Update app.py

app.py

@@ -1,12 +1,11 @@
 # app.py
 import gradio as gr
-from PIL import Image
+from PIL import Image, ImageOps
 import numpy as np
 import pandas as pd
 import io
 import os
 import cv2
-import glob
 from pathlib import Path
 from typing import List, Dict, Any, Tuple
 
@@ -35,6 +34,21 @@ def _to_np(img: Image.Image) -> np.ndarray:
     return np.array(_ensure_rgb(img))
 
 
+def _thumbnail(img: Image.Image, max_side: int = 320) -> Image.Image:
+    """
+    Create a *small* thumbnail for gallery display so you don't need to scroll.
+    Keeps aspect ratio; pads to square so grids look neat.
+    """
+    img = _ensure_rgb(img.copy())
+    img.thumbnail((max_side, max_side))
+    # Pad to square with white (looks nicer for product photos)
+    w, h = img.size
+    side = max(w, h)
+    bg = Image.new("RGB", (side, side), (255, 255, 255))
+    bg.paste(img, ((side - w) // 2, (side - h) // 2))
+    return bg
+
+
 def _hsv_hist_features(img: Image.Image) -> np.ndarray:
     """Return simple features for matching and scoring.
     - Hue histogram (18 bins)
@@ -69,7 +83,9 @@ def _complementary_hue_score(q_hist: np.ndarray, w_hist: np.ndarray) -> float:
     q_h = q_hist[:hb]
     w_h = w_hist[:hb]
     q_shift = np.roll(q_h, hb // 2)
-    hue_sim = float(np.dot(q_shift, w_h) / (np.linalg.norm(q_shift) * np.linalg.norm(w_h) + 1e-8))
+    # cosine similarity on shifted hues
+    denom = (np.linalg.norm(q_shift) * np.linalg.norm(w_h) + 1e-8)
+    hue_sim = float(np.dot(q_shift, w_h) / denom)
 
     # Encourage pairing items with different edge density (texture contrast)
     q_ed = q_hist[-1]
@@ -254,7 +270,9 @@ def _get_embedder() -> _Embedder:
 # ----------------------------
 # State schema:
 # {
-#   "wardrobe": [ {"id": int, "name": str, "image": PIL.Image, "features": np.ndarray, "embedding": np.ndarray, "rating": int|None} ],
+#   "wardrobe": [ {"id": int, "name": str, "image": PIL.Image, "thumb": PIL.Image,
+#                  "features": np.ndarray, "embedding": np.ndarray, "category": str,
+#                  "rating": int|None} ],
 #   "selected_idx": int|None
 # }
 
@@ -262,7 +280,8 @@ def _blank_state() -> Dict[str, Any]:
     return {"wardrobe": [], "selected_idx": None}
 
 
-
+# ----------------------------
+# Wardrobe management
 # ----------------------------
 
 def add_wardrobe(files: List[Any], state: Dict[str, Any]):
@@ -293,6 +312,7 @@ def add_wardrobe(files: List[Any], state: Dict[str, Any]):
                 "id": next_id,
                 "name": name,
                 "image": img,
+                "thumb": _thumbnail(img),
                 "features": feats,
                 "embedding": emb,
                 "category": category,
@@ -307,9 +327,7 @@ def add_wardrobe(files: List[Any], state: Dict[str, Any]):
 
 
 def add_wardrobe_from_dir(example_dir: str, state: Dict[str, Any]):
-    """Load all images in a folder into the wardrobe and auto-rate/classify them.
-    Used by gr.Examples. Accepts relative paths in the Space repo.
-    """
+    """Load all images in a folder into the wardrobe and auto-rate/classify them."""
     if not example_dir:
         return state, _render_gallery(state), _ratings_df(state)
     p = Path(example_dir)
@@ -325,34 +343,13 @@ def clear_wardrobe(state: Dict[str, Any]):
     return state, [], _ratings_df(state)
 
 
-def _render_gallery(state: Dict[str, Any]) -> List[Image.Image]:
-    return [w["image"] for w in state.get("wardrobe", [])]
-
-
-def on_select_item(item_label: str, state: Dict[str, Any]):
-    if not item_label:
-        return state, None, gr.update(value=50)
-    # label format: "#<id> · <name>"
-    try:
-        item_id = int(item_label.split(" ")[0][1:])
-    except Exception:
-        return state, None, gr.update(value=50)
-
-    idx = next((i for i, w in enumerate(state["wardrobe"]) if w["id"] == item_id), None)
-    state["selected_idx"] = idx
-    if idx is None:
-        return state, None, gr.update(value=50)
-    w = state["wardrobe"][idx]
-    current_rating = w["rating"] if w["rating"] is not None else 50
-    return state, w["image"], gr.update(value=int(current_rating))
-
-
-def save_rating(rating: int, state: Dict[str, Any]):
-    idx = state.get("selected_idx", None)
-    if idx is None:
-        return state, _ratings_df(state)
-    state["wardrobe"][idx]["rating"] = int(rating)
-    return state, _ratings_df(state)
+def _render_gallery(state: Dict[str, Any]) -> List[Tuple[Image.Image, str]]:
+    """Return list of (thumbnail, caption)."""
+    out = []
+    for w in state.get("wardrobe", []):
+        caption = f"#{w['id']} · {w['name']} · {w.get('category','?')} · {w.get('rating','-')}/100"
+        out.append((w["thumb"], caption))
+    return out
 
