Setup kubeflow on AWS. Copy-pastable

aws_kubeflow.sh

# mac os x specific:
# install dependencies:
brew install awscli
brew install kubectl
brew install weaveworks/tap/eksctl
brew install ksonnet/tap/ks
brew install docker
brew install helm

# configure connection to docker daemon
# export DOCKER_HOST=ssh://user@docker-machine.com

# export environment variables to access AWS or configure ~/.aws/credentials
export AWS_DEFAULT_REGION=us-east-1
export AWS_ACCESS_KEY_ID=KEY_ID
export AWS_SECRET_ACCESS_KEY=KEY

# create Kubernetes cluster:
# -N - number of nodes
# -n - name of the cluster
# -t instance type
# --ssh-public-key - optional path to your public ssh key to install on cluster nodes

eksctl create cluster -N 3 -r us-east-1 -n kubeflow-1 -t m5.large --ssh-public-key ~/.ssh/id_rsa_4096.pub --timeout=90m

# check status of the cluster

kubectl describe nodes

# configure persistent volumes for Jupyter
# check if storage classes exist

kubectl get storageclass

# if it doesn't create storage class

cat <<EOF | kubectl create -f -
kind: StorageClass
apiVersion: storage.k8s.io/v1
metadata:
  name: gp2
  annotations:
    storageclass.beta.kubernetes.io/is-default-class: "true"
provisioner: kubernetes.io/aws-ebs
parameters:
  type: gp2
reclaimPolicy: Delete
mountOptions:
  - debug
EOF

# check that storageclass is created

kubectl get storageclass

# kubeflow specific installation steps

export NAMESPACE=kubeflow
kubectl create namespace ${NAMESPACE}

# download kubeflow deploy script

export KUBEFLOW_VERSION=0.2.5
export KUBEFLOW_DEPLOY=false
curl https://raw.githubusercontent.com/kubeflow/kubeflow/v${KUBEFLOW_VERSION}/scripts/deploy.sh | bash

# launch installation

cd kubeflow_ks_app/
ks env set default --namespace ${NAMESPACE}
ks apply default

# check kubeflow deployment status

kubectl get pod -n ${NAMESPACE}

# at this point the installation should be finished

### Jupyter specific steps

# Create docker repository
export ACCOUNTID=`aws iam get-user|grep Arn|cut -f6 -d:`
aws ecr get-login --no-include-email --region us-east-1 > docker_login.sh
bash docker_login.sh
aws ecr create-repository --repository-name tensorflow-notebook-cpu --region us-east-1

# fetch the model
curl -o train.py https://raw.githubusercontent.com/kubeflow/examples/master/github_issue_summarization/notebooks/train.py
curl -o seq2seq_utils.py https://raw.githubusercontent.com/kubeflow/examples/master/github_issue_summarization/notebooks/seq2seq_utils.py

# Build, tag and push Jupyter notebook docker image to registy
docker build -t $ACCOUNTID.dkr.ecr.us-east-1.amazonaws.com/tensorflow-notebook-cpu:latest . -f-<<EOF
FROM gcr.io/kubeflow-images-public/tensorflow-1.9.0-notebook-cpu
RUN pip uninstall msgpack thinc
RUN pip install thinc
RUN pip install ktext annoy sklearn h5py nltk pydot matplotlib
COPY train.py /workdir/train.py
COPY seq2seq_utils.py /workdir/seq2seq_utils.py
EOF

docker push $ACCOUNTID.dkr.ecr.us-east-1.amazonaws.com/tensorflow-notebook-cpu:latest

# Connect to Jupyter Notebook

kubectl port-forward svc/tf-hub-lb -n ${NAMESPACE} 8080:80
# open http://google.com
# configure spawner options

# Image: image we created perviously - AWS_ID.dkr.ecr.us-east-1.amazonaws.com/tensorflow-notebook-cpu:latest
# CPU, Memory - up to you
# Then click "spawn"

