MNIST DIGITS CLASSIFICATION EXAMPLE

Create code repo

• Name: dkube-examples
• Project source: Git
• Git URL: https://github.com/oneconvergence/dkube-examples.git
• Branch: tensorflow

Create dataset repo

• Name: mnist
• Dataset source: Other
• URL: https://s3.amazonaws.com/img-datasets/mnist.pkl.gz

Create a model

• Name: mnist
• Keep default for others

Launch Notebook

• Create Jupyterlab IDE with tensorflow framework.
• Select the Code dkube-examples.
• Repos->Inputs->Datasets: select mnist and enter mountpath as /mnist.
• Run workspace/dkube-examples/mnist/train.ipynb
• You can experient in the notebook and develop your code. Once you are ready for a formal run, export your code into python script(s)

Run training job

• Runs->+Training Run.
• Code: dkube-examples
• Framework: Tensorflow
• Version: 2.0.0
• Start-up script: python mnist/train.py
• Repos->Inputs->Datasets: select mnist and enter mountpath as /mnist
• Repos->Outputs->Model: select mnist and enter mountpath as /model
• Submit

Katib based Hyperparameter Tuning

1. Create a Run same as explained above, except that now a tuning file also needs to be uploaded in the configuration tab.

• For hyperparameter tuning upload the https://github.com/oneconvergence/dkube-examples/blob/tensorflow/mnist/tuning.yaml under upload tuning definition.
• Submit the run.

Tuning.yaml file Details:

1. objective: The metric that you want to optimize.

2. goal parameter is mandatory in tuning.yaml file.

3. objectiveMetricName: Katib uses the objectiveMetricName and additionalMetricNames to monitor how the hyperparameters work with the model. Katib records the value of the best objectiveMetricName metric.

4. parameters : The range of the hyperparameters or other parameters that you want to tune for your machine learning (ML) model.

5. parallelTrialCount: The maximum number of hyperparameter sets that Katib should train in parallel. The default value is 3.

6. maxTrialCount: The maximum number of trials to run.

7. maxFailedTrialCount: The maximum number of failed trials before Katib should stop the experiment.

8. algorithm: Search algorithm to find the best hyper parameters. Value must be one of following:

   • random
   • bayesianoptimization
   • hyperband
   • cmaes
   • enas

Deploy Model (DKube version 2.1.x.x)

• Repos->Models->mnist: select a model version
• Deploy
• Name: mnist
• Type: Test
• Transformer: True
• Transformer script: mnist/transformer.py
• Submit

Publish and Deploy Model (Dkube version 2.2.x.x)

• Repos->Models->mnist: select a model version
• Click on Publish model icon under ACTIONS column.
• Name: mnist
• Transformer: True
• Transformer script: mnist/transformer.py
• Submit

Deploy model

• Click on Model catalog and select the published model.
• Click on the deploy model icon under ACTIONS column.
• Enter the deploy model name and select CPU and click Submit.
• Check in Model Serving and wait for the deployed model to change to running state.

Publish and Deploy Model (Dkube version 3.0.x.x)

• Models->mnist: select a model version
• Click on Publish model icon under ACTIONS column
• Transformer: True
• Transformer script: mnist/transformer.py
• Submit

Deploy model

• Click on Models in the navigation pane
• Click on the drop down next to 'Owned by me' and select 'Published'
• Click on the published model 'mnist'
• Select the published version and click on the deploy model icon under ACTIONS column
• Enter the deploy model name, select Deployment / Test and select Deploy using / CPU. Click Submit
• Check in Deployments and wait for the deployed model to change to running state

Test inference

• Go to
  • Deployments in 2.1.x.x version
  • Model Serving in 2.2.x.x version
  • Deployments in 3.0.x.x version
• Copy the prediction Endpoint for the model
• Create a browser tab and go to https://<dkube_url>/inference
• Paste the Endpoint URL
• Copy Auth token from Developer settings in Dkube page and Paste in inference page
• Choose mnist
• Upload 3.png from repo
• Click predict

Automate using pipelines

Run this pipeline to automate training and serving using kubeflow pipelines.