app.py

import torch
import numpy as np
import gradio as gr
import cv2
import time
import os
from pathlib import Path

# Create cache directory for models if it doesn't exist
os.makedirs("models", exist_ok=True)

# Check device availability - Hugging Face Spaces often provides GPU
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")

# Load YOLOv5x model with caching for faster startup
model_path = Path("models/yolov5x.pt")
if model_path.exists():
    print(f"Loading model from cache: {model_path}")
    model = torch.hub.load("ultralytics/yolov5", "yolov5x", pretrained=True, 
                          source="local", path=str(model_path)).to(device)
else:
    print("Downloading YOLOv5x model and caching...")
    model = torch.hub.load("ultralytics/yolov5", "yolov5x", pretrained=True).to(device)
    # Cache the model for faster startup next time
    torch.save(model.state_dict(), model_path)

# Optimization configurations
model.conf = 0.3  # Confidence threshold of 0.3 as specified
model.iou = 0.3   # NMS IoU threshold of 0.3 as specified
model.classes = None  # Detect all 80+ COCO classes

# Optimize for GPU if available
if device.type == "cuda":
    # Use mixed precision for performance boost
    model.half()
else:
    # On CPU, optimize operations
    torch.set_num_threads(os.cpu_count())

# Set model to evaluation mode for inference
model.eval()

# Assign fixed colors to each class for consistent visualization
np.random.seed(42)  # For reproducible colors
colors = np.random.uniform(0, 255, size=(len(model.names), 3))

# Track performance metrics
total_inference_time = 0
inference_count = 0

def detect_objects(image):
    """
    Process input image for object detection using YOLOv5
    
    Args:
        image: Input image as numpy array
        
    Returns:
        output_image: Image with detection results visualized
    """
    global total_inference_time, inference_count
    
    if image is None:
        return None
    
    start_time = time.time()
    
    # Create a copy for drawing results
    output_image = image.copy()
    
    # Fixed input size for optimal processing
    input_size = 640
    
    # Perform inference with no gradient calculation
    with torch.no_grad():
        # Convert image to tensor for faster processing
        results = model(image, size=input_size)
    
    # Record inference time (model processing only)
    inference_time = time.time() - start_time
    total_inference_time += inference_time
    inference_count += 1
    avg_inference_time = total_inference_time / inference_count
    
    # Extract detections from first (and only) image
    detections = results.pred[0].cpu().numpy()
    
    # Draw each detection on the output image
    for *xyxy, conf, cls in detections:
        # Extract coordinates and convert to integers
        x1, y1, x2, y2 = map(int, xyxy)
        class_id = int(cls)
        
        # Get color for this class
        color = colors[class_id].tolist()
        
        cv2.rectangle(output_image, (x1, y1), (x2, y2), color, 4)
        
        # Create label with class name and confidence score
        label = f"{model.names[class_id]} {conf:.2f}"
        
        font_scale = 0.8
        font_thickness = 2
        
        (w, h), _ = cv2.getTextSize(label, cv2.FONT_HERSHEY_SIMPLEX, font_scale, font_thickness)
  
        cv2.rectangle(output_image, (x1, y1 - h - 10), (x1 + w + 10, y1), color, -1)
        
        cv2.putText(output_image, label, (x1 + 5, y1 - 5),
                    cv2.FONT_HERSHEY_SIMPLEX, font_scale, (0, 0, 0), font_thickness + 1)
        cv2.putText(output_image, label, (x1 + 5, y1 - 5),
                    cv2.FONT_HERSHEY_SIMPLEX, font_scale, (255, 255, 255), font_thickness)
    
    # Calculate FPS
    fps = 1 / inference_time
    
    fps_overlay = output_image.copy()
    cv2.rectangle(fps_overlay, (5, 5), (250, 80), (0, 0, 0), -1)
    # Apply the overlay with transparency
    alpha = 0.7
    output_image = cv2.addWeighted(fps_overlay, alpha, output_image, 1 - alpha, 0)
    
    # Display FPS with larger font
    cv2.putText(output_image, f"FPS: {fps:.2f}", (10, 35),
                cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)
    cv2.putText(output_image, f"Avg FPS: {1/avg_inference_time:.2f}", (10, 70),
                cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2)
    
    return output_image

# Define example images - these will be stored in the same directory as this script
example_images = [
    "examples/spring_street_after.jpg", 
    "examples/pexels-hikaique-109919.jpg"
]

# Make sure example directory exists
os.makedirs("examples", exist_ok=True)

# Create Gradio interface - optimized for Hugging Face Spaces
with gr.Blocks(title="Optimized YOLOv5 Object Detection") as demo:
    gr.Markdown("""
    # Optimized YOLOv5 Object Detection
    
    This system utilizes YOLOv5 to detect 80+ object types from the COCO dataset.
    
    **Performance Features:**
    - Processing speed: Optimized for 30+ FPS at 640x640 resolution
    - Confidence threshold: 0.3
    - IoU threshold: 0.3
    
    Simply upload an image and click Submit to see the detections!
    """)
    
    with gr.Row():
        with gr.Column(scale=1):
            input_image = gr.Image(label="Input Image", type="numpy")
            with gr.Row():
                submit_button = gr.Button("Submit", variant="primary")
                clear_button = gr.Button("Clear")
                
        with gr.Column(scale=1):
            output_image = gr.Image(label="Detected Objects", type="numpy")
    
    # Example gallery
    gr.Examples(
        examples=example_images,
        inputs=input_image,
        outputs=output_image,
        fn=detect_objects,
        cache_examples=True  # Cache for faster response
    )
    
    # Set up button event handlers
    submit_button.click(fn=detect_objects, inputs=input_image, outputs=output_image)
    clear_button.click(lambda: None, None, [input_image, output_image])
    

# Launch for Hugging Face Spaces
demo.launch()