Researchers from Google have developed and released a new tool that helps developers to easily write TensorFlow code.
Basically, the new tool called TensorFlow Coder which is part of Tensorflow 2.0 is able to write TensorFlow expressions that achieve the desired tensor transformation. The user is asked for an input-output example of the transformation and the tool performs a combinatorial search to find expressions that can match and execute such a transformation.
The tool was developed in order to help researchers and developers find the correct combination of TensorFlow expressions that can achieve a particular transformation and preserve mathematical correctness. Having a large number of operations TensorFlow sometimes poses a challenge to developers to find the right combination, and TensorFlow Coder solves exactly this problem.
The tool is easy-to-use and works well for transformations that require solutions consisting of 3-4 operations and results are returned within a minute of search. However, for more complex combinations of many operations, the tool requires a lot of time making this a limitation of Tensorflow Coder. In any case, the tool can be a useful asset for developers and researchers who can now write TensorFlow code by example.
The official release blog post contains details on how to use TensorFlow Coder and what are the potential limitations. The implementation of TFCoder was open-sourced and it is available on Github.