"""
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
# ============================================================
print("Loading CLIP model...")
MODEL_NAME = "openai/clip-vit-base-patch32"
device = "cuda" if torch.cuda.is_available() else "cpu"
clip_model = CLIPModel.from_pretrained(MODEL_NAME).to(device)
clip_model.eval()
processor = CLIPProcessor.from_pretrained(MODEL_NAME)
print("Loading dataset...")
ds = load_dataset("chavajaz/wonders_dataset")
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)
print(f"Ready. {len(full_ds)} images, embeddings {image_embeddings.shape}, on {device}")
# ============================================================
# CORE FUNCTIONS
# ============================================================
@torch.no_grad()
def embed_image(pil_image):
img = pil_image.convert("RGB")
inputs = processor(images=img, return_tensors="pt").to(device)
feats = clip_model.get_image_features(**inputs)
if not isinstance(feats, torch.Tensor):
if hasattr(feats, "image_embeds") and feats.image_embeds is not None:
feats = feats.image_embeds
elif hasattr(feats, "pooler_output") and feats.pooler_output is not None:
feats = feats.pooler_output
else:
feats = feats[0]
feats = feats / feats.norm(dim=-1, keepdim=True)
return feats
@torch.no_grad()
def embed_text(text):
inputs = processor(text=[text], return_tensors="pt", padding=True, truncation=True).to(device)
feats = clip_model.get_text_features(**inputs)
if not isinstance(feats, torch.Tensor):
if hasattr(feats, "text_embeds") and feats.text_embeds is not None:
feats = feats.text_embeds
elif hasattr(feats, "pooler_output") and feats.pooler_output is not None:
feats = feats.pooler_output
else:
feats = feats[0]
feats = feats / feats.norm(dim=-1, keepdim=True)
return feats
def recommend(query_embedding, top_k=3, diversity_threshold=0.98):
sims = (query_embedding @ EMBEDDINGS_TENSOR.T).squeeze(0)
top_scores, top_indices = sims.topk(min(top_k * 20, len(sims)))
top_scores = top_scores.cpu().tolist()
top_indices = top_indices.cpu().tolist()
results = []
chosen_embeddings = []
for score, idx in zip(top_scores, top_indices):
candidate_emb = EMBEDDINGS_TENSOR[idx]
too_similar = any(
(candidate_emb @ prev_emb).item() > diversity_threshold
for prev_emb in chosen_embeddings
)
if too_similar:
continue
item = full_ds[idx]
results.append({
"index": idx,
"score": score,
"image": item["image"],
"label_name": class_names[item["label"]],
})
chosen_embeddings.append(candidate_emb)
if len(results) >= top_k:
break
return results
def recommend_from_image(input_image):
if input_image is None:
return [], "Please upload an image to find matching wonders."
query_emb = embed_image(input_image)
results = recommend(query_emb, top_k=3)
gallery_items = [
(r["image"], f"{r['label_name'].replace('_', ' ').title()} • match {r['score']*100:.1f}%")
for r in results
]
medals = ["🥇", "🥈", "🥉"]
summary = "Your top 3 wonder matches:\n\n" + "\n".join(
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."
query_emb = embed_text(text_query)
results = recommend(query_emb, top_k=3)
gallery_items = [
(r["image"], f"{r['label_name'].replace('_', ' ').title()} • match {r['score']*100:.1f}%")
for r in results
]
medals = ["🥇", "🥈", "🥉"]
summary = f'Best matches for "{text_query}":\n\n' + "\n".join(
f"{medals[i]} {r['label_name'].replace('_', ' ').title()} — similarity {r['score']:.3f}"
for i, r in enumerate(results)
)
return gallery_items, summary
# ============================================================
# UI
# ============================================================
CUSTOM_CSS = """
@import url('https://fonts.googleapis.com/css2?family=Quicksand:wght@400;500;600;700&family=Nunito:wght@400;600;700;800&display=swap');
.gradio-container {
background: linear-gradient(135deg, #F5EBDD 0%, #EDE0CC 100%) !important;
font-family: 'Nunito', 'Quicksand', -apple-system, sans-serif !important;
}
h1, h2, h3, h4 {
font-family: 'Quicksand', sans-serif !important;
letter-spacing: 0.3px !important;
}
#header-block {
background: linear-gradient(135deg, #8B4513 0%, #A0522D 50%, #CD853F 100%);
padding: 36px 28px;
border-radius: 20px;
margin-bottom: 28px;
box-shadow: 0 8px 24px rgba(139, 69, 19, 0.25);
text-align: center;
}
#header-block h1 {
color: #FFF8E7 !important;
font-size: 2.8em !important;
font-weight: 700 !important;
margin: 0 !important;
text-shadow: 2px 2px 4px rgba(0,0,0,0.2);
}
#header-block h3 {
color: #FFE4B5 !important;
font-weight: 500 !important;
margin: 10px 0 0 0 !important;
