Skip to content

Evaluate pure HTR metric

When we are evaluating a HTR+NER model, we are usually interested in CER/WER without NER tokens (HTR) and with NER tokens (HTR+NER). The current evaluation manager automatically removes tokens and only compute pure HTR metric. We need to do just like *_no_punct metrics and simply compute both values.

We introduce two new supported metric names:

See https://gitlab.teklia.com/atr/dan/-/blob/dbfffbec104dcb7740f138a5f3df74882490a861/dan/ocr/manager/metrics.py#L102

Edited by Yoann Schneider