Deep learning memory allocation optimization method and system

The invention discloses a deep learning memory allocation optimization method and system. The method comprises the steps of recording a memory allocation size and memory application and release time of a tensor involved in current round calculation to obtain a tensor size and life cycle information; performing memory allocation planning according to the acquired tensor size and life cycle information to generate a memory allocation plan; and according to the memory allocation plan, performing tensor memory allocation of subsequent rounds. According to the method, the characteristic that memory allocation among multiple rounds of iteration has natural similarity due to the fact that operation data of each round of iteration of a deep learning training task is different, but execution calculation graphs are the same is utilized, tensor size and life cycle information of the first several rounds of iteration are collected, a memory allocation plan capable of being used for subsequent iteration is generated, and