heat
heat copied to clipboard
Feature: DataParallel: async SGD
asynchronous SGD also has many options, 1 option which looks promising now is a ring approach
there is a version of this implemented in heat now however it is based on nonblocking hooks but it still does traditional synchronization SGD. that is not traditional ASGD. there is another parallelization method in DASO (#728 ) but that is a different optimization method.
What should be done here @coquelin77 ?
In principle of interest, if we want to extend/improve NN support. Could be a good student project if a sufficiently detailed description could be added.
Reviewed within #1109