Upload 3 files

README.md

@@ -2,10 +2,10 @@
 title: Wonder Finder
 emoji: 🌍
 colorFrom: yellow
-colorTo: yellow
+colorTo: red
 sdk: gradio
-sdk_version: 4.44.0
-python_version: "3.12"
+sdk_version: 4.44.1
+python_version: "3.10"
 app_file: app.py
 pinned: false
 license: mit
@@ -22,18 +22,15 @@ Visual recommender for the 12 Wonders of the World, powered by CLIP embeddings.
 ## How it works
 1. The catalog (11,544 images across 12 wonder classes) is pre-embedded using CLIP ViT-B/32.
 2. User input (image or text) is embedded into the same 512-D space.
-3. Cosine similarity ranks the catalog; the top 3 results are returned with a diversity filter to avoid duplicates.
+3. Cosine similarity ranks the catalog; top 3 results are returned with a diversity filter to avoid duplicates.
 
 ## Dataset
 [chavajaz/wonders_dataset](https://huggingface.co/datasets/chavajaz/wonders_dataset) — CC0-1.0 licensed, ~960 images per class on average.
 
 ## Model
-[openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32) — chosen for its joint image-text embedding space, which enables both image and text input through a single model.
+[openai/clip-vit-base-patch32](https://huggingface.co/openai/clip-vit-base-patch32) — chosen for its joint image–text embedding space, which enables both image and text input through a single model.
 
 ## Cluster analysis
 K-Means at k=12 on the embeddings achieved **ARI = 0.890** and **NMI = 0.927** against ground-truth wonder labels, indicating CLIP's pretrained space already separates the 12 wonders almost perfectly without supervision.
 
-## Files
-- `app.py` — the Gradio application
-- `requirements.txt` — pinned dependencies
-- `wonders_embeddings.parquet` — precomputed CLIP embeddings (one row per catalog image, column `embedding`, aligned 1:1 with the dataset in `train → validation → test` order)
+Built as the final project for Assignment 3.

app.py

@@ -1,12 +1,77 @@
+"""
+Wonder Finder — Visual recommender for the 12 Wonders of the World.
+HF Spaces deployment.
+
+Notes on defensive patches:
+- gradio 4.44.x has a known bug in gradio_client/utils.py where api_info
+  schema generation crashes on `additionalProperties: True` (boolean schema).
+- The bug raises gradio_client.utils.APIInfoParseError, which is NOT a
+  subclass of TypeError/KeyError/AttributeError — so naive try/except misses it.
+- We patch at THREE layers: get_type, _json_schema_to_python_type, and
+  Blocks.get_api_info. Each catches Exception (the broadest possible).
+"""
+
+import os
+
+# Belt-and-suspenders env vars
+os.environ["GRADIO_ANALYTICS_ENABLED"] = "False"
+os.environ["GRADIO_SERVER_NAME"] = "0.0.0.0"
+
+# ============================================================
+# DEFENSIVE PATCHES — must run before any Gradio component init
+# ============================================================
+
+import gradio_client.utils as _gcu
+
+# Patch 1: _json_schema_to_python_type — the inner recursive function
+_original_json_schema = _gcu._json_schema_to_python_type
+def _safe_json_schema(schema, defs=None):
+    # Handle boolean schemas (the actual bug trigger)
+    if isinstance(schema, bool):
+        return "Any"
+    if not isinstance(schema, dict):
+        return "Any"
+    try:
+        return _original_json_schema(schema, defs)
+    except Exception:
+        return "Any"
+_gcu._json_schema_to_python_type = _safe_json_schema
+
+# Patch 2: get_type — wraps the entry-point type checker
+_original_get_type = _gcu.get_type
+def _safe_get_type(schema):
+    if not isinstance(schema, dict):
+        return "Any"
+    try:
+        return _original_get_type(schema)
+    except Exception:
+        return "Any"
+_gcu.get_type = _safe_get_type
+
+# Patch 3: top-level api_info generator — safety net for anything we missed
 import gradio as gr
+import gradio.blocks as _gradio_blocks
+_original_get_api_info = _gradio_blocks.Blocks.get_api_info
+def _safe_get_api_info(self):
+    try:
+        return _original_get_api_info(self)
+    except Exception:
+        return {"named_endpoints": {}, "unnamed_endpoints": {}}
+_gradio_blocks.Blocks.get_api_info = _safe_get_api_info
+
+# ============================================================
+# REGULAR IMPORTS
+# ============================================================
+
 import torch
 import numpy as np
 import pandas as pd
+from PIL import Image
 from transformers import CLIPProcessor, CLIPModel
 from datasets import load_dataset, concatenate_datasets
 
