR = load_mm_predictor()
+except Exception as e:
+    PREDICTOR = None
+    print("Failed to load predictor:", e)
 
 # =========================
 # Preprocess (for visualization)
@@ -71,7 +83,6 @@ TARGET_SIZE = 224
 RESIZE_SHORT = 256
 
 def preprocess_for_viz(pil_img: Image.Image) -> Image.Image:
-    # Keep aspect ratio, resize shorter side to 256, then center crop 224x224
     img = pil_img.convert("RGB")
     img = ImageOps.exif_transpose(img)  # respect orientation EXIF
     w, h = img.size
@@ -92,10 +103,14 @@ ALLOWED_EXT = {".jpg", ".jpeg", ".png"}
 MAX_BYTES   = 8 * 1024 * 1024  # 8 MB
 MAX_SIDE    = 4096
 
-def validate_image(fileobj: gr.File) -> str:
+def validate_image(fileobj) -> str:
     if fileobj is None:
         raise ValueError("Please upload an image.")
-    path = fileobj.name if hasattr(fileobj, "name") else str(fileobj)
+    # gr.Image with type="filepath" returns a str path in Spaces
+    path = getattr(fileobj, "name", fileobj)
+    if not isinstance(path, str):
+        path = str(path)
+
     ext = pathlib.Path(path).suffix.lower()
     if ext not in ALLOWED_EXT:
         raise ValueError("Unsupported file type. Please upload a PNG or JPG/JPEG.")
@@ -112,38 +127,33 @@ def validate_image(fileobj: gr.File) -> str:
 # =========================
 def infer(image_file, threshold=0.5, tta=False):
     try:
+        if PREDICTOR is None:
+            raise RuntimeError("Model failed to load. Check model artifacts and environment.")
+
         # 1) Validate & open
         path = validate_image(image_file)
         orig = Image.open(path).convert("RGB")
         vis = preprocess_for_viz(orig.copy())
 
         # 2) Build test dataframe for predictor
-        tmp_path = path  # predictor wants file path
-        df = pd.DataFrame([{"image": tmp_path}])
+        df = pd.DataFrame([{"image": path}])
 
         # 3) Predict proba for lipstick=1 (binary)
         proba_main = PREDICTOR.predict_proba(df)
-        # MultiModalPredictor returns a DataFrame; get prob for positive class if present
-        # If the column names are [0,1] or ["0","1"], handle generically:
         row = proba_main.iloc[0]
-        # Try to locate the positive class (1 / "1" / "lipstick")
         p1 = None
         for k in [1, "1", "lipstick", "Lipstick", "positive", "True"]:
             if k in row.index:
-                p1 = float(row[k])
-                break
+                p1 = float(row[k]); break
         if p1 is None:
-            # fallback: if only two columns, choose the higher prob and assume it's lipstick for display
             if len(row.index) == 2:
                 p1 = float(max(row.values))
             else:
-                # last resort: if single-prob (e.g., sigmoid), cast to float
                 p1 = float(row.values[0])
 
         # 4) Optional simple TTA: average with horizontally flipped image prediction
         if tta:
             flipped = orig.transpose(Image.FLIP_LEFT_RIGHT)
-            # save flipped temporarily
             flip_tmp = pathlib.Path(path).with_suffix(".flip_tmp.jpg")
             flipped.save(flip_tmp, format="JPEG", quality=95)
             df_flip = pd.DataFrame([{"image": str(flip_tmp)}])
@@ -164,13 +174,11 @@ def infer(image_file, threshold=0.5, tta=False):
         p0 = float(1.0 - p1)
         decision = "Lipstick" if p1 >= float(threshold) else "No Lipstick"
 
-        # 5) Return: original image, preprocessed image, class probabilities, decision text
         return orig, vis, {"Lipstick": p1, "No Lipstick": p0}, f"Prediction: {decision} (p1={p1:.3f})"
 
     except Exception as e:
         tb = traceback.format_exc(limit=1)
-        return None, None, {"error": f"{type(e).__name__}: {e}"}, f"Failed: {type(e).__name__}: {e}
-{tb}"
+        return None, None, {"error": f"{type(e).__name__}: {e}"}, f"Failed: {type(e).__name__}: {e}\n{tb}"
 
 # =========================
 # Build examples from dataset (if available)
@@ -190,17 +198,14 @@ def get_examples(n=3):
     cnt = min(n, len(ds))
     for i in range(cnt):
         rec = ds[i]
-        # Assume the image column is "image" with PIL Image or path
         img = rec.get("image", None)
         if img is None:
             continue
         if isinstance(img, Image.Image):
-            # save to temp file
             p = f"example_{i}.jpg"
             img.convert("RGB").save(p, "JPEG", quality=95)
             ex.append([p, 0.5, False])
         else:
-            # could be dict/path-like
             ex.append([img, 0.5, False])
     return ex
 
@@ -210,14 +215,8 @@ EXAMPLES = get_examples(3)
 # Gradio UI
 # =========================
 with gr.Blocks() as demo:
-    gr.Markdown("# 💄 Lipstick Detection (AutoGluon — EfficientNet-B0)")
-    gr.Markdown(
-        "- Upload a face image; the model predicts **Lipstick** vs **No Lipstick**.  
-"
-        "- Left: original; Right: the **preprocessed** 224×224 view seen by the model.  
-"
-        "- This is a **teaching demo**; don’t use for real decisions."
-    )
+    gr.Markdown(f"# {TITLE}")
+    gr.Markdown(DESC)
 
     with gr.Row():
         with gr.Column():
@@ -240,4 +239,6 @@ with gr.Blocks() as demo:
         cache_examples=False
     )
 
-demo.launch()
+# In Spaces, share=True is not required; leaving default.
+if __name__ == "__main__":
+    demo.launch()