rsilvery / kf_demo_mnist

Demo of KubeFlow

This demo is to train and serve a TensorFlow MNIST model. Training is done in the notebook container on data in the MapR volume and model is output back to MapR storage.

Train Model

• Launch Jupyter instance with TensorFlow from JupyterHub (located on port 8000)
• Open mnist.ipynb notebook
• Set your training_data and model_output directories to the correct location in the MapR filesystem.
  • Note that the mount point is under /home/jovyan so if your directory was mounted as training_data it will be /home/jovyan/training_data
• Run training job

Serve Model

• Copy model.py executable code to the same directory as your model: (ex. /user/mapr/kubeflow/models/mnist)
• Run Kubectl with serve_mnist.yaml

  kubectl create -f serve_mnist.yaml
  

  This uses the mount point from the kf-pvc Persistent Volume which must point to the directory in MapR-FS which contains the model and model executable code (model.py).

Test Model

...to be continued.