conformer ocr

conformer_ocr is a drop-in replacement for kraken's CNN-LSTM line text recognizer based on a slightly modified conformer architecture.

Installation

::

    $ pip install .

Training

Options are largely identical to those offered by ketos test, including possible data set formats.

::

    $ cocr -d cuda train -f binary --workers 32 *.arrow

Default hyperparameters are optimized for large datasets (~50k lines), trained with reasonably large batch sizes (32) on a GPU with at least 16Gb memory.

Inference

Inference is supported with:

::

    cocr ocr -i input output -m model.ckpt ...

Inputs can be defined as with kraken's inference tools. A segmentation must be provided from XML files in ALTO or Page XML format. Outputs in any of kraken's formats are supported.