Extra validation datasets
It would be nice to be able to feed extra validation datasets, that are tracked during training, but are not used during "checkpoint" validation.
The easiest would be to:
- Modify the run() function at the trainer engine to accommodate a new parameter called
extra_valid(or similar), that if present would contain a list of extra validation datasets that are tracked during training.
- The output CSV (trainlog.csv) would need to be modified to contain an extra column in this case, with the validation loss for each dataset in that list (so an "array" column).
- Finally, the "train" script needs to be modified to take the extra parameter into consideration.