Kubeflow 0.1 boasts support for TensorFlow, Jupyter Hub & more
It’s been a while since we last checked in on Kubeflow, the open source option for making ML stacks easier. Kubeflow 0.1 boasts a number of technical improvements, including support for TensorFlow, Jupyter Hub, and more!
Kubeflow has been pretty busy. Since their initial announcement last year, this mix of Kubernetes and TensorFlow has grown incredibly fast to one of the top 2% of projects on GitHub. It’s not difficult to see why so many developers and organizations are contributing.
The Kubeflow team wanted to offer developers and data scientists a scalable ML stack that is easy to use and very portable. The system takes care of the ML stack, making it easier for developers to focus on the data itself. Kubeflow gives users simple manifests, meaning they can use an ML stack anywhere Kubernetes is already running. The self-configure option makes it even simpler to use.
Now available on GitHub, Kubeflow 0.1 provides a basic set of packages for developing, training, and deploying machine learning models. This set is minimal, but packs a big punch in terms of tooling.
Kubeflow 0.1 now offers a Jupyter Hub to help create interactive Jupyter notebooks for collaborative and interactive model training. These notebooks can be shared between any numbers of collaborators, making development more congenial.
Of course, it’s not an open source machine learning project if it doesn’t acknowledge TensorFlow. Kubeflow 0.1 also offers a number of TensorFlow tools, like the Training Controller for native distributed training. The Training Controller can be configured to CPUs or GPUs; it also can be adjusted to fit the size of a cluster with a single click. Additionally, TensorFlow Serving container makes it easy to export finished models to Kubernetes.
Thanks to community contributions, Kubeflow now supports Argo, for managing ML workflows. Other help includes Seldon Core, for running complex model deployments as well as non-TensorFlow serving.
Future plans for Kubeflow include a 0.2 release later this summer, including new features like a simplified setup via bootstrap container, improved accelerator integration, and support for more ML frameworks like Spark ML, XKGBoost, and sklearn.
Getting Kubeflow 0.1
Want to try out Kubeflow? It’s pretty simple. There’s a walkthrough on Katacoda, a guided tutorial, and, of course, you can start a model of your own with any Kubernetes conformant cluster. More information is available on GitHubS.