ABSTRACT
This chapter discusses the evaluation in MetaNN. Evaluation is a fairly important step in deep learning frameworks. From the point of view of programming, the essence of a deep learning system is to transform the input data and produce the output result. Transformations like this require a large amount of computing resources. Layers of MetaNN call the corresponding functions in the process of forward and backward propagation to construct operation templates. Evaluation in MetaNN is the process of converting operation template objects to the corresponding principal type. Operations of MetaNN construct the corresponding operation templates, which actually describe the relationships among parameters and results. A parameter of a certain operation may be the result of another operation. Accordingly, operation templates form a hierarchy among parameters and results.
