Google has released a new version of its AI development platform TensorFlow. New features include an experimental API and more functionality for TensorFlow profiler.
The update of the AI development platform arrived two months after the previous release of version 2.0. In version 2.3.0, more attention is paid to understanding and limiting system resource usage. New mechanisms are offered within the data library, and the tool now has up-to-date profiler tools.
This results in an experimental snapshot API in tf.data. This API stores the data output of a preprocessing pipeline to disk. This way, all the used data can be reused. This, in turn, reduces CPU usage for recalculating the data for other steps in the process.
In addition, this experimental TFD service should speed up the tracing process in case the attached host cannot track the data consumption of the AI model. For example, if a model can process more images than the host can generate, the service ensures that a cluster of employees can prepare the necessary training data.
Furthermore, version 2.3.0 also offers several new tools in TensorFlow profiler, including a memory profiler and a Python tracer. The memory profiler provides users with more insight into how their machine learning-based model uses memory in a certain period of time. This is especially useful when optimizing their models or to get a better idea of its purpose.
The newly added Python tracer helps developers with tracking all Python functionality in their TensorFlow programs. It is also possible to visualize debugging data for the programming language through a new Debugger V2 dashboard. This gives a more in-depth insight into programs and shows graphical structures, history of torques, tensor competition and the location of code.