Google developers behind Swift for TensorFlow, which tunes the Apple-designed Swift programming language for machine learning applications, shared project roadmap information in a recent talk. Future plans for Swift for TensorFlow include capabilities such as C++ interoperability, improved automatic differentiation, and support for distributed training.
Swift for TensorFlow is an early-stage, Google-led project that integrates Google’s TensorFlow machine learning library with Swift, the modern general purpose language created by Apple. The use of Swift enables more powerful algorithms to be expressed in a new manner, and easy differentiation of functions via generalized differentiation APIs, according to the Swift for TensorFlow developers.
Open source Swift has been described on the Swift for TensorFlow project website as easy to use and elegant, with advantages such as a strong type system, which can help developers catch errors earlier and promotes good API design. Building on TensorFlow, Swift for TensorFlow APIs provide transparent access to low-level TensorFlow operators.
Swift for TensorFlow is focused on two sets of users: advanced researchers limited by current machine learning frameworks, and machine learning learners just getting started. Extensions to the Swift language provide interoperability between Swift and Python, a popular language in machine learning. Python can be imported within a Swift Jupyter Notebook and TensorFlow itself is Python-friendly. Developers can write Swift to call into Python libraries, with no wrappers and no additional overhead.
Where to download Swift for Tensorflow
You can download Swift for TensorFlow from GitHub. Tutorials, documentation, and instructions for community participation in the project can be found at tensorflow.org/swift.