Bespoke: A Block-Level Neural Network Optimization Framework for Low-Cost Deployment

As deep learning models become popular, there is a lot of need for deploying them to diverse device environments. Because it costly develop and optimize neural network every single environment, line research search networks multiple target environments efficiently. However, existing works such situation still suffer from requiring many GPUs expensive costs. Motivated by this, we propose novel optimization framework named Bespoke low-cost deployment. Our searches lightweight model replacing parts an original with randomly selected alternatives, each which comes pretrained or the model. In practical sense, has two significant merits. One that requires near zero cost designing space networks. The other merit exploits sub-networks public networks, so total minimal compared works. We conduct experiments exploring Bespoke's merits, results show finds efficient targets meager cost.