The plan for actually implementing model ensembling (#306) is to write a class that takes several model parameter files, loads them all, calls score_dataset on all of them, and averages the result. In order for this to work well, we really should have the random seed as part of the parameter file.
The only real trouble here is that the random seeds have to be set before keras is imported, so this has to be implemented carefully.
The plan for actually implementing model ensembling (#306) is to write a class that takes several model parameter files, loads them all, calls
score_dataset
on all of them, and averages the result. In order for this to work well, we really should have the random seed as part of the parameter file.The only real trouble here is that the random seeds have to be set before keras is imported, so this has to be implemented carefully.