}
#header-block p {
color: #FFF8E7 !important;
margin-top: 14px !important;
font-size: 1.05em !important;
opacity: 0.95;
}
.tab-nav {
background: transparent !important;
border-bottom: none !important;
gap: 12px !important;
padding: 0 4px !important;
margin-bottom: 8px !important;
}
.tab-nav button {
background: #FFF8E7 !important;
border: 2px solid #D2B48C !important;
color: #8B4513 !important;
font-family: 'Nunito', sans-serif !important;
font-size: 1.15em !important;
font-weight: 700 !important;
padding: 14px 32px !important;
border-radius: 14px !important;
margin: 0 !important;
box-shadow: 0 2px 6px rgba(139, 69, 19, 0.12) !important;
transition: all 0.25s ease !important;
cursor: pointer !important;
}
.tab-nav button:hover {
background: #FFE8C8 !important;
border-color: #A0522D !important;
transform: translateY(-2px);
box-shadow: 0 4px 12px rgba(139, 69, 19, 0.25) !important;
}
.tab-nav button.selected {
background: linear-gradient(135deg, #8B4513 0%, #A0522D 100%) !important;
border-color: #8B4513 !important;
color: #FFF8E7 !important;
box-shadow: 0 6px 16px rgba(139, 69, 19, 0.4) !important;
transform: translateY(-2px);
}
button.primary, .gr-button-primary {
background: linear-gradient(135deg, #8B4513 0%, #A0522D 100%) !important;
border: none !important;
color: #FFF8E7 !important;
font-family: 'Nunito', sans-serif !important;
font-weight: 700 !important;
font-size: 1.08em !important;
padding: 14px 30px !important;
border-radius: 12px !important;
box-shadow: 0 4px 12px rgba(139, 69, 19, 0.3) !important;
transition: all 0.2s ease !important;
}
button.primary:hover, .gr-button-primary:hover {
transform: translateY(-2px);
box-shadow: 0 6px 16px rgba(139, 69, 19, 0.45) !important;
}
.gr-box, .gr-form, .gr-panel {
background: #FFF8E7 !important;
border: 2px solid #D2B48C !important;
border-radius: 14px !important;
}
label, .gr-input-label {
color: #5C4033 !important;
font-family: 'Nunito', sans-serif !important;
font-weight: 700 !important;
font-size: 1em !important;
}
textarea, input[type="text"] {
background: #FFFAF0 !important;
border: 2px solid #D2B48C !important;
color: #3E2723 !important;
font-family: 'Nunito', sans-serif !important;
font-size: 1.02em !important;
border-radius: 10px !important;
padding: 12px !important;
}
textarea:focus, input[type="text"]:focus {
border-color: #8B4513 !important;
outline: none !important;
box-shadow: 0 0 0 3px rgba(139, 69, 19, 0.15) !important;
}
.gr-gallery {
background: #FFF8E7 !important;
border: 2px solid #D2B48C !important;
border-radius: 14px !important;
padding: 10px !important;
}
#footer-block {
margin-top: 28px;
padding: 22px 24px;
background: rgba(139, 69, 19, 0.08);
border-radius: 14px;
border-left: 5px solid #8B4513;
color: #5C4033 !important;
font-family: 'Nunito', sans-serif !important;
line-height: 1.7;
}
#footer-block a {
color: #8B4513 !important;
font-weight: 700;
text-decoration: none;
border-bottom: 1px dashed #8B4513;
}
#footer-block a:hover {
color: #A0522D !important;
}
"""
with gr.Blocks(css=CUSTOM_CSS, theme=gr.themes.Soft(
primary_hue="orange", secondary_hue="amber", neutral_hue="stone",
), title="Wonder Finder") as demo:
gr.HTML("""
""")
with gr.Tabs():
with gr.Tab("📷 Search by Image"):
gr.Markdown("### Upload your travel photo, and we'll find the wonders that look most like it.")
with gr.Row():
with gr.Column(scale=1):
img_input = gr.Image(type="pil", label="Drop your photo here", height=320)
img_btn = gr.Button("✨ Find Similar Wonders", variant="primary", size="lg")
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)
img_btn.click(recommend_from_image, inputs=img_input, outputs=[img_gallery, img_summary])
with gr.Tab("💬 Search by Description"):
gr.Markdown("### Describe a place in your own words — CLIP translates language into visual matches.")
with gr.Row():
with gr.Column(scale=1):
text_input = gr.Textbox(
label="Describe a wonder",
placeholder='e.g. "an ancient stone temple in the jungle" or "a tall tower at sunset"',
lines=3,
)
text_btn = gr.Button("✨ Find Matching Wonders", variant="primary", size="lg")
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)
text_btn.click(recommend_from_text, inputs=text_input, outputs=[text_gallery, text_summary])
with gr.Accordion("🎬 Watch the project presentation", open=False):
gr.Video(value="Video.mp4", label=None, show_label=False)
gr.HTML("""
""")
if __name__ == "__main__":
demo.launch(
server_name="0.0.0.0",
server_port=7860,
show_api=False,
share=False,
)