Example - PyTorch Job

What is it about?

Distributed training with PyTorch. You can select level of parallelism, number of workers and resources for each worker (GPU, CPU).

How to run?

• Navigate to: cd pytorchjob .
• Build docker image with custom code:
  • docker build -f Dockerfile -t @username/pytorch-job . .
• Push docker image: docker push @username/pytorch-job .
• Edit pytorch.yaml (or other provided .yaml files).
  • Define PyTorch master and workers.
  • Make sure to run from the image you have just built image: registry.hub.docker.com/@username/pytorch-job .
• In Terminal, run: kubectl apply -f pytorch.yaml.
• Monitor pytorchjob with: kubectl logs pytorch-dist-mnist-nccl-master-0