Keywords: machine learning, neural networks, tensorflow, software, productivity, compilers
We will describe a new software library for TensorFlow that automatically converts plain Python code into its TensorFlow equivalents, using source code transformation. For example, instead of tf.cond, and tf.while_loop, users would write imperative-style code with ifs and whiles, and expect it would be converted into its Graph equivalent. We aim to support other convenient Python constructs like lists, recurrent function calls, and list comprehensions, although we will not support the full Python language. Our approach is complementary with the new TensorFlow Eager project and will allow using the imperative style of Eager mode, while retaining the benefits of graph mode. By using automatic code conversion, developers can write code that’s more concise, efficient and robust.