Error running 23_recog.sh on pre-trained dataset

[rwth-i6]

Error running 23_recog.sh on pre-trained dataset

[albertz]

23_recog.sh uses tools/recommend-recog-epochs.py to get epochs to do recognition on. That script uses the training/cv scores, which are saved by Returnn in the learning_rate_file.
See in returnn.config: learning_rate_file = "data/exp-%s/train-scores.data" % setup_name.
(In other example configs, you often see learning_rate_file = "newbob.data".)
You do not have that file train-scores.data. You incorrectly renamed the model data file to that name.

A simple workaround for now: In 23_recog.sh, replace this line:

epochs=$(./tools/recommend-recog-epochs.py --experiment $experiment)

By:

epochs=238

[albertz]

I think this should resolve it. Closing the issue now.

[aaronrr]

Yes, this has resolved it. Thanks for the help!