2.1.3

• Extended and updated docs
• Support for rule files for character regularization
• Fix of auto rotation of PageXML
• Upgrade to tfaip 1.2.5
• Changed default parameters: enabling EMA weights of 0.99, weight decay of 1e-5
• support for parallel post-processing (and disabling it)

2.1.2

• Fixed definition of custom network architectures
• Fixed prediction positions

2.1.1

• fixed cmd line interface
• fixed migration of old models using BIDI
• fixed pagexml file sorting

2.1.0

• Completely changed the command line interface (see README and docs and tests)
• tests as runnable python unittests, see calamari_ocr/test
• Update of tfaip version
• Updated documentation/README

2.0.2

• Enforcing usage of a validation dataset
• Update text regularizer
• Fixed random blobs augmentation

2.0.1

• Support for custom data generators
• Added support for --validation_split_ratio splitting the provided --files into training and validation data
• Changed image manipulation operations to opencv which resulted in a massive speedup (thanks to @andbue )
• Fixed ScaleToHeightProcessor if used stand-alone
• Fixes for hdf5 dataset: closing files properly, shuffling data withing a file and files names during training

2.0.0

We proudly announce Calamari 2.0. There were great changes in the Code-bases which leads to cleaner code and a more robust usage. Thereto we rely on the tfaip-package providing a lot of functionality to Calamari. Old models will automatically be converted.

• Full rework and cleanup of the code-bases
• Training is now per epoch (as provided by keras) instead of by iteration
• Command line interfaces with only minor changes (epochs instead of max_iters)
• New checkpoint version 3 (models get automatically converted, old ones will be backed up)
• Using Tensorflow 2.3 as default (bugs in model upgrade in 2.4)

v1.0.5

Version update

v1.0.3

version update to 1.0.3

v1.0.1

• Fixed bug with orientation
• Data set viewer shows lines without GT (to support to show prediction datasets)
• parameter to stop training at a fixed accuracy