This project has been merged into https://github.com/arc-community/arc and can be used by installing the arc package
requires https://github.com/arc-community/arc installed
is interoperable with torchvision transforms so you can do things like:
to install, clone the repo and run pip install -e . in the root directory of the repo
from torchvision import transforms
transform = transforms.compose([
transforms.RandomOrder([
RandomCropInputAndOuput(0.5, same_aug_for_all_pairs=True),
RandomDoubleInputBoard(0.5, same_aug_for_all_pairs=True),
]),
RandomColor(0.5, same_aug_for_all_pairs=True)
])
aug_training_riddles = transform(training_riddles)
See the demonstrations notebook for more details.
Each augmentation class has a corresponding function in arcmentations.functional that take in params to do a specific non random transformation. which is also useful for people that just want to use the same augmentation for all their data.
you can sample params for a specific function in arcmentations.functional by calling the Random* class's static method get_params()
for example:
from arcmentations.augmentations import RandomCropInputAndOuput
params = RandomCropInputAndOuput.get_params(*args, **kwargs)
for riddle in riddles:
riddle = functional.cropInputAndOutput(riddle, *params)
- Add padding to riddles
- Octagonal grid rotations, much more determinisic and accurate than the current RandomFloatRotate
- Super resolution (2x or 3x resolution increase)
- Torus translate (wraps back around) (can be combined with pad augmentation to make it more interesting)
- Mask augmentation (random pixels masked with a sparsity parameter)
- Unique mapping between color and pattern
- Static noise augmentation (add static noise)
- Static noise augmentation that tries to add static with colors not in the puzzle
- Reorder puzzle training boards (randomly)
- change the test board to one of the training ones and a training board to the test one, ranadomly
- Some augmentation that demonstrates the concept of correspondence?
- Object detector in input and output and finding objects that are in both, then maybe changing the entire object
- Repeat board with augmentation such as reflect
- Join different boards from same riddle into single board (with and without added augmentations)
- Grow input or output board by 2x horizontally and or vertically by spacing out the original board with a background color or random color (maybe this helps with adding a correspondance prior?)
- super res scale everything except background color by a random factor, maybe only across a specifix x/y axis
- flag to specify input/output/both on a transformation (impl using helpers not base class)
- flag or helper to artificially generate more board pairs with AND without an augmentation applied. Useful especially in the case where the augmentation is only applied to input board.
- update demo notebook with full riddle augs
- row_col_shuffler aug take a boardpair from train and copy it then shuffle some cols around.
- same pair as one in the train (better copy task!)