[feat] Added inverse for augmentations. #1013
Conversation
|
Let's get this merged if the API is okay for everyone so that we can include it in the release for next week /cc @vadimkantorov |
kornia/augmentation/augmentation.py
Outdated
| def inverse_transform( | ||
| self, input: torch.Tensor, transform: Optional[torch.Tensor] = None, | ||
| size: Optional[Tuple[int, int]] = None, **kwargs | ||
| ) -> torch.Tensor: | ||
| return hflip(input) |
There was a problem hiding this comment.
Should we be based on transformation matrix for hlip and vflip? The direct hlip will be much faster.
There was a problem hiding this comment.
how different is this from what we have already ?
There was a problem hiding this comment.
No no. It is for the inverse. restore back the flip that has been applied.
kornia/augmentation/augmentation.py
Outdated
| def inverse_transform( | ||
| self, input: torch.Tensor, transform: Optional[torch.Tensor] = None, | ||
| size: Optional[Tuple[int, int]] = None, **kwargs | ||
| ) -> torch.Tensor: |
There was a problem hiding this comment.
could you add some docs here ? what are the kwarg ? could you make it explicit ?
There was a problem hiding this comment.
I was thinking that if it is needed to overwrite some of the parameters during the inverse conversion.
Say, I used bilinear for the forward but I may want to use nearest for inverse? I have not decided do I need to do some overwriting here. I need some real-world feedback.
kornia/augmentation/augmentation.py
Outdated
| align_corners = self.align_corners if "align_corners" not in kwargs else kwargs['align_corners'] | ||
| padding_mode = 'zeros' if "padding_mode" not in kwargs else kwargs['padding_mode'] | ||
| transform = cast(torch.Tensor, transform) | ||
| return affine(input, transform.inverse()[..., :2, :3], mode, padding_mode, align_corners) |
There was a problem hiding this comment.
remove boilerplate code and reuse in both places the affine function (or use directly warp_affine)
| ) -> torch.Tensor: | ||
| if self.cropping_mode != 'resample': | ||
| raise NotImplementedError(f"`inverse` is only applicable for resample cropping mode. Got {self.cropping_mode}.") | ||
| mode = self.resample.name.lower() if "mode" not in kwargs else kwargs['mode'] |
There was a problem hiding this comment.
a lot of boilerplate across augmentations for the flags too
| to the batch form (False). Default: False. | ||
| """ | ||
|
|
||
| def compute_transformation(self, input: torch.Tensor, params: Dict[str, torch.Tensor]) -> torch.Tensor: |
There was a problem hiding this comment.
I guess params could Optional
| params = self._params | ||
| if 'batch_prob' not in params: | ||
| params['batch_prob'] = torch.tensor([True] * batch_shape[0]) | ||
| warnings.warn("`batch_prob` is not found in params. Will assume applying on all data.") |
There was a problem hiding this comment.
do you have any mechanism to make warnings optional ?
edgarriba
added
enhancement 🚀
|
Should be okay now. |
Description
Related to #1000.
@edgarriba, @ducha-aiki I am not sure if we should enable the inverse operation for perspective and cropping transforms?