Method and system for model conversion between deep learning frameworks based on minimum execution cost

Abstract

The present invention provides a method and system for model conversion between deep learning frameworks based on the minimum execution cost, including: on the basis of the original technology, adding an operation conversion cost value, while considering the situation that multiple independent operations can be fused, and supplementing the fusion mapping; The concrete realization of the model is reflected in the operation transformation of the model. At this stage, according to the model transformation mapping table, the transformed model structure with the lowest execution cost is obtained through the dynamic programming algorithm. The invention can reduce the reading and writing process of intermediate results between multiple operations through operation fusion, thereby optimizing computing performance and storage space, and further reducing the execution cost of the converted model. At the same time, when there are multiple fusion options, the model conversion method with the smallest execution cost is obtained through the dynamic programming algorithm.

Description

technical field [0001] The present invention relates to the field of deep learning, in particular to a method and system for model conversion between deep learning frameworks based on minimum execution cost, and is specifically applied to model conversion of different frameworks. Background technique [0002] Current deep learning models can be implemented in different frameworks. After a model is trained on framework A, users can convert it to a model under framework B so that it can be directly inferred under framework B without retraining the model. like figure 1 As shown, each model Model consists of a variety of different operations. In general, each operation OP (operator) has operation type, input (support zero to multiple), output (support 1 to multiple), attributes, etc. content. like figure 2 As shown, each operation is associated with its respective input and output, figure 2 The model structure contains 5 OPs, namely Conv, SpatialBN, Relu, AveragePool and ...

AI Technical Summary

Problems solved by technology

[0010] The purpose of the present invention is to solve the problem in the prior art that the deep learning model is converted between different frameworks, and the execution cost of the converted model is too high, and a minimum execution cost-based method that can generate a minimum execution cost model is proposed Method and system for model conversion between deep learning frameworks

Images