The precision value that we log during training is a good approximation for RCA. Alternatively, you can run the training code without direcitons update and use the validate_classifier (apparently, I should enable it with a flag option in the run_train.py). MOS is actually a human-labeled based metrics.
The
precisionvalue that we log during training is a good approximation for RCA. Alternatively, you can run the training code without direcitons update and use thevalidate_classifier(apparently, I should enable it with a flag option in therun_train.py). MOS is actually a human-labeled based metrics.