[ROADMAP] Kornia v0.1.5

This is the roadmap for kornia v0.1.5 focuses mainly importing in the data augmentation module.

Feel free to volunteer yourself if you are interested in trying out some items(they do not have to be on the list). Please, take a read on the CONTRIBUTING notes.

• Data augmentation https://github.com/kornia/kornia/issues/298 -> ticket PR https://github.com/kornia/kornia/pull/309 Make differentiable versions of albumentations, torchvision. Please propose and we will assign tasks.

  • BaseAugmentation @kshitij12345 @edgarriba d729d7c4357ca73e4915a42285a0771bca4436ce
  • Denormalize @dvd42 0972e55b221d4d2cedd3c44c72ac55dfe7bd5d50
  • ColorJitter @priba 98b875289cc05e61fc12bcf8775640f57fd4d273
  • RandomHorizontalFlip @dvd42 ba1e89c6d7ff0a38bb171c0cfbd7bdc7726018c8
  • MotionBlur @mronta 63a34dfbbbc2fcf1bcbfe94547ad6c29896058b9
  • RandomVerticalFlip @kshitij f6f1c1164805a8b60a4e5f28d0dc9146360c07fd
  • RandomErasing @mronta 85ef4c0a1e58a023dbae840b233677b37ac32fdc
  • RandomGrayscale @priba 1782eb35bc61600df65c17087c7a03e3aa93d1ee
  • Resize @edgarriba d0f596eee9d49d8a5cdd46d9ced9d02ae80423ab
  • CenterCrop @edgarriba https://github.com/kornia/kornia/pull/409
  • RandomAffine @edgarriba https://github.com/kornia/kornia/pull/403
  • RandomPerspective @edgarriba https://github.com/kornia/kornia/pull/403
  • RandomRotation @dvd42 https://github.com/kornia/kornia/pull/397
  • RandomResizedCrop @kshitij https://github.com/kornia/kornia/pull/408
  • Grayscale @dvd42
    • call rgb_to_grayscale + https://pytorch.org/docs/stable/torchvision/transforms.html#torchvision.transforms.Grayscale

• Color transforms Match missing OpenCV conversion. Please propose !

  • rgb_to_ycbcr @justanhduc a0bafcdc8f959ad1ddcae5b050a8d786e1e66000
  • ycbcr_to_rgb @justanhduc a0bafcdc8f959ad1ddcae5b050a8d786e1e66000
  • rgb_to_yuv @ananyamukh6 eefca4f443703c133d4276b83d0f09aae85042de
  • yuv_to_rgb @ananyamukh6 eefca4f443703c133d4276b83d0f09aae85042de
  • rgb_to_rgba @ucalyptus @edgarriba e6689bec210c23f385d076e6dee2bcd47272d5de
  • rgba_to_rgb @ucalyptus @edgarriba e6689bec210c23f385d076e6dee2bcd47272d5de
  • bgr_to_bgra @ucalyptus @edgarriba e6689bec210c23f385d076e6dee2bcd47272d5de
  • bgra_to_bgr @ucalyptus @edgarriba e6689bec210c23f385d076e6dee2bcd47272d5de
  • lab_to_rgb https://github.com/kornia/kornia/issues/239 @HassanAlsamahi
  • rgb_to_lab https://github.com/kornia/kornia/issues/239 @HassanAlsamahi
  • rgb_to_xyz @priba
  • xyz_to_rgb @priba
  • rgb_to_luv @priba
  • luv_to_rgb @priba

• GPU Tests

  • parametrize test functions to accept torch.device cpu/cuda @edgarriba @ducha-aiki da793cd48e17eca7e78b042240a9464065611d9d 0fcb85e25f0d37042ff426a2541cfb296b03cd30
  • setup CI @edgarriba @chadell

• Framework

  • Update to PyTorch 1.4.0 @edgarriba https://github.com/kornia/kornia/pull/402

[shijianjian]

For data augmentation tasks, I think it is better to mimic PyTorch style implementation for better reproducibility.

Take the ColorJitter as an example, the current implementation cannot reproduce the same result via a function call. Currently, we have: 1. colorJitter method, which will generate random params and apply function. 2. ColorJitter class, which does nothing but calls the colorJitter method.

For better reproducibility, I think the PyTorch style is more clear that for ColorJitter class, it will have a get_params static method that generates random params with a def __call__() implementation. Thus, the colorJitter method will apply the transformations only.

One of the use cases of reproducing the transformations can be for segmentation tasks, that we will wish to apply the exact same affine transformations for both input image and mask.

[ananyamukh6]

Volunteering for Morphology, some color conversions.