import gradio as gr import torch import yaml import numpy as np from PIL import Image from roboflow import Roboflow from ultralytics import YOLO def load_model(file_path): # Load the Roboflow model rf = Roboflow(api_key="K1TXQnJq7EE7yoCf1g3C") project = rf.workspace("fyp-l87nq").project("bone-fracture-detection-rkuqr") model = project.version(3).model # Load the model weights into a PyTorch model pytorch_model = YOLO('data.yaml') pytorch_model.load_state_dict(torch.load(file_path, map_location=torch.device('cpu'))) pytorch_model.eval() return pytorch_model file_path = 'yolov8s.pt' model = load_model(file_path) def predict_fracture(image): # Preprocess the image for the Roboflow model img = Image.fromarray(image) img_tensor = to_tensor(img).unsqueeze(0) # Convert image to tensor and add batch dimension # Perform inference with the Roboflow model with torch.no_grad(): output = model(img_tensor) # Postprocess the inference output results = output[0] img_with_boxes = image.copy() for box in results: label = int(box[5]) score = float(box[4]) if label == 0: # Assuming 0 corresponds to the bone fracture class color = "red" if score > 0.5 else "orange" # Adjust the threshold as needed xmin, ymin, xmax, ymax = box[:4].int().tolist() img_with_boxes.rectangle([xmin, ymin, xmax, ymax], outline=color, width=2) img_with_boxes.text((xmin, ymin), f"Fracture: {score:.2f}", font_size=12, color=color) return Image.fromarray((np.uint8(img_with_boxes))) # Define the to_tensor function def to_tensor(image): image = np.array(image) / 255.0 return torch.from_numpy(image.transpose((2, 0, 1))).float() # Gradio Interface iface = gr.Interface( fn=lambda *args, **kwargs: predict_fracture(args[0], load_model), inputs=gr.Image(), outputs=gr.Image(), live=True, title="Bone Fracture Detection", description="Upload an X-ray image to detect bone fractures using Roboflow's YOLOv8 model.", ) iface.launch()