Upload 2 files

Adding the main files

app.py

@@ -0,0 +1,309 @@
+#!/usr/bin/env python3
+"""
+GeoVLM - AI-Powered Geolocation
+Upload any image and predict where it was taken using Vision-Language Models
+"""
+
+import gradio as gr
+from PIL import Image
+from transformers import AutoProcessor, AutoModelForImageTextToText
+import torch
+import re
+import math
+from dataclasses import dataclass
+
+# ============================================================================
+# Simplified Geolocation Parser (from vlm-gym)
+# ============================================================================
+
+@dataclass(frozen=True)
+class Coords:
+    """Geographic coordinates"""
+    lat: float
+    lon: float
+
+@dataclass(frozen=True)
+class ParsedResponse:
+    """Structured model output"""
+    city: str | None
+    region: str | None
+    country: str | None
+    coords: Coords | None
+    raw_text: str
+    format_valid: bool
+
+PROMPT_TEMPLATE = (
+    "Look at the image and guess the location.\n"
+    "Respond with EXACTLY these 5 lines, no extra text:\n"
+    "City: <city name>\n"
+    "Region: <state or region>\n"
+    "Country: <country name or ISO-2 code>\n"
+    "Latitude: <number between -90 and 90>\n"
+    "Longitude: <number between -180 and 180>\n"
+)
+
+KEY_ALIASES = {
+    "city": "city",
+    "country": "country",
+    "region": "region",
+    "state": "region",
+    "province": "region",
+    "latitude": "lat",
+    "lat": "lat",
+    "longitude": "lon",
+    "lon": "lon",
+}
+
+def parse_response(text: str) -> ParsedResponse:
+    """Parse structured 5-line format"""
+    parsed = {}
+    
+    if not text:
+        return ParsedResponse(None, None, None, None, text, False)
+    
+    # Parse key-value lines
+    key_pattern = re.compile(
+        r'^\s*(?:[-*+\u2022]\s*)?(?P<key>[A-Za-z][A-Za-z0-9\s\-/_.]*?)\s*:\s*(?P<value>.+)$'
+    )
+    
+    for line in text.splitlines():
+        match = key_pattern.match(line)
+        if not match:
+            continue
+        
+        key_raw = match.group("key").strip().lower()
+        key_raw = key_raw.strip("*_`\"' ")
+        key_raw = re.sub(r"\s+", " ", key_raw)
+        canonical = KEY_ALIASES.get(key_raw)
+        
+        if canonical is None:
+            continue
+        
+        value_raw = match.group("value").strip()
+        value_raw = value_raw.strip("`\"' \t")
+        value_raw = re.sub(r"^[*_`]+", "", value_raw)
+        value_raw = re.sub(r"[*_`]+$", "", value_raw)
+        value_raw = value_raw.strip()
+        
+        if canonical in {"city", "region", "country"}:
+            if value_raw and canonical not in parsed:
+                parsed[canonical] = value_raw
+        elif canonical in {"lat", "lon"}:
+            if canonical not in parsed:
+                match_num = re.search(r"-?\d+(?:[.,]\d+)?", value_raw)
+                if match_num:
+                    try:
+                        parsed[canonical] = float(match_num.group(0).replace(",", "."))
+                    except ValueError:
+                        pass
+    
+    # Build coords if available
+    coords = None
+    if "lat" in parsed and "lon" in parsed:
+        try:
+            lat = parsed["lat"]
+            lon = parsed["lon"]
+            if -90 <= lat <= 90 and -180 <= lon <= 180:
+                coords = Coords(lat=lat, lon=lon)
+        except (ValueError, TypeError):
+            pass
+    
+    format_valid = bool(len(parsed) >= 2)
+    
+    return ParsedResponse(
+        city=parsed.get("city"),
+        region=parsed.get("region"),
+        country=parsed.get("country"),
+        coords=coords,
+        raw_text=text,
+        format_valid=format_valid,
+    )
+
+# ============================================================================
+# Model Setup
+# ============================================================================
+
+model = None
+processor = None
+MODEL_NAME = "Qwen/Qwen2-VL-2B-Instruct"
+
+def load_model():
+    """Load model once on startup"""
+    global model, processor
+    if model is None:
+        print(f"Loading model: {MODEL_NAME}")
+        processor = AutoProcessor.from_pretrained(MODEL_NAME)
+        model = AutoModelForImageTextToText.from_pretrained(
+            MODEL_NAME,
+            torch_dtype=torch.float16 if torch.cuda.is_available() else torch.float32,
+            device_map="auto" if torch.cuda.is_available() else "cpu"
+        )
+        print("Model loaded successfully!")
+
+def predict_location(image):
+    """Predict geolocation from an image"""
+    if image is None:
+        return "Please upload an image.", ""
+    
+    # Ensure model is loaded
+    load_model()
+    
+    # Convert to PIL if needed
+    if not isinstance(image, Image.Image):
+        image = Image.fromarray(image).convert("RGB")
+    else:
+        image = image.convert("RGB")
+    
+    # Prepare prompt
+    messages = [
+        {
+            "role": "user",
+            "content": [
+                {"type": "image"},
+                {"type": "text", "text": PROMPT_TEMPLATE}
+            ]
+        }
+    ]
+    
+    # Process inputs
+    text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
