group predictions by source peptides before scoring their quality, added logging of each hyperparameter's performance

using proper cross-validation to estimate accuracy of just synthetic data (previously small leakage of information via shared similarity matrix)

moved helpers for allele_similarities and synthetic_data into own modules, added script for best synthetic smoothing coef which 0.01

Fixes #9 peptide_length parameter is not passed

Python3 pickle error with bytes

False start of a NIPS paper