This folder contains an example where mnist is trained. This example is also used for e2e testing.
The python script used to train mnist with pytorch takes in several arguments that can be used to switch the distributed backends. The manifests to launch the distributed training of this mnist file using the pytorch operator are under the respective version folders: v1. Each folder contains manifests with example usage of the different backends.
Note: PyTorch job doesn’t work in a user namespace by default because of Istio automatic sidecar injection. In order to get it running, it needs annotation sidecar.istio.io/inject: "false" to disable it for either PyTorch pods or namespace.
Build Image
The default image name and tag is kubeflow/pytorch-dist-mnist-test:1.0
.
docker build -f Dockerfile -t kubeflow/pytorch-dist-mnist-test:1.0 ./
NOTE: If you you are working on Power System, Dockerfile.ppc64le could be used.
Create the mnist PyTorch job
The below example uses the gloo backend.
kubectl create -f ./v1/pytorch_job_mnist_gloo.yaml