To optimize this code to run on the GPU, we can make the following changes:

1. Convert the numpy arrays to PyTorch tensors before sending them to the GPU.
2. Use PyTorch's built-in image resizing and normalization functions instead of using OpenCV.
3. Remove the unnecessary numpy operations.

Here's the optimized code:

from torchvision import transforms import torch

class ImagePreprocess: def init(self, device="cpu"): self.transform = transforms.Compose([ transforms.Resize((img_size, img_size)), transforms.ToTensor(), transforms.Normalize((0.485, 0.456, 0.406), (0.229, 0.224, 0.225)) ]) self.device = device

def preprocess(self, imgs, img_size):
    imgs_tensor = torch.stack([self.transform(img) for img in imgs]).to(self.device)
    return imgs_tensor.cpu().detach().numpy()

Example usage

preprocessor = ImagePreprocess(device="cuda") imgs = [cv2.imread(img_path) for img_path in img_paths] preprocessed_imgs = preprocessor.preprocess(imgs, img_size=224)