A Hybrid Parallelization Approach for Distributed and Scalable Deep Learning