JBrainy is supposed to do the normalization by default.

Sometimes, the hardware feature names don't start with "PAPI"

This reverts commit d8a32258.

Compiling to run a JMH benchmark actually did not work.

It was set to the wrong commit.

Uses filtering to check that the method name contains the keys
to match.

Fixed error if matching regex does not work.

Test now tries to load methods with columns in some other order.

Accepts two different formats (including Christoph's)
and ignores non-matched collections if there is
no method matching.

The column indices were used instead of the column names,
so re-ordering the column broke the trace loader.

Previous version did not build with sm-java/container-sub

There is a classifier for each collection instead of one global
classifier.

The data is split and each classifier is trained on a subset of it.

Added some experiments with interface and collection specific
classifiers instead of a global one for all collections.

Erroneously ran the square of the needed benchmarks, and duplicated
data.

Switched from exporting benchmark results as nested maps to exporting
lists of triplets with benchmark, counter name, and values.

Instead of taking a boolean as a parameter, added methods.

Re-trained with new data.

Uses a software counters file that has been regenerated with new
formats for method names.

based on the jmh-results-9307f70f.csv

Avoids taking too much memory.