Containers have quickly become a common method of developing and deploying AI and data center applications in the Enterprise space. Given their lightweight design and deterministic software environment, they are able to reliably transition functional applications between systems and clusters without requiring significant system reconfiguration.
NVIDIA’s container registry, NGC, contains several Docker images that have been tuned and tested to work on NVIDIA hardware with topics ranging from different AI frameworks, like TensorFlow, MXNet, and PyTorch, to healthcare, like NVIDIA CLARA SDKs, to robotics with NVIDIA Isaac Sim, and much more. These images are great starting points for working on accelerating your applications on NVIDIA platforms.
NGC also contains a private container registry allowing users to publish modified images which can be used on any machine capable of running Docker which has access to the registry. This makes it seamless to migrate images from a local workstation to a cluster of DGX nodes to accelerate your workload.
This tutorial is designed to be a guide for setting up Docker on a local workstation, authenticating with NGC, pulling and running Docker containers, modifying a base image from NGC, running a sample image classification application, and pushing a modified image to NGC’s private registry. This is in preparation for creating a modified image for your own application which can later be deployed on LaunchPad for training at scale. While the tutorial has a demo for a specific use-case that can be followed, you are encouraged to adapt the steps as needed to use a different base image, install specific packages, or change the image name for your specific workload.