Neural network training by including for implementation costs

摘要

An exemplary method of implementing a neural network is to select the first neural network architecture from the search space, and train the neural network with the first neural network architecture for accuracy and implementation cost. Including doing. The implementation cost is based on the programmable device of the inference platform. The method further selects a second neural network architecture from the search space based on accuracy and implementation cost, and outputs the weights and hyperparameters of the neural network with the second neural network architecture. including. [Selection diagram] Fig. 3