 # ============================================================
-# LOAD EVERYTHING ON STARTUP (runs once when the Space boots)
+# LOAD EVERYTHING ON STARTUP
 # ============================================================
 
 print("Loading CLIP model...")
@@ -18,28 +83,13 @@ processor = CLIPProcessor.from_pretrained(MODEL_NAME)
 
 print("Loading dataset...")
 ds = load_dataset("chavajaz/wonders_dataset")
-# Concatenate whatever splits exist, in a stable order, instead of assuming
-# train/validation/test are all present.
-split_order = [s for s in ["train", "validation", "test"] if s in ds]
-split_order += [s for s in ds.keys() if s not in split_order]
-splits = [ds[s] for s in split_order]
-full_ds = concatenate_datasets(splits) if len(splits) > 1 else splits[0]
+full_ds = concatenate_datasets([ds["train"], ds["validation"], ds["test"]])
 class_names = full_ds.features["label"].names
 
 print("Loading precomputed embeddings...")
 embeddings_df = pd.read_parquet("wonders_embeddings.parquet")
 image_embeddings = np.array(embeddings_df["embedding"].tolist(), dtype=np.float32)
 EMBEDDINGS_TENSOR = torch.tensor(image_embeddings, device=device, dtype=torch.float32)
-# Defensively L2-normalize so cosine similarity and the diversity threshold
-# are correct even if the stored vectors weren't normalized.
-EMBEDDINGS_TENSOR = EMBEDDINGS_TENSOR / EMBEDDINGS_TENSOR.norm(dim=-1, keepdim=True).clamp_min(1e-12)
-
-if len(full_ds) != EMBEDDINGS_TENSOR.shape[0]:
-    print(
-        f"WARNING: dataset has {len(full_ds)} images but the embeddings file has "
-        f"{EMBEDDINGS_TENSOR.shape[0]} rows. They must line up 1:1 and be in the "
-        f"same order for results to be correct."
-    )
 
 print(f"Ready. {len(full_ds)} images, embeddings {image_embeddings.shape}, on {device}")
 
@@ -106,7 +156,7 @@ def recommend(query_embedding, top_k=3, diversity_threshold=0.98):
 
 def recommend_from_image(input_image):
     if input_image is None:
-        return [], "✋ Please upload an image to find matching wonders."
+        return [], "Please upload an image to find matching wonders."
     query_emb = embed_image(input_image)
     results = recommend(query_emb, top_k=3)
     gallery_items = [
@@ -115,14 +165,14 @@ def recommend_from_image(input_image):
     ]
     medals = ["🥇", "🥈", "🥉"]
     summary = "Your top 3 wonder matches:\n\n" + "\n".join(
-        f"{medals[i]}  {r['label_name'].replace('_', ' ').title():<22}  similarity {r['score']:.3f}"
+        f"{medals[i]}  {r['label_name'].replace('_', ' ').title()}  —  similarity {r['score']:.3f}"
         for i, r in enumerate(results)
     )
     return gallery_items, summary
 
 def recommend_from_text(text_query):
     if not text_query or not text_query.strip():
-        return [], "✋ Please describe what you're looking for."
+        return [], "Please describe what you're looking for."
     query_emb = embed_text(text_query)
     results = recommend(query_emb, top_k=3)
     gallery_items = [
@@ -131,7 +181,7 @@ def recommend_from_text(text_query):
     ]
     medals = ["🥇", "🥈", "🥉"]
     summary = f'Best matches for "{text_query}":\n\n' + "\n".join(
-        f"{medals[i]}  {r['label_name'].replace('_', ' ').title():<22}  similarity {r['score']:.3f}"
+        f"{medals[i]}  {r['label_name'].replace('_', ' ').title()}  —  similarity {r['score']:.3f}"
         for i, r in enumerate(results)
     )
     return gallery_items, summary
@@ -148,13 +198,11 @@ CUSTOM_CSS = """
     font-family: 'Nunito', 'Quicksand', -apple-system, sans-serif !important;
 }
 
-/* Headings get the rounder, friendlier Quicksand */
 h1, h2, h3, h4 {
     font-family: 'Quicksand', sans-serif !important;
     letter-spacing: 0.3px !important;
 }
 
