Intermediate representation method and device for neural network model calculation

Abstract

The invention discloses a neural network model calculation-oriented intermediate representation method and device. The method comprises the following steps of S1, analyzing an input model file to obtain topological structure information of a neural network; s2, constructing a logic calculation graph; s21, deducing physical layout information of each operator in the logic calculation graph; s22, deriving the element attribute of each operator in the logic calculation graph; s23, deducing description information of an input and output logic tensor of each operator in the logic calculation graph; s3, constructing a physical calculation graph; s31, generating a physical calculation graph; according to the meta-attribute-based intermediate representation for neural network model calculation disclosed by the invention, data parallelism, model parallelism and pipeline parallelism are originally supported from an operator level. According to the neural network model calculation-oriented intermediate representation method and device disclosed by the invention, the calculation expressions are taken as basic units, the tensors are taken as flowing data in the calculation graph formed by the whole calculation expressions, and the calculation process of the neural network model is realized in a composition manner.

Description

technical field [0001] The invention relates to the field of deep learning, in particular to an intermediate representation method and device for neural network model calculation. Background technique [0002] With the rapid development of artificial intelligence industrial applications, the demand for large models in practical application scenarios has become more and more urgent. Most of the existing deep learning frameworks provide efficient interfaces for the expression of neural network model computation and the training of neural network models on a single device. However, existing deep learning operating systems may not be flexible and effective when targeting new distributed devices for large-scale deep neural network model training, because distributed devices require more complex parallelism than single devices. In addition, the distributed training interface that has been developed strengthens the parallelism of the models of the existing deep learning framework,...

AI Technical Summary

Problems solved by technology

However, existing deep learning operating systems may not be flexible and effective when targeting new distributed devices for large-scale deep neural network model training, because distributed devices require more complex parallelism th...

Images