An efficient conversion method and device for a deep learning model

Abstract

An efficient conversion method for a deep learning model provided by the embodiment of the invention is used to solve the technical problem that the development efficiency and operation efficiency of a deep learning model are low. The method includes the following steps: building a data standardization framework corresponding to an NPU (Neural-Network Processing Unit) model according to a general deep learning framework; using the data standardization framework to convert the parameters of a deep learning model into the standard parameters of the data standardization framework; and converting the standard parameters into the parameters of the NPU model. According to the invention, a unified data standardization framework is built for a specific processor according to the parameter structures of general deep learning frameworks. Standard data can be formed using the unified data structure of the data standardization framework according to the parameters of a deep learning model formed by a general deep learning framework. Thus, the process of data analysis by the processor depends much less on the structure of the deep learning model, and the development of the processing process of the processor and the development of the deep learning model can be separated. A corresponding efficient conversion device is also provided.

Description

technical field [0001] The invention relates to data processing of a deep learning framework, in particular to an efficient conversion method and device for a deep learning model. Background technique [0002] In the field of deep learning technology, general deep learning frameworks such as Caffe, TensorFlow, and Torch are usually used to define deep learning models for massive data processing and analysis. The training and data analysis of deep learning models require a large number of computing resources. Due to the large differences in the storage and analysis methods of data information in different general-purpose deep learning frameworks, and due to the differences in the hardware structure of processors, the data processing and data scheduling processes of deep learning models often cannot be consistent with the structural characteristics of processors. Adaptation reduces the computing efficiency of the processor. [0003] For example, in existing technical solution...

AI Technical Summary

Problems solved by technology

[0004] In view of this, the embodiments of the present invention provide a method and device for efficiently converting deep learning models, which are used to solve the technical problems of low development efficiency and computing efficiency of deep learning models

Images