pipelines/samples/contrib/nvidia-resnet

A simple GPU-accelerated ResNet Kubeflow pipeline

Overview

This example demonstrates a simple end-to-end training & deployment of a Keras Resnet model on the CIFAR10 dataset utilizing the following technologies:

• NVIDIA-Docker2 to make the Docker containers GPU aware.
• NVIDIA device plugin to allow Kubernetes to access GPU nodes.
• TensorFlow-19.03 containers from NVIDIA GPU Cloud container registry.
• TensorRT for optimizing the Inference Graph in FP16 for leveraging the dedicated use of Tensor Cores for Inference.
• TensorRT Inference Server for serving the trained model.

System Requirements

• Ubuntu 16.04 and above
• NVIDIA GPU

Quickstart

• Install NVIDIA Docker, Kubernetes and Kubeflow on your local machine (on your first run):
  • sudo ./install_kubeflow_and_dependencies.sh
• Build the Docker image of each pipeline component and compile the Kubeflow pipeline:
  • First, make sure IMAGE variable in build.sh in each component dir under components dir points to a public container registry
  • Then, make sure the image used in each ContainerOp in pipeline/src/pipeline.py matches IMAGE in the step above
  • Then, make sure the image of the webapp Deployment in components/webapp_launcher/src/webapp-service-template.yaml matches IMAGE in components/webapp/build.sh
  • Then, sudo ./build_pipeline.sh
  • Note the pipeline.py.tar.gz file that appears in your working directory
• Determine the ambassador port:
  • sudo kubectl get svc -n kubeflow ambassador
• Open the Kubeflow UI on:
  • https://[local-machine-ip-address]:[ambassador-port]/
  • E.g. https://10.110.210.99:31342/
• Click on Pipeline Dashboard tab, upload the pipeline.py.tar.gz file you just compile and create a run
• Training takes about 20 minutes for 50 epochs and a web UI is deployed as part of the pipeline so user can interact with the served model
• Access the client web UI:
  • https://[local-machine-ip-address]:[kubeflow-ambassador-port]/[webapp-prefix]/
  • E.g. https://10.110.210.99:31342/webapp/
• Now you can test the trained model with random images and obtain class prediction and probability distribution

Cleanup

Following are optional scripts to cleanup your cluster (useful for debugging)

• Delete deployments & services from previous runs:
  • sudo ./clean_utils/delete_all_previous_resources.sh
• Uninstall Minikube and Kubeflow:
  • sudo ./clean_utils/remove_minikube_and_kubeflow.sh