DeePMD-kit is a realization of Deep Potential, it is a deep learning package for many-body potential energy representation and molecular dynamics, which can written in Python/C++. In this challenge, we are required to make improvements on the training procedure. We end up with the speed results that are 7.946(watewr), 3.121(mgalcu), 1.728(copper) compared to the original results, through the optimization of custom operators, Compressed Training.
We analyse the workflow of DeePMD-kit. Moreover, we design the strategies used to optimize the performance in the process of the training process. In the end, we evaluate the efficiency of our approaches and come to a conclusion.