helmut-hoffer-von-ankershoffen/jetson
Helmut Hoffer von Ankershoffen experimenting with arm64 based NVIDIA Jetson (Nano and AGX Xavier) edge devices running Kubernetes (K8s) for machine learning (ML) including Jupyter Notebooks, TensorFlow Training and TensorFlow Serving using CUDA for smart IoT.
8 Releases
Latest: 6y ago
TensorFlow Serving meta and public endpoint0.4.4Latest
helmut-hoffer-von-ankershoffen·6y ago·September 12, 2019📋 Changes
- enh: Provide meta information in response of webservice containing TF model name, Jetson model name, timestamp and duration in milliseconds
- enh: Provide make target `make tensorflow-serving-predict-public` showing access to public SSL endpoint on `https://tensorflow-serving.polarize.ai provided` via CloudFront -> K8s/Traefic loadbalancer -> K8s/service -> webservice on Jetson Node
Automation0.4.3
helmut-hoffer-von-ankershoffen·6y ago·September 12, 2019refactor: **Automatically repack CUDA ml libraries** as part of provisioning thus reducing the number of commands to enter refactor: **Automatically create K8s token** as part of provisioning K8s on (new) nodes thus reducing the number of commands to enter enh: Provide **more build meta in Docker images**
Versioned images0.4.2
helmut-hoffer-von-ankershoffen·6y ago·September 10, 2019📋 Changes
- enh: Images pushed to Docker Hub are now tagged with release versions using `make publish-all [tag]`
- enh: All images contain build metadata in `/meta` directory
Enhancements and fix0.4.1
helmut-hoffer-von-ankershoffen·6y ago·September 10, 2019📋 Changes
- feat: Provide `make publish-all` to publish all images to DockerHub
- enh: Set additional CUDA compute capability 6.2 for jetson/xavier/tensorflow-serving-base allowing run on NVIDIA Jetson TX2
- fix: Set compute CUDA capability to 5.3 for jetson/nano/tensorflow-serving-base
- enh: Allow passing any options from Skaffold to Docker Build via custom builders
- doc: Add hyperlinks to README
Jetson AGX Xavier Developer Kit0.4.0
helmut-hoffer-von-ankershoffen·6y ago·September 9, 2019📋 Changes
- Automated provisioning of guest VM (Ubuntu) using Vagrant+Virtual Box+Ansible as required for using NVIDIA SDK Manager on Mac
- Automated build of custom Kernel for Xavier supporting Kubernetes inc. Weave Networking
- Automated build of custom rootfs for Xavier providing SDK components on-flash
- Support for headless OEM setup of Xavier
- Automated provisioning of Xavier after flash with identical featureset as for Jetson Nano
- Automated integration of NVMe SSD during provisioning for adequate storage of Docker images and volumes
- Automated and persistent entering of performance mode and set max frequencies for CPU/GPU/EMC frequencies for improved performance
- Automated build, test, deploy and publish of Docker images and Kubernetes deployments for Xavier with identical featureset as for Nano
TensorFlow Serving0.3.0
helmut-hoffer-von-ankershoffen·6y ago·July 30, 2019📋 Changes
- Base image for TensorFlow Serving inc. latest TensorFlow core, adapted to CUDA capabilities of Jetson devices
- Python / Fast API based webservice as facade for TensorFlow serving inc. health check for K8s probes, interactive OAS 3 documentation, request and response validation, accessing TensorFlow Serving via its REST or gRPC endpoint
- Integration of Googles container structure tests in all builds
NVIDIA JetPack 4.2.10.2.0
helmut-hoffer-von-ankershoffen·6y ago·July 23, 2019📋 Changes
- Works with NVIDIA JetPack 4.2.1
- Automatically repackage libraries bundled with the JetPack image such as CUDNN, TensorRT and python bindings for Docker Builds
- Semi-automatic setup of a SATA SSD drive connected via USB 3.0 as boot device thus largely increasing throughput on disk heavy workloads
Basics for ML0.1.0
helmut-hoffer-von-ankershoffen·6y ago·July 14, 2019📋 Changes
- Auto-bootstrap of macOS development environment
- Auto-provisioning of basics on Jetson node
- Auto-deploy of Jupyter server supporting CUDA accelerated Tensorflow+Keras running in Kubernetes Pod on Jetson node