This code is released primarily for the purpose of documentation.
Implements 2-fold cross validation for which species to train with for the Helixer deep learning gene calling tool. This repository:
- sets up data
- starts training and evaluation of Helixer modesl via
nnictl - summarizes results.
Note that while the code allows extended iteration and refinement of a promising set, in practice compute was better spent in one round with many random starts and then a remix.
General process is as follows
# setup data & start many training runs with randomly selected
spselec setup <parameters>
# for as many rounds as makes sense (stop at status 'seeds evaluating', probably once for recommended half round only)
# repeat the following (run after previous round finishes, manual check requried)
spselec next <parameters>
# when ready
# combine training species from best two models in each fold, and all other species as validation
spselec remix-train
# evaluate on all species that were in none of the remix training sets above
spselec remix-eval