pytorch/tnt

Feature-Request: Ability to save metric-values

sauercrowd opened this issue · 2 comments

It might be useful to be able to save the metrics to files, in order to be able to use them somewhere else for e.g. visualisation.

I'm thinking about something like:

mlog.save_meter(dir='train_metrics', meters={'loss':'loss.csv', 'accuracy':'accuracy.csv', 'confusion':'confusion.csv'})

For me that would be really beneficial, what do you think about it?

Yes I agree that this is an important feature! tnt.utils.ResultsWriter does log values to a file and may be what you’re looking for.

FYI In future versions this class will be renamed to “tnt.logger.FileLogger” since the class fits logging semantics.

Sorry, didn't see that! Looks really promising and I think that should be exactly what I need