Methods and equipment for training deep learning models, electronic devices, computer-readable storage media and computer programs

摘要

The embodiments of the present disclosure disclose methods and devices for training deep learning models. One specific embodiment of the method is to divide the model description information into at least two sections based on the step of acquiring the model description information and the configuration information of the deep learning model and the division point variable in the configuration information, and the model. The step of loading the descriptive information into the corresponding resource and executing it, and after inputting the training sample of a batch into the resource corresponding to the model descriptive information in the first section, the training is started, and the obtained context information is applied to the following. When the step of inputting the resource corresponding to the model description information of the section, the step of repeating as described above until the execution result of the resource corresponding to the model description information of the last section is obtained, and the training end condition are satisfied. Includes a step of outputting a trained deep learning model and, if the training end condition is not met, a step of taking a training sample of the next batch and executing the training step until the training end condition is met. The embodiment can realize a free combination of different types of devices, fully exert the computing power of different computing devices, and improve the training speed. [Selection diagram] Fig. 2