bobye
bobye
It is not as easy as other operators: - MIRROR operator should also work for Operational objects - It should be compatible with other operators
Currently I am focusing on other parts
Another workaround is to consider pass derivative as explicit outputs, and use aggregate to obtain overall gradient.
I like the second solution to handle data parallelization: It always keeps the status immutable.
implemented training for RAE
parallelzing bug for RAE
May include mnist data for experiments, this summer.
Need some further unit tests
Why not? I haven't updated this project for more than two years.
memory footprint should be fixed in the first place.