 
 def _ratings_df(state: Dict[str, Any]) -> pd.DataFrame:
@@ -376,25 +373,6 @@ def export_ratings(state: Dict[str, Any]):
     df.to_csv(buf, index=False)
     buf.seek(0)
     return buf
-    buf = io.BytesIO()
-    df.to_csv(buf, index=False)
-    buf.seek(0)
-    return buf
-
-
-def import_ratings(file_obj, state: Dict[str, Any]):
-    # Deprecated in auto-rating flow; keep no-op for compatibility
-    return state, _ratings_df(state)
-    try:
-        df = pd.read_csv(file_obj.name if hasattr(file_obj, "name") else file_obj)
-        names_to_rating = {str(row["name"]): int(row["rating"]) if not pd.isna(row["rating"]) else None
-                               for _, row in df.iterrows()}
-        for w in state.get("wardrobe", []):
-            if w["name"] in names_to_rating:
-                w["rating"] = names_to_rating[w["name"]]
-    except Exception:
-        pass
-    return state, _ratings_df(state)
 
 
 # ----------------------------
@@ -403,7 +381,9 @@ def import_ratings(file_obj, state: Dict[str, Any]):
 
 def rate_and_recommend(query_img: Image.Image, top_k: int, matching_mode: str, state: Dict[str, Any]):
     if query_img is None:
-        return 0, "No image provided.", []
+        return 0, "No image provided.", [], pd.DataFrame(columns=[
+            "rank", "name", "category", "model_rating", "match_score"
+        ])
 
     query_img = _ensure_rgb(query_img)
 
@@ -452,7 +432,7 @@ def rate_and_recommend(query_img: Image.Image, top_k: int, matching_mode: str, s
         else:  # Complementary color + style
             final = 0.5 * ((cos + 1.0) / 2.0) + 0.5 * comp
 
-        # Quality prior
+        # Quality prior (prefer items you generally like)
         if w.get("rating") is not None:
             qual = 0.5 + 0.5 * (w["rating"] / 100.0)
             final *= qual
@@ -470,73 +450,112 @@ def rate_and_recommend(query_img: Image.Image, top_k: int, matching_mode: str, s
             final = 0.5 * ((cos + 1.0) / 2.0) + 0.5 * comp
             candidates.append((final, w))
 
+    # Rank and prepare outputs
     candidates.sort(key=lambda x: x[0], reverse=True)
-    recs = [w["image"] for _, w in candidates[: max(0, top_k)]]
-
-    if len(recs) == 0:
-        txt = f"Predicted rating: {pred}/100. No matches found in your wardrobe."
-        return pred, txt, []
+    chosen = candidates[: max(0, top_k)]
+
+    if not chosen:
+        txt = f"**Predicted wear score:** {pred}/100  \n" \
+              f"**Detected category:** {qcat}  \n" \
+              f"**Results:** No compatible matches found in your wardrobe."
+        return pred, txt, [], pd.DataFrame(columns=[
+            "rank", "name", "category", "model_rating", "match_score"
+        ])
 
-    top_names = ", ".join([f"{w['name']} ({w.get('category')})" for _, w in candidates[: max(0, top_k)]])
+    # Human-friendly summary
+    top_names = ", ".join([f"{w['name']} ({w.get('category')})" for _, w in chosen])
     txt = (
-        f"Predicted rating: {pred}/100. Query category: {qcat}. Suggested pairings: {top_names}."
-        f"Logic: {matching_mode.lower()} with category filtering and quality prior."
+        f"**Predicted wear score:** {pred}/100  \n"
+        f"**Detected category:** `{qcat}`  \n"
+        f"**Matching mode:** `{matching_mode}`  \n"
+        f"**Suggested pairings:** {top_names}  \n"
+        f"_Scoring blends style similarity (CLIP) with color/texture complement and your model ratings._"
     )
-    return pred, txt, recs
 
