rnn code testing method and device

Abstract

The present invention provides an RNN (Recurrent Neural Network) code testing method and apparatus. The method comprises the steps: using a hidden layer and an output layer in an RNN that controls RNN code implementation as to-be-tested layers; using an initial matrix as an input of a code segment that achieves the functions of the to-be-tested layers in the RNN code, and controlling the operation of the code segment to obtain an initial result matrix output by the code segment; according to the preset test times, changing an element value of the initial matrix every time to obtain a reference matrix, using the reference matrix as the input of the code segment, controlling the operation of the code segment again to obtain a reference result matrix output by the code segment, and judging whether element values corresponding to the positions of the changed element values in the initial matrix are changed in the reference result matrix and the initial result matrix; and if every judgment result is yes, determining that the logic of the code segment is correct. The test method and apparatus realize the testing of the RNN code and fill up the blank in the prior art in the RNN code testing.

Description

【Technical field】 [0001] The invention relates to the field of software technology, in particular to a method and device for testing RNN codes. 【Background technique】 [0002] Deep learning is a popular technology in the field of big data analysis, and is widely used in artificial intelligence fields such as image recognition, speech recognition, and natural language understanding. The core of the deep learning algorithm is its prediction module, and a classic model to realize the prediction module is the Recurrent Neural Network (RNN). [0003] From a code point of view, in the code implementation of deep learning-based applications, the code that implements the RNN function (RNN code for short) is the core part of the entire code, so the RNN code is tested to ensure its correctness, and the entire code Quality plays a key role. However, since deep learning algorithms output probability values, traditional black-box testing methods, etc. are not suitable for RNN codes. I...

AI Technical Summary

Problems solved by technology

However, since the output of the deep learning algorithm is a probability value, traditional black-box testing methods, etc. are not suitable for RNN codes

In addition, since RNN is a nonlinear model, the transformation test method for linear models is not suitable for RNN codes

Images