+    inputs = processor(text=[text], images=[image], return_tensors="pt", padding=True)
+    
+    # Move to device
+    inputs = {k: v.to(model.device) for k, v in inputs.items()}
+    
+    # Generate
+    with torch.no_grad():
+        output_ids = model.generate(
+            **inputs,
+            max_new_tokens=256,
+            do_sample=False,
+        )
+    
+    # Decode
+    generated_ids = output_ids[0][inputs['input_ids'].shape[1]:]
+    response = processor.decode(generated_ids, skip_special_tokens=True).strip()
+    
+    # Parse
+    parsed = parse_response(response)
+    
+    # Format output
+    output = f"""
+## 🤖 Raw Model Response:
+```
+{response}
+```
+
+---
+
+## 📍 Parsed Prediction:
+
+**City:** {parsed.city or "Not provided"}  
+**Region:** {parsed.region or "Not provided"}  
+**Country:** {parsed.country or "Not provided"}  
+**Coordinates:** {f"{parsed.coords.lat:.6f}, {parsed.coords.lon:.6f}" if parsed.coords else "Not provided"}  
+**Format Valid:** {"✅ Yes" if parsed.format_valid else "❌ No"}
+"""
+    
+    # Create map embed
+    map_html = ""
+    if parsed.coords:
+        map_html = f"""
+        <div style="margin-top: 20px;">
+            <iframe 
+                width="100%" 
+                height="450" 
+                frameborder="0" 
+                scrolling="no" 
+                marginheight="0" 
+                marginwidth="0" 
+                src="https://www.openstreetmap.org/export/embed.html?bbox={parsed.coords.lon-0.1},{parsed.coords.lat-0.1},{parsed.coords.lon+0.1},{parsed.coords.lat+0.1}&marker={parsed.coords.lat},{parsed.coords.lon}"
+                style="border: 2px solid #ddd; border-radius: 8px;">
+            </iframe>
+            <div style="margin-top: 10px; text-align: center;">
+                <a href="https://www.google.com/maps?q={parsed.coords.lat},{parsed.coords.lon}" target="_blank" style="margin: 0 10px; color: #4285f4; text-decoration: none; font-weight: bold;">
+                    🗺️ View on Google Maps
+                </a>
+                <span style="color: #666;">|</span>
+                <a href="https://www.openstreetmap.org/?mlat={parsed.coords.lat}&mlon={parsed.coords.lon}#map=12/{parsed.coords.lat}/{parsed.coords.lon}" target="_blank" style="margin: 0 10px; color: #7ebc6f; text-decoration: none; font-weight: bold;">
+                    🌍 View on OpenStreetMap
+                </a>
+            </div>
+        </div>
+        """
+    else:
+        map_html = "<div style='text-align: center; padding: 20px; color: #666;'>No valid coordinates found</div>"
+    
+    return output, map_html
+
+# ============================================================================
+# Gradio Interface
+# ============================================================================
+
+with gr.Blocks(title="GeoVLM - AI Geolocation", theme=gr.themes.Soft()) as demo:
+    gr.Markdown(
+        """
+        # 🌍 GeoVLM - AI-Powered Geolocation
+        
+        Upload any image and let AI predict where it was taken using vision-language models!
+        
+        ### How it works:
+        - Analyzes visual features: architecture, vegetation, road signs, landscape
+        - Uses state-of-the-art vision-language models (Qwen2-VL)
+        - Predicts city, region, country, and GPS coordinates
+        
+        **Powered by [vlm-gym](https://github.com/sdan/vlm-gym)** | Model: Qwen2-VL-2B-Instruct
+        """
+    )
+    
+    with gr.Row():
+        with gr.Column(scale=1):
+            image_input = gr.Image(
+                type="pil",
+                label="📸 Upload Image",
+                height=400
+            )
+            predict_btn = gr.Button("🔍 Predict Location", variant="primary", size="lg")
+            
+            gr.Markdown(
+                """
+                ### 💡 Tips:
+                - Outdoor images work best
+                - Street views are ideal
+                - Clear photos with visible landmarks
+                - Unique architectural or natural features help
+                """
+            )
+        
+        with gr.Column(scale=1):
+            output_text = gr.Markdown(label="Results")
+            map_output = gr.HTML(label="Map")
+    
+    gr.Markdown(
+        """
+        ---
+        ### 🎯 Use Cases:
+        - **OSINT Research** - Verify photo locations for investigations
+        - **GeoGuessr Training** - Practice location identification
+        - **Education** - Learn about geographic features and cultures
+        - **Travel Planning** - Identify interesting locations from photos
+        
+        ---
+        
+        **Note:** This is a demo. Predictions may not always be accurate. Use responsibly for educational and research purposes.
+        
+        Built with ❤️ using [Gradio](https://gradio.app) and [Hugging Face Transformers](https://huggingface.co/transformers)
+        """
+    )
+    
+    # Event handlers
+    predict_btn.click(
+        fn=predict_location,
+        inputs=image_input,
+        outputs=[output_text, map_output]
+    )
+
+if __name__ == "__main__":
+    print("🚀 Starting GeoVLM...")
+    load_model()
+    demo.launch()

requirements.txt

@@ -0,0 +1,7 @@
+gradio
+torch
+torchvision
+transformers
+accelerate
+Pillow
+qwen-vl-utils