-    # Model rating
-    try:
-        scorer = _get_scorer()
-        pred = scorer.predict_1to100(query_img)
-    except Exception:
-        # Fallback if model unavailable
-        pred = 50
-
-    # Compute compatibility with wardrobe
-    qfeat = _hsv_hist_features(query_img)
-    candidates = []
-    for w in state.get("wardrobe", []):
-        comp = _complementary_hue_score(qfeat, w["features"])
-        # Weight by user rating if available
-        user_w = 1.0
-        if w["rating"] is not None:
-            user_w = 0.5 + 0.5 * (w["rating"] / 100.0)  # 0.5..1.0
-        final = comp * user_w
-        candidates.append((final, w))
+    # Gallery (thumbnails) + table
+    rec_gallery = [(w["thumb"], f"{w['name']} · {w.get('category')} · {w.get('rating','-')}/100")
+                   for _, w in chosen]
 
-    candidates.sort(key=lambda x: x[0], reverse=True)
-    recs = [w["image"] for _, w in candidates[: max(0, top_k)]]
+    rec_table = pd.DataFrame([{
+        "rank": i + 1,
+        "name": w["name"],
+        "category": w.get("category"),
+        "model_rating": w.get("rating"),
+        "match_score": round(float(s), 3)
+    } for i, (s, w) in enumerate(chosen)])
 
-    if len(recs) == 0:
-        txt = f"Predicted rating: {pred}/100. No matches found in your wardrobe."
-        return pred, txt, []
-
-    top_names = ", ".join([w["name"] for _, w in candidates[: max(0, top_k)]])
-    txt = (
-        f"Predicted rating: {pred}/100. Suggested pairings from your wardrobe: {top_names}."
-        f"Logic: complementary hues + texture contrast + your cached ratings."
-    )
-    return pred, txt, recs
+    return pred, txt, rec_gallery, rec_table
 
 
 # ----------------------------
 # Gradio UI
 # ----------------------------
-with gr.Blocks(title="Wardrobe Rater + Recommender", css=".gradio-container {max-width: 1200px}") as demo:
-    gr.Markdown("# Wardrobe Rater + Recommender\nUpload, auto-rate with your HF model, and get AI pair suggestions. Category-aware matching avoids recommending the same type.")
+
+APP_DESCRIPTION = """
+# Wardrobe Rater + Recommender
+
+**What this app does (1 minute):**
+- **Upload** photos of items in your wardrobe (tops, pants, jackets, shoes).
+- The app **auto-rates** each item with your model and tags its category.
+- When you upload a **new item** (e.g., a shopping photo), it:
+  1) Predicts how likely **you** are to wear it (1–100),
+  2) Suggests the **best matches** from your wardrobe using style similarity and color/texture complement.
+
+**How to use it:**
+1. Go to **“1) Wardrobe Manager”** and upload several wardrobe images (front-on, good lighting).
+2. Then open **“2) Rate + Recommend New Item”**, upload your candidate item, pick the number of matches, and click **Rate + Recommend**.
+3. Review the **score**, the **explanation**, the **top matches** (thumbnails), and the **table**.
+"""
+
+WARDROBE_TIPS = """
+**What to upload:** clear product-style photos (JPG/PNG).  
+Avoid group photos or cluttered backgrounds when possible.  
+You can click thumbnails to preview full size.
+"""
+
+QUERY_TIPS = """
+**Upload a single item** you’re considering (e.g., a screenshot or photo from a listing).  
+Good lighting and centered framing helps the detector and embeddings.
+"""
+
+with gr.Blocks(title="Wardrobe Rater + Recommender",
+               css="""
+               .gradio-container {max-width: 1100px}
+               .small-note {font-size: 0.9em; color: #4b5563}
+               """) as demo:
+    gr.Markdown(APP_DESCRIPTION)
 
     app_state = gr.State(_blank_state())
 
     with gr.Tab("1) Wardrobe Manager"):
+        gr.Markdown("### Upload your wardrobe")
         with gr.Row():
-            wardrobe_uploader = gr.File(label="Upload wardrobe images", file_types=["image"], file_count="multiple")
+            wardrobe_uploader = gr.File(
+                label="Upload wardrobe images",
+                file_types=["image"],
+                file_count="multiple"
+            )
+        gr.Markdown(WARDROBE_TIPS, elem_classes=["small-note"])
         with gr.Row():
-            add_btn = gr.Button("Add to wardrobe (auto-rate + auto-category)")
-            clear_btn = gr.Button("Clear wardrobe")
-        gallery = gr.Gallery(label="Current wardrobe", columns=6, height=250)
-        ratings_table = gr.Dataframe(headers=["id", "name", "category", "model_rating"], interactive=False)
+            add_btn = gr.Button("➕ Add to wardrobe (auto-rate + auto-category)", variant="primary")
+            clear_btn = gr.Button("🗑️ Clear wardrobe", variant="secondary")
+
+        gallery = gr.Gallery(
+            label="Current wardrobe (click to preview)",
+            columns=6,
+            height=220,
+            allow_preview=True,
+            show_label=True
+        )
+
+        gr.Markdown("### Item summary")
+        ratings_table = gr.Dataframe(
+            headers=["id", "name", "category", "model_rating"],
+            interactive=False,
+            wrap=True
+        )
 
