abstract |
The invention provides a deep learning memory management method based on Tensor access. The method obtains the memory space overhead and time overhead under relevant decisions by collecting the execution information of the neural network and the performance information of the hardware platform, and establishes an integer linear programming model. By optimizing and solving the optimal Tensor scheduling strategy under constraints, it can solve the problem of insufficient memory and obtain high deep learning training performance. Compared with the prior art, under the same hardware performance, the present invention can realize the neural network training with larger batch size. At the same time, the invention also proposes a memory management system, which includes a profile module, a decision module and an execution module; the system can be directly added to the deep learning framework and is convenient to use. |