Processing unit, calculation device, and calculation graph processing method of deep learning model

Abstract

The invention discloses a processing unit, a calculation device and a calculation graph processing method of a deep learning model. The processing unit comprises: an instruction fetching unit for fetching computer instructions from a memory external to the processing unit; an instruction decoding unit which is used for decoding the fetched computer instruction; an instruction execution unit which is used for executing the decoded computer instruction so as to extract the attribute of an operator from a computational graph used for a deep learning model under the first deep learning framework; converting the computational graph into an intermediate expression conforming to the following of the acceleration unit, wherein the attributes of operators in the intermediate expression are different from those of operators in the computational graph; performing model processing on the intermediate expression; and converting the processed intermediate expression back to a computational graph of the first deep learning framework, wherein the attribute of the operator in the intermediate expression is replcaed with the extracted attribute of the operator of which the operator identifier is not changed. Compared with the prior art, the embodiment of the invention needs fewer mapping functions to be maintained.

Description

technical field [0001] The present disclosure relates to the field of chips, and in particular, to a processing unit, a computing device, and a calculation graph processing method for a deep learning model. Background technique [0002] The current mainstream deep learning frameworks include TensorFlow framework, MxNet framework, Caffe framework, MxNet framework, and so on. These frameworks all define different sets of operators, and even operators with similar functions have different attributes in different frameworks. An operator is the basic unit of operation in a deep learning model, such as a convolution operator. In TensorFlow, the filling attribute of the convolution operator is "VALID" or "SAME", which respectively means that after the convolution operation is completed, "VALID" means that only valid convolution is performed, and the boundary data is not processed, and "SAME" means The convolution result at the boundary is retained; however, in MxNet and Caffe, th...

AI Technical Summary

Problems solved by technology

The disadvantage of this is that numerous mapping functions need to be developed and maintained

Images