For DL4J there is a clear example of external error implementation deeplearning4j-examples/MultiLayerNetworkExternalErrors.java at master · eclipse/deeplearning4j-examples · GitHub
It has nice property of having separate feed-forward and back-propagation steps that are public APIs. Is there any similar example for SameDiff, where feed-forward, back-propagation and model update steps are separate and are using public stable APIs?
My problem is that I need to perform network evaluation during training (as a step of network output evaluation) and for DL4J it is absolutely possible to do that using public APIs in a clear and concise way, but for SameDiff I’m lost as there is no explicit feed-forward, backdrop steps that do not modify internal states