Deep learning reasoning engine testing method based on differential evaluation

Abstract

The invention discloses a deep learning reasoning engine test method based on differential evaluation. A model import check module, an intermediate representation acquisition module and a result evaluation module are included. In the model import checking module, whether the engine supports all operators and related parameters involved in the model is analyzed, and whether the structure of the model is kept consistent after the model is imported into the engine is judged in combination with an exception capture mechanism in the engine and model structure comparison before and after import. Inthe middle representation acquisition module, IR data acquisition paths provided by different compilers are arranged, engine information and corresponding acquisition instructions are packaged throughmanual processing, and a unified calling interface is generated. In the result evaluation module, a compiler list suitable for the differential test task is obtained and includes a compiler name anda corresponding model source, and an intermediate representation obtaining interface is called to obtain IR data.

Description

technical field [0001] The invention belongs to the fields of software engineering and machine learning, in particular for model processing of a deep learning reasoning engine. For the intermediate process and output results processed by the deep learning model, evaluate the support of the inference engine for the specific deep learning framework. Background technique [0002] With the rapid development of artificial intelligence, neural network models based on deep learning (DL) technology have emerged and are widely used in cutting-edge fields such as autonomous driving and medical diagnosis. In order to provide convenient model training and deployment services, deep learning frameworks such as TensorFlow, PyTorch, and Caffe have emerged as needed. However, due to a series of problems such as platform support differences, there are many difficulties in deploying DL models trained by specific frameworks on various hardware, which promotes the research and development of DL...

AI Technical Summary

Problems solved by technology

[0006] The problem to be solved by the present invention is: the test Oracle of the deep learning compiler is difficult to solve, and the validity of a single output result is difficult to evaluate

Images