mittagessen/kraken 7.0.3 on GitHub

What's Changed

The python-native BiDi implementation of python-bidi used by the recognition container classes was highly outdated and breaks on some modern text, in addition to not implementing the 6.3 paired bracket feature that reduces issues with bracket mirroring at directional boundaries. The library has been replaced by an internal implementation following Unicode 17.
Recognition models now extract lines in the data model they have been trained on from XML files, i.e., bounding box lines for bbox models and baselines for baseline models. The default can be overridden with the --linetype option. During recognition training the same option can be used to select the data model to train with when training either from XML files directly (ketos train -f xml --linetype [bbox|baselines] ...) or with compiled datasets (ketos compile -f xml --linetype [bbox|baselines] ...).
When setting a Unicode normalization in ketos test both the prediction and ground truth are now normalized. In previous versions the transformations was only applied to the ground truth.
ketos train now works with DDP

GroupNorm was computed with padding, changing recognition results when changing the batch size. The layer strips padding before now, making recognition (mostly) invariant of batch size. The default recognition architecture doesn't use GroupNorm so the erroneous behavior was limited to custom VGSL specs.
A bug in the 7.0 CoreML loader broke loading of pre-1.0 models that did not contain a kraken_meta field.
TensorBoard logging in training is now functional again
The imagename attribute in Segmentation container objects is now set by the legacy bbox segmenter.
Binary training datasets with bbox data have been fixed.
Insertion and deletions were inversed in the ketos test report.
A regression in ImageTransforms caused crashes on operating systems that default to multiprocessing with spawn.
Some small changes in load_safetensors that allow deserialization of shared weights directly on cuda.