predict: ctc_decoder parameters are never applied

[bertsky]

calamari/calamari_ocr/scripts/predict.py

Line 114 in f0139d6

prepare_ctc_decoder_params(args.ctc_decoder)

This merely post-processes some of the command-line choices. It never actually instantiates a CTCDecoderProcessor, or replaces the default one in the postprocessor pipeline.

How is this supposed to have worked in the first place?

Also, the parameterization of this postprocessor begs more questions. Assuming some tests have been done with the dictionary feature (word beam search):

• Why is non_word_chars not automatically configured to all the punctuation characters in the entire charset during training? (All the public models I see merely contain the default characters, i.e. only ASCII punctuation.)
• Why is word_separator only whitespace by default – shouldn't that allow more cases like hyphen (esp. in German)?