The problem of choosing hyperparameters for deep neural networks is subject of ongoing research. This session presents a specific tool to approach this issue: propulate (github paper)
Propulate is an asynchronous population-based search framework for hyperparameter and neural architecture search on high-performance computers. It utilizes HPC native foundations like MPI and HDF5 to leverage supercomputer performance at scale without sacrificing user friendliness.
We will introduce propulate and its mechanism and get some hands-on time on the LUMI supercomputer.
Participants are welcome to bring their own models to optimize.
Read the full description here on GitHub.
