Jian Shi

It gonna be `AugmentationBase2D(_BasicAugmentationBase, GenericAugmentationBase)` and `AugmentationSequential(ImageSequential, GenericAugmentationBase)`

> I would really love such an implementation, it would make it so much easier to truly use kornia for keypoint augmentation which is a mess now with unsupported transformations...

LGTM. Can you add some tests that address the failing scenarios?

By having a dtype flag for each test case "def test(...,dtype)", then you may run "pytest --dtype float16" to do that. But you may implement yourself for the AMP stuff....

> Is there any inplace ops happening inside @shijianjian ? Yes. All the bbox operations are inplace right now.

Hi @copaah, Thanks for the proposal. Those losses are common and not only for stereo images. Since we are a computer vision library, I personally think we should include losses...

Hi @copaah, Is the MeanBad{n}Error more close to be stereo-only? I think probably we can include some functions around computing disparity maps? @edgarriba what do you think? Best, Jian

I would avoid using names like ```metrics``` and ```losses``` since they look like more DL related., probably go for ```kornia.measure.stereo```. BTW, what is the plan on the stereo? To have...

> @shijianjian : What is the motivation behind avoiding DL nomenclature? Not a strong objection. The term of "loss" in DL is more or less ready to be called with...

TorchVision augmentation indeed can perform in a batch but the exact same augmentation will be performed on all the images across the whole batch, while in kornia, we perform different...