Train a GT4HistOCR Calamari model

train.sh trains a Calamari 1 model based on GT4HistOCR. Or rather 5 using cross-validation to use for confidence voting. This repository mainly serves as documentation of the provenance of the model published at https://qurator-data.de/calamari-models/, not as the definitive guide to training such a model.

Trained models

For a finished model have a look here: https://qurator-data.de/calamari-models/

Training your own model

If you really want to, you can use this script to train your own. It takes about 1 week on a Nvidia RTX 2080 GPU. Please use requirements.txt in that case to setup a virtualenv.

train.sh is able to download GT4HistOCR from the web if the data submodule is not available, that is if you're not a member of the Qurator team at SBB.