Transforms and data augmentation: use Kornia

[remtav]

Kornia offers great transforms for data augmentation. GDL could replace its in-house augmentations from augmentation.py for Kornia's ones. Initially, the problem was that torchvision's transforms were based on PIL and did not accept more than 3 bands. That is not the case with Kornia as it is much more flexible and offers lots of features (GPU-support, apply transforms on entire batch, etc.) that our in-house augmentations don't offer.

[remtav]

(draft)

CURRENT - Torchvision approach

Instantiation:

• SegmentationDataset()
• transforms as torchvision.Compose object (with list of custom transforms classes each including minimally an init and call methods)
  • ends with a ToTensorTarget() converting numpy array to pytorch tensor (and a bit more...)

During training:

1. SegmentationDataset's get_item returns a "sample" as dict contianing numpy arrays (keys: 'sat_img', 'map_img')
2. transforms are applied sequentially
   • input: sample (dict with values as np.arrays)
   • output: sample (dict with values as torch.Tensors)

PROPOSED - translate to Kornia

Requirements:

• input as torch.Tensor, can be dict with values as torch.Tensors if working with torchgeo's version of AugmentationSequential
• all transforms, custom or not, must be subclasses from nn.Module

Instantiation:

• SegmentationDataset()
• transforms as list of transforms passed to torchgeo.transforms.AugmentationSequential()
  • Radiometric transforms
  • Geometric transforms

Implementation: SegmentationDataset: _load_image() in get_item SegmentationDataset: _load_target() in get_item SegmentationDataset: subclass torchgeo's VisionDataset! augmentations.py: class Augmentation(), which accepts variable length list of transforms as nn.Module (ex.: Intensity)

Test:

1. Test with GDL's input (dict of numpy arrays) but with Kornia's transforms inside a torchgeo.transforms.AugmentationSequential()
2. Returns transformed sample with same keys?