# Check that we have jupyter-"something" pods running

kubectl get pod -n ${NAMESPACE} | grep jupyter

# launch jupyter web terminal and clone example kubeflow repo:
# git clone https://github.com/kubeflow/examples
#
# in jupyter console browse to the Training ipynb adapt training parameters and launch it
# set DATA_DIR to /home/jovyan/github-issues-data
# set training_data_size to amount of rows you want to process
# to launch select cell-run all


### Serve trained model

# create microservice with trained model and IssueSummarization.py script

# copy Dockerfiles and deps from github
git clone https://github.com/kubeflow/examples serve/

# run docker build
cd serve/github_issue_summarization/notebooks
docker run -v $(pwd):/my_model seldonio/core-python-wrapper:0.7 /my_model IssueSummarization 0.1 gcr.io --base-image=python:3.6 --image-name=gcr-repository-name/issue-summarization

# copy models from kubernetes
cd build
sudo chown `id -u` .
PODNAME=`kubectl get pods --namespace=${NAMESPACE} --selector="app=jupyterhub" --output=template --template="{{with index .items 0}}{{.metadata.name}}{{end}}"`

kubectl --namespace=${NAMESPACE} cp ${PODNAME}:/home/jovyan/examples/github_issue_summarization/notebooks/seq2seq_model_tutorial.h5 .
kubectl --namespace=${NAMESPACE} cp ${PODNAME}:/home/jovyan/examples/github_issue_summarization/notebooks/body_pp.dpkl .
kubectl --namespace=${NAMESPACE} cp ${PODNAME}:/home/jovyan/examples/github_issue_summarization/notebooks/title_pp.dpkl .

# There's an issue with ip settings of jupyter notebook scripts in some environments.
# See https://github.com/codenvy/codenvy/issues/2427 for details.

# take the start-notebook.sh from the kubeflow repo and put it to the build directory.
cp PATH_TO/kubeflow_repo/components/tsorflow-notebook-image/start-notebook.sh ./

# then edit this script and add the argument --ip=0.0.0.0 to jupyter command.
# to put the script in our serve model image open the Dockerfile and add the following line:
COPY start-notebook.sh /usr/local/bin/start-notebook.sh
before WORKDIR /microservice line

# also you need to copy again scripts and data from notebooks directory (the one inside build dir)
cp notebooks/IssueSummarization.py ./
cp notebooks/Training.ipynb ./

# than correct seldon requirements to make it compatible with current pandas version:
# edit seldon_requirements.txt and change the line with numpy to just numpy without any versions

# build and push service image to our registry
aws ecr create-repository --repository-name github-issue-summarization --region us-east-1
docker build --force-rm=true -t $ACCOUNTID.dkr.ecr.us-east-1.amazonaws.com/github-issue-summarization:latest .
docker push $ACCOUNTID.dkr.ecr.us-east-1.amazonaws.com/github-issue-summarization:latest

# serve the model
# go back to kubeflow_ks_app directory

ks generate seldon seldon --name=seldon
ks apply default -c seldon

# ensure it's up and running

kubectl get pods -n ${NAMESPACE} | grep seldon-cluster-manager

# deploy trained docker image as a deployment

ks generate seldon-serve-simple issue-summarization-model-serving \
--name=issue-summarization \
--image=$ACCOUNTID.dkr.ecr.us-east-1.amazonaws.com/github-issue-summarization:0.1 \
--replicas=3

ks apply default -c issue-summarization-model-serving

# check that we can access the model API

kubectl port-forward svc/ambassador -n ${NAMESPACE} 8081:80

# try to predict summary for an issue
curl -X POST -H 'Content-Type: application/json' -d '{"data":{"ndarray":[[".pyenv/versions/2.7.13/envs/ENV2/lib/python2.7/site-packages/keystoneauth1/adapter.py:136: UserWarning: Using keystoneclient sessions has been deprecated. Please update your software to use keystoneauth1. warnings.warn(Using keystoneclient sessions has been deprecated. Determining IP Address to use with a ping test. Checking ... IP to be used is: INFO: Connecting to Instance at IP: Warning: Identity file gvonlasz not accessible: No such file or directory."]]}}' http://localhost:8081/seldon/issue-summarization/api/v0.1/predictions

Training.ipynb