ABSTRACT
The internal network structure of a practical deep learning system can be quite complicated to perform challenging tasks like machine translation and autonomous driving. The training processes for different network structures and loss functions are different, but in general, the training of a whole model includes following steps: It is not difficult to be aware that the deep learning framework plays an important role in every step of model training. An excellent deep learning framework should not only support the training of models, but also further optimize the prediction using models in order to improve the speed of prediction and reduce the use of resources. Therefore, the core of a deep learning framework is matrix operations. Deep learning systems are complex and a good deep learning framework must be able to cover up the complexity to some extent, providing users with interfaces easy to configure and use.
