Microsoft has made a new tool available in the open source community. It concerns TensorWatch, an internally developed tool to make artificial intelligence (AI) projects less complex. The tool specifically focuses on debugging AI models.

Searching for and resolving errors in code takes a lot of time, regardless of which software project is being worked on. When it comes to AI development, it is often even more time-consuming, because machine learning models are very complex. As a result, there are many more ways to destroy them than with a traditional program.

TensorWatch aims to make it easier to find errors, writes Silicon Angle. It does this by enabling developers to visualize their models as interactive graphs. The tool generates those graphs by using the data that an AI produces when it is tested. According to Microsoft, each source of information is displayed as a “stream”. Such a stream can contain the output of a model, as well as statistics on how much processing power it uses and TensorWatch charts.

The advantage of this approach is that it is easier to work with the data. A developer can use the same stream several times in different visualizations, but also create a graph showing different information streams next to each other. TensorWatch lets developers create everything from simple bar charts to complex, three-dimensional maps that virtualize potential errors in an interactive, virtual space.

Jupyter Notebook

Users can zoom in on parts by writing queries that can be used to manipulate the graph. This feature is powered by Jupyter Notebook, a popular open source experimental programming application built into TensorWatch by Microsoft.

With TensorWatch, developers can not only search for errors in a virtual way, but the tool must also make it more hardware efficient to search for the errors. This is possible thanks to a feature that Microsoft researchers call “lazy logging mode”. TensorWatch can reduce processing overhead by reducing the amount of data that needs to be entered to find problem patterns. The tool can then only observe the core variables that show how well a model performs during testing.

The code for TensorWatch is on GitHub.

This news article was automatically translated from Dutch to give Techzine.eu a head start. All news articles after September 1, 2019 are written in native English and NOT translated. All our background stories are written in native English as well. For more information read our launch article.