A library which turns torchvision transformations invertible and replayable.
Project description
invertransforms
A library which turns torchvision transformations invertible and replayable.
Installation
pip install invertransforms
Usage
Simply replace previous torchvision import statements and enjoy the new features.
# from torchvision import transforms as T
import invertransforms as T
transform = T.Compose([
T.RandomCrop(size=256),
T.ToTensor(),
])
img_tensor = transform(img)
# invert
img_again = transform.invert(img_tensor)
# replay
img_tensor2 = transform.replay(img2)
# track
for i in range(n):
img_tensor_i = transform.track(img_i)
# ...
inverse_tf = transform.get_inverse(j) # or transform[j]
img_j = inverse_tf(img_tensor_j)
All transformations have an inverse
transformation attached to it.
inv_transform = transform.inverse()
img_inv = inv_transform(img)
Notes:
If a transformation is random, it is necessary to apply it once before calling invert
or inverse()
. Otherwise it will raise InvertibleError
.
On the otherhand, replay
can be called before, it will simply set the randomness on its first call.
One can create its own invertible transforms either by using the
practical Lambda
class function or by extending the Invertible
class available
in the invertransforms.lib
module.
Documentation
The library's documentation contains the full list of transformations which includes all the ones from torchvision and more.
Use Cases
This library can be particularly useful in following situations:
-
Reverting a NN-model output in order to stack predictions
-
Applying the same (random) transformations the same way on different inputs
Features
-
Invert any transformations, even random ones
-
Replay any transformations, even random ones
-
Track all transformations to invert them long after
-
All classes extend its torchvision transformation equivalent class. This means, you can just replace your previous torchvision import statements and it will not break your code.
-
Extensive unit testing (100% coverage, be safe, hopefully)
Requirements
python>=3.6
torch>=1.2.0
torchvision>=0.4.0
Future Improvements
-
[WIP] Extend the number of tranformations (e.g. random rotation and cropping (within the rotated area))
-
Make the transformations on tensors directly (data augmentation/transformation on GPU)
Contribute
You found a bug, think a feature is missing or just want to help ?
Please feel free to open an issue, pull request or contact me mail@gregunz.io
Project details
Download files
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.