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
os.makedirs("models", exist_ok=True)

# Select device
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")

# Use YOLOv5 Nano for speed
model_path = Path("models/yolov5n.pt")

if model_path.exists():
    print(f"Loading model from cache: {model_path}")
    model = torch.hub.load("ultralytics/yolov5", "custom", path=str(model_path), source="local").to(device)
else:
    print("Downloading YOLOv5n model and caching...")
    model = torch.hub.load("ultralytics/yolov5", "yolov5n", pretrained=True).to(device)
    torch.save(model.state_dict(), model_path)

# Optimize model for speed
model.conf = 0.3  # Confidence threshold
model.iou = 0.3   # IoU threshold for Non-Maximum Suppression (NMS)
model.classes = None  # Detect all classes
model.eval()

if device.type == "cuda":
    print("Using FP16 precision for inference (high speed, lower accuracy)")
    model.half()  # Enable FP16 for faster inference

torch.set_num_threads(os.cpu_count())  # Optimize CPU threading

# Pre-generate colors for bounding boxes
np.random.seed(42)
colors = np.random.uniform(0, 255, size=(len(model.names), 3))

# FPS tracking
total_inference_time = 0
inference_count = 0

def detect_objects(image):
    global total_inference_time, inference_count

    if image is None:
        return None

    start_time = time.time()

    # Convert image to BGR format for OpenCV
    image_bgr = cv2.cvtColor(image, cv2.COLOR_RGB2BGR)

    # Get image dimensions
    h, w, _ = image.shape

    with torch.inference_mode():  # Faster than torch.no_grad()
        results = model(image_bgr)

    inference_time = time.time() - start_time
    total_inference_time += inference_time
    inference_count += 1
    avg_inference_time = total_inference_time / inference_count

    detections = results.xyxy[0].cpu().numpy()  # Use xyxy format

    output_image = image.copy()

    for *xyxy, conf, cls in detections:
        x1, y1, x2, y2 = map(int, xyxy)
        class_id = int(cls)
        color = colors[class_id].tolist()

        # Keep bounding boxes within image bounds
        x1, y1, x2, y2 = max(0, x1), max(0, y1), min(w, x2), min(h, y2)

        # Draw bounding box
        cv2.rectangle(output_image, (x1, y1), (x2, y2), color, 3, lineType=cv2.LINE_AA)

        label = f"{model.names[class_id]} {conf:.2f}"
        font_scale, font_thickness = 0.9, 2
        (tw, th), _ = cv2.getTextSize(label, cv2.FONT_HERSHEY_SIMPLEX, font_scale, font_thickness)

        # Label background
        cv2.rectangle(output_image, (x1, y1 - th - 10), (x1 + tw + 10, y1), color, -1)
        cv2.putText(output_image, label, (x1 + 5, y1 - 5),
                    cv2.FONT_HERSHEY_SIMPLEX, font_scale, (255, 255, 255), font_thickness, lineType=cv2.LINE_AA)

    fps = 1 / inference_time

    # Display FPS
    overlay = output_image.copy()
    cv2.rectangle(overlay, (10, 10), (300, 80), (0, 0, 0), -1)
    output_image = cv2.addWeighted(overlay, 0.6, output_image, 0.4, 0)
    cv2.putText(output_image, f"FPS: {fps:.2f}", (20, 40),
                cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2, lineType=cv2.LINE_AA)
    cv2.putText(output_image, f"Avg FPS: {1/avg_inference_time:.2f}", (20, 70),
                cv2.FONT_HERSHEY_SIMPLEX, 1, (0, 255, 0), 2, lineType=cv2.LINE_AA)

    return output_image

# Gradio UI
example_images = ["spring_street_after.jpg", "pexels-hikaique-109919.jpg"]
os.makedirs("examples", exist_ok=True)

with gr.Blocks(title="YOLOv5 Object Detection (High Quality, High FPS)") as demo:
    gr.Markdown("""
    # YOLOv5 Object Detection - High Quality & High FPS  
    Detects objects with full-resolution output and ultra-fast performance.
    """)

    with gr.Row():
        with gr.Column(scale=1):
            input_image = gr.Image(label="Input Image", type="numpy")
            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")

    gr.Examples(
        examples=example_images,
        inputs=input_image,
        outputs=output_image,
        fn=detect_objects,
        cache_examples=True
    )

    submit_button.click(fn=detect_objects, inputs=input_image, outputs=output_image)
    clear_button.click(lambda: (None, None), None, [input_image, output_image])

demo.launch()