Abstract: Aiming at the problem of  the performance degradation of the automatic operator selection algorithm in convolutional neural network training under high memory pressure, we modelled offloading, recomputing and convolutional operator selecting in a unified manner and proposed an intelligent operator selection algorithm. The algorithm weighed the time overhead introduced by offloading and recomputing against the time saved by faster convolutional operators, found the scheduling of offloading, recomputing and convolutional operator selecting, and solved the performance degradation problem of the automatic operator selection algorithm. The experimental results  show that the intelligent operator selection algorithm reduces training time by 13.53% over the recomputing-automatic operator selection algorithm and by 4.36% over the existing offloading/recomputing-automatic operator selection algorithm.

Key words: memory, convolutional neural network training, convolutional operator, offloading, recomputing