Bug in train-multi-stages.sh (individual strategy training), in the last part: re-training selected models using all cases and re-cross-validating in order to use predictions for meta-training stage

[erwanm]

issue migrated from original private gitlab repo

The system frequently crashes in this part, I'm not sure why:

• the problem happens in the directory best, apparently always with the GI strategy
• the part which doesn't work is the "re-cross-validation", not the recomputing of the global model: missing .answers files for certain configs, (whereas the number of .model directories seems always correct)
• apparently removing the dir "best" isn't enough: is it because the error happens every time, or because something outside this part causes the issue?

[erwanm]

Discovered that GI also fails regularly in the first training stage (indiv); this points to a bug in train-cv.sh with GI. probably GI fails and train-cv doesn't recover even in failsafe mode?