Changelog

With the supporting features from TaggingBackends!11 (merged), if predict is called on a directory instead of a file:

a primary input file is guessed among the files with extensions: .spine, .outline, .csv, .mat or .hdf5;
only the files related to the primary input file are taken as input, in addition to the primary input file.

As a consequence, the extra files in the input directory are no longer listed in the dependencies section of the resulting .label file.

Additional change:

the path specified by the training_dataset element in the clf_config.json file of a tagger is now relative;
a requirements.txt file is shipped to freeze the version number of most Python dependencies.