TensorWatch, a debugging tool with many advanced capabilities that researchers and engineers will find pretty helpful in their work, introduced by Microsoft at the 2019 ACM SIGCHI Symposium on Engineering Interactive Computing Systems, is now open source.
While the tool is also helpful in reducing the complexities of AI projects, with more focus on the crucial part of the development process, which includes debugging, making it easier to get rid of errors and thus minimize one of the most time-consuming jobs in software projects.
As deep learning is increasingly accompanied by model complexity, like large datasets and training times for models, researchers will often need to understand the training metrics when working on novel concepts, and available tools for machine learning training have hitherto focused on a rather garbage in, garbage out approach.
But with TensorWatch, it will become easier to spot bugs by visualizing the models in interactive graphs, and through the data made available by the AI model during testing. The tool is a Python library, and also allow users to build own custom UIs or use it within the vast Python data science ecosystem, with support for several standard visualization types, such as bar/pie charts, histograms, and 3D variations.
Additionally, there is the lazy logging mode, which doesn’t require explicit logging of information.
TensorWatch reduces overall processing overhead by cutting down on the amount of data that’s processed to find problem patterns. If you wish to give it a spin, the code for TensorWatch is available on GitHub.