First download the CAR-A and CAR-B dataset from http://www.orand.cl/en/icfhr2014-hdsr/#datasets.
Or use your own dataset and adapt its structure to match that of the CAR datasets.

The training can be started by running the train.py script. For example

python train.py --data '/path/to/ORAND-CAR-2014/CAR-A/' --epochs 20

Check python train.py -h for all possible options.

TODO

After training 100 epochs:

CAR-A Test | avg_dist: 0.108615 | accuracy: 0.905127 | avg_loss: 0.485466

CAR-B Test | avg_dist: 0.079289 | accuracy: 0.928230 | avg_loss: 0.315886