PREDICTION OF VARIATIONAL AUTOENCODER EXECUTION TIME BASED ON ARIMA FAMILY SERIES MODELS

Currently, neural networks of the «variational autoencoder» type are used in various fields, including image compression. In FPV control of UAVs, they are used as a part of neural network video stream codecs. Neural network codecs allow for a high degree of frame compression and transmitting FPV video streams over low-bandwidth channels. Execution time of variational autoencoders is inconsistent and depends on the equipment and its load level. In this article, the execution time of variational autoencoders is considered as a time series. Predictive models of the execution time of variational autoencoders are studied both at the stage of video stream encoding and at the stage of decoding. As the result, it is shown that ARIMA models does may not be used to predict variational autoencoders operating time. Prediction models for encoder and decoder execution time are also provided.