This repository is work in progress. If builds upon the batchgenerators framework but makes several key changes to the transforms:
- Transforms now explicitly distinguish between data types: images, segmentation, pixel-wise regression target, keypoints, bbox
- All transforms have been reimplemented from scratch with a focus on performance. In case of performance parity between previous numpy and new torch-based implementations, preference is given to pytorch.
- Transforms are applied on a sample level, not a batch level as was done previously!
Caveats:
- performance is optimized for CPU. GPU-based data augmentation is not supported (implementation may use numpy etc) and will not be supported
- currently this repository only covers a small subset of the transforms available in batchgenerators. Feel free to contribute more
We are happy to accept PRs that further optimize performance and extend the available transformations!
- Please provide benchmarking results relative to the old batchgenerators implementation (if applicable)
- Please stick to the current transform template!
batchgeneratorsv2 developed and maintained by the Applied Computer Vision Lab (ACVL) of Helmholtz Imaging and the Division of Medical Image Computing at the German Cancer Research Center (DKFZ).