Skip to content

Releases: Calamari-OCR/calamari

2.3.0.post1

01 Oct 23:41
Compare
Choose a tag to compare

(without tests, docs and other non-resource data files)

2.3.0

01 Oct 21:05
Compare
Choose a tag to compare
  • 🔥 Switched from Apache 2 to GPL 3 license (which was effectively the case ever since Calamari 2.0)
  • 🔥 Update to ocrd-fork-tfaip==1.2.7 (supporting Tensorflow up to and including 2.15)
  • Adapt to changes in Numpy, Pillow, Python 3.9 and 3.11
  • 🔥 Switch to Tensorflow SavedModel format as backend for models (because they are compatible across Python versions)
    new checkpoint format version identifier is 6, updates will happen automatically (as usual), if possible
    to migrate older models, you might need to go back to Python 3.8 once
  • Implement --non_existing_pred_handling_mode evaluator param (providing new skip and error, in addition to current and now default empty)
    (fixes missing --non_existing_file_handling_mode from Calamari 1)
  • Fix warmstarting: keep network architecture
  • Introduce new model parameter temperature (inactive by default, can be applied ex post)
    for pre-scaling the logits, controlling steepness of probability distribution prior to CTC)
  • Respect directory of PageXML datasets: output as input, unless --output-dir is given
  • Change label PERCENT to PERCENT_CER in evaluator's confusion table
  • Fully git-controlled versioning (via setuptools_scm)

1.0.7

05 Jul 15:51
Compare
Choose a tag to compare

adapt to numpy, scipy, re changes

1.0.6

18 Aug 11:26
40103de
Compare
Choose a tag to compare

Fixes some issues with versions of requirements

2.2.2

21 Mar 11:52
Compare
Choose a tag to compare
  • Remove alpha channel from LA images

2.2.1

21 Mar 11:52
Compare
Choose a tag to compare
  • Fix calling calamari-predict --help without --checkpoints
  • Don't divide by zero if there are no predictions

2.1.5

21 Mar 11:30
Compare
Choose a tag to compare
  • Added predifined network architectures. Support to specify them via parameter
  • Tensorboard display fixed for retraining on original
  • Added a no_train flag to cross_fold_train.py to only create folds without training

2.2.0

21 Mar 11:51
Compare
Choose a tag to compare

2.1.4

02 Oct 07:50
f1cdbb4
Compare
Choose a tag to compare
  • Support to run cross-fold-train on distributes systems (slurm)
  • Added parameter for maximum line length
  • PageXML can now emit glyphs, words, and their confidences
  • Fixed support for lists files (files that comprise a list of filenames)

2.1.3

08 Sep 13:23
eba8e9c
Compare
Choose a tag to compare
  • 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)