-        # Add Gradio Examples for example wardrobe
+        # Optional Example loader (kept hidden input textbox)
         gr.Markdown("### Or load an example wardrobe")
         example_dir = gr.Textbox(label="Example folder path", value="examples/wardrobe_basic", visible=False)
         gr.Examples(
@@ -548,23 +567,61 @@ with gr.Blocks(title="Wardrobe Rater + Recommender", css=".gradio-container {max
             run_on_click=True,
         )
 
-
     with gr.Tab("2) Rate + Recommend New Item"):
+        gr.Markdown("### Upload a candidate item")
         with gr.Row():
-            query_img = gr.Image(label="Upload or take photo", sources=["upload", "webcam","clipboard"], type="pil")
+            query_img = gr.Image(
+                label="Upload or paste image (single item)",
+                sources=["upload", "webcam", "clipboard"],
+                type="pil"
+            )
+            controls_col = gr.Column()
+        with controls_col:
             topk = gr.Slider(1, 6, value=3, step=1, label="# of matches to return")
-            matching_mode = gr.Radio(["Complementary color+style", "Similar style"], value="Complementary color+style", label="Matching mode")
-        go_btn = gr.Button("Rate + Recommend")
+            matching_mode = gr.Radio(
+                ["Complementary color+style", "Similar style"],
+                value="Complementary color+style",
+                label="Matching mode"
+            )
+            go_btn = gr.Button("⭐ Rate + Recommend", variant="primary")
+        gr.Markdown(QUERY_TIPS, elem_classes=["small-note"])
+
         with gr.Row():
-            pred_score = gr.Number(label="Predicted rating (1-100)")
-        rec_text = gr.Markdown()
-        rec_gallery = gr.Gallery(label="Matches in your wardrobe", columns=6, height=250)
+            pred_score = gr.Number(label="Predicted wear score (1–100)")
+        rec_text = gr.Markdown()  # human-readable summary
+
+        gr.Markdown("### Top matches (click to preview)")
+        rec_gallery = gr.Gallery(
+            columns=6,
+            height=220,
+            allow_preview=True,
+            show_label=False
+        )
+
+        gr.Markdown("### Details")
+        rec_table = gr.Dataframe(
+            headers=["rank", "name", "category", "model_rating", "match_score"],
+            interactive=False,
+            wrap=True
+        )
 
     # --- Wiring ---
-    add_btn.click(add_wardrobe, inputs=[wardrobe_uploader, app_state], outputs=[app_state, gallery, ratings_table])
-    clear_btn.click(clear_wardrobe, inputs=[app_state], outputs=[app_state, gallery, ratings_table])
+    add_btn.click(
+        add_wardrobe,
+        inputs=[wardrobe_uploader, app_state],
+        outputs=[app_state, gallery, ratings_table]
+    )
+    clear_btn.click(
+        clear_wardrobe,
+        inputs=[app_state],
+        outputs=[app_state, gallery, ratings_table]
+    )
 
-    go_btn.click(rate_and_recommend, inputs=[query_img, topk, matching_mode, app_state], outputs=[pred_score, rec_text, rec_gallery])
+    go_btn.click(
+        rate_and_recommend,
+        inputs=[query_img, topk, matching_mode, app_state],
+        outputs=[pred_score, rec_text, rec_gallery, rec_table]
+    )
 
 # Lightweight tests. Run only when RUN_TESTS=1
 if __name__ == "__main__":
@@ -586,9 +643,14 @@ if __name__ == "__main__":
         # Test recommend path with small wardrobe
         st = _blank_state()
         img_b = solid(32)
-        st["wardrobe"].append({"id":0, "name":"test.png", "image":img_b, "features":_hsv_hist_features(img_b), "embedding":_get_embedder().embed(img_b), "rating":50})
-        pred, txt, recs = rate_and_recommend(solid(200), 1, "Similar style", st)
+        st["wardrobe"].append({
+            "id":0, "name":"test.png", "image":img_b, "thumb":_thumbnail(img_b),
+            "features":_hsv_hist_features(img_b), "embedding":_get_embedder().embed(img_b),
+            "rating":50, "category":"pants"
+        })
+        pred, txt, recs, tbl = rate_and_recommend(solid(200), 1, "Similar style", st)
         assert isinstance(pred, int) and isinstance(txt, str) and isinstance(recs, list)
+        assert isinstance(tbl, pd.DataFrame)
         print("Tests passed.")
     else:
-        demo.launch()
+        demo.launch()