Common issues¶
Note
If you don’t find an answer here, check out the issue tracker on Github.
Config files are ignored¶
Install PyYAML with pip install pyyaml.
Beam decoder with beam=12 does not create 12-best lists¶
The beam decoder stops when the best scoring hypothesis ends with the end-of-sentence symbol. Therefore, n-best lists
do not contain hypotheses longer than the best hypothesis. If you want to create full n-best lists, use
--early_stopping false. Thereby, the decoding does not stop until all active hypotheses end with the end-of-sentence symbol.
UnicodeDecodeError: ‘ascii’ codec can’t decode byte...¶
Python is not properly set up to read your word maps or source sentence files. Make sure that your Anaconda environment and your
locals are set up correctly (e.g. LC_ALL points to a value installed on your system). Consider processing indexed sentences
(default preprocessing and postprocessing options) and handle unicode separately.
Decoding continues after the end-of-sentence symbol¶
Double-check that the reserved IDs in your predictor are consistent with SGNMT. You can change the
reserved IDs used by SGNMT with the indexing_scheme parameter, or mask predictors with their
own reserved IDs with the idxmap predictor wrapper.
‘No complete hypothesis found’ when combining models with different tokenizations¶
Make sure that all your word maps use the standard names for reserved tokens (<eps>, <s>, </s>). Also,
double-check that your indexing_scheme parameter is consistent with the word maps and the indexing schemes
in your models.
‘int’ object is not callable (fst, nfst, rtn predictors)¶
You are likely to use an outdated OpenFST version. The minimum required version is 1.5.4.