using loss as metric

[jagstein]

when using softmax_loss as the metric, the "best" trial is reported as the one with the highest loss. Need to fix this and also make sure glimr is exploiting the trials with the lowest loss.

[jagstein]

when using loss as a metric with different losses that give different values (e.g. hinge & CCE), how to figure out which is "best"

[jagstein]

tune allows you to specify the metric mode ("min" or "max") -- https://docs.ray.io/en/latest/tune/api/doc/ray.tune.search.Searcher.html
glimr supports passing metric mode, but when listing "best" trial, it still gives highest loss, not lowest. This may just be a reporting issue, but need to check. Also, issue with different losses needs solving