Design pattern for data augmentation

1

Hi all, a bit of a basic question here - I am trying to work with a somewhat small dataset, and apply augmentation techniques to increase my training set size. My DataLoader is currently very basic, of the form:

class Dataset(torch.utils.data.Dataset):
  
  def __init__(self, images_path):
    super(Dataset, self).__init__()
    self.images_path = images_path
  
  def __len__(self):
    return len(self.images_path)

  def __getitem__(self, idx):
    img_path = self.image_paths[idx]
    img = cv2.imread(str(img_path))
    return img

This Dataset doesn’t really suit itself well to augmentation, as images are iterated / specified by a specific index, whereas I want one single image to give rise to multiple training examples with different augmentations. I’m wondering what the usual solution in Pytorch is for this type of augmentation.

torchvisions.transforms is very well documented, but it talks about taking training samples and applying random transformations to them, rather than generating multiple samples. I always could do it manually, by applying random transformations to the batch multiple times at training time, but this seems ugly, and fails to spread out the same sample with different augmentations through different batches - is there an easy, Pytorchy solution for multiple training samples from a single image w/ augmentation?

2

You could transform the img multiple times and just return all of them. Inside the DataLoader loop you could then reshape the additional images to the batch dimension and reduce the batch_size of the DataLoader by the same factor.