On the problem of predicting calculation time needed for neural network executed by means of gpu in case of convolution neural networks

Computation performance of GPU devices has grown significantly in recent time. After CUDA architecture has appeared researchers could make active use of GPU devices in their work including nanotechnology area. However in many cases it is difficult to predict acceleration factor for an algorithm after its implementation by using GPU and consequently to estimate computational efficiency of this algorithm. Thus the task of computational performance prediction of an algorithm implemented using GPU is crucial.
This work describes computational performance prediction model for algorithms based on artificial neural networks. Neural network depends on large amount of hyperparameters, which are defined on the architecture design stage, and affect its execution speed and results accuracy. A process of selecting these parameters values could take a long time. Application of prediction approaches allows to reduce time needed for the selection stage and to increase precision of hyperparameters' estimations.