Skip to content

Trvlsplit par replicas

Created by: goord

First ad-hoc version of training-validation split per parallel replica. Basically, the network no longer stacks replicas together but rather connects each replica to its own set of observable wrappers.

This should be considered as a proof of concept. A final implementation would have to be more memory-efficient and leverage 'lazy' masking of the FK-tables within the network.

Merge request reports

Loading