-/* ---------- HEADER ---------- */
 #header-block {
     background: linear-gradient(135deg, #8B4513 0%, #A0522D 50%, #CD853F 100%);
     padding: 36px 28px;
@@ -182,7 +230,6 @@ h1, h2, h3, h4 {
     opacity: 0.95;
 }
 
-/* ---------- TABS (the big upgrade) ---------- */
 .tab-nav {
     background: transparent !important;
     border-bottom: none !important;
@@ -221,7 +268,6 @@ h1, h2, h3, h4 {
     transform: translateY(-2px);
 }
 
-/* ---------- BUTTONS ---------- */
 button.primary, .gr-button-primary {
     background: linear-gradient(135deg, #8B4513 0%, #A0522D 100%) !important;
     border: none !important;
@@ -239,7 +285,6 @@ button.primary:hover, .gr-button-primary:hover {
     box-shadow: 0 6px 16px rgba(139, 69, 19, 0.45) !important;
 }
 
-/* ---------- INPUTS / PANELS ---------- */
 .gr-box, .gr-form, .gr-panel {
     background: #FFF8E7 !important;
     border: 2px solid #D2B48C !important;
@@ -276,7 +321,6 @@ textarea:focus, input[type="text"]:focus {
     padding: 10px !important;
 }
 
-/* ---------- FOOTER ---------- */
 #footer-block {
     margin-top: 28px;
     padding: 22px 24px;
@@ -298,9 +342,6 @@ textarea:focus, input[type="text"]:focus {
 }
 """
 
-# Build sample image indices defensively so a smaller dataset can't crash startup.
-SAMPLE_IDX = [i for i in [50, 2000, 5000, 7500, 10000] if i < len(full_ds)]
-
 with gr.Blocks(css=CUSTOM_CSS, theme=gr.themes.Soft(
     primary_hue="orange", secondary_hue="amber", neutral_hue="stone",
 ), title="Wonder Finder") as demo:
@@ -324,12 +365,6 @@ with gr.Blocks(css=CUSTOM_CSS, theme=gr.themes.Soft(
                 with gr.Column(scale=2):
                     img_gallery = gr.Gallery(label="Top 3 Matches", columns=3, rows=1, height=320, object_fit="cover")
                     img_summary = gr.Textbox(label="📊  Match Details", lines=6, show_copy_button=True)
-            if SAMPLE_IDX:
-                gr.Examples(
-                    examples=[[full_ds[i]["image"]] for i in SAMPLE_IDX],
-                    inputs=img_input,
-                    label="✨  Or try these sample images:",
-                )
             img_btn.click(recommend_from_image, inputs=img_input, outputs=[img_gallery, img_summary])
 
         with gr.Tab("💬  Search by Description"):
@@ -345,19 +380,6 @@ with gr.Blocks(css=CUSTOM_CSS, theme=gr.themes.Soft(
                 with gr.Column(scale=2):
                     text_gallery = gr.Gallery(label="Top 3 Matches", columns=3, rows=1, height=320, object_fit="cover")
                     text_summary = gr.Textbox(label="📊  Match Details", lines=6, show_copy_button=True)
-            gr.Examples(
-                examples=[
-                    ["ancient stone pyramid in the desert"],
-                    ["tall modern skyscraper at night"],
-                    ["waterfall in the tropical jungle"],
-                    ["ancient Roman amphitheater"],
-                    ["statue of a religious figure with outstretched arms"],
-                    ["a misty stone monument at sunrise"],
-                    ["white marble palace with a dome"],
-                ],
-                inputs=text_input,
-                label="✨  Or try these example queries:",
-            )
             text_btn.click(recommend_from_text, inputs=text_input, outputs=[text_gallery, text_summary])
 
     gr.HTML("""
@@ -369,6 +391,10 @@ with gr.Blocks(css=CUSTOM_CSS, theme=gr.themes.Soft(
     </div>
     """)
 
-
 if __name__ == "__main__":
-    demo.launch()
+    demo.launch(
+        server_name="0.0.0.0",
+        server_port=7860,
+        show_api=False,
+        share=False,
+    )

requirements.txt

@@ -1,10 +1,9 @@
-gradio==4.44.0
-pydantic==2.9.2
+gradio==4.44.1
+gradio-client==1.3.0
 transformers==4.45.2
 torch==2.4.1
 datasets==3.0.0
 pillow==10.4.0
 numpy==1.26.4
 pandas==2.2.2
-pyarrow==17.0.0
-huggingface-hub==0.25.0
+huggingface-hub==0.25.2