Auto groups generation && auto metric2name

[stared]

I looked at the code, and while there is some progress, it is not what I had in my mind.

• Please no hard-coded values. There should be loop over options. Regex would be the best, but startswith is OK-ish. [('', 'Training'), ('^val_', 'Validation')]. Or in the 'Training' actually need negative look before: https://stackoverflow.com/questions/899422/regular-expression-for-a-string-that-does-not-start-with-a-sequence (but maybe not, just leaving the first option as default). Note that thanks to the pattern we can easily use re.replace
• Not sure why log_history got an underscore. It shouldn't.
• In this grouped options, with names change we DON'T get "Validation Accuracy". We get "Validation", that's it. Yes, "Validation" won't be unique anymore.
• At the same time, group names should be named accordingly.
• We spell Negative Log-Likelihood. But in out case, we actually want Log Loss.

Look at the previous outputs, it should look similarish.

[Bartolo1024]

1) regex looks better a lot, first value as default is less clear. 2) I unnecessary added it to all and not reverted this one. 3) Can you write something more? I did not understand. Currently, we have "Validation Accuracy"? 4) It is done since commit 3, I did not change it in comment.
5) ok

[stared]

Ad 3 and 4 - the result should be as in:

[stared]

After this thing, I think we can merge, release, and move on. :)

[Bartolo1024]

I did it, but I have to get metrics_to_name from logger in ExtremaPrint and do transformation there, because it requires unique names. Optionally I can do cache for each group in ExtremaPrint. Which approach seems to be better?

[Bartolo1024]

• Ok, so I had to do a rebase because of directory name change.
• I extended tests - now raw_names and raw_group_names are set to True for old cases. I've also added some new test cases.
• The option with separate cache for groups in extrema printer will be better imho.