Adaptive Distributed Training of Deep Learning Models