Build a predictive model with Kubeflow Notebooks

Practice with a sample dataset in Jupyter Notebook.

Step 1. Prerequisites

Download the training data and use the dataset to complete the exercise in Jupyter Notebook.

• Check the minimum specifications for the node pool used in the exercise.
• Minimum node pool specifications: at least 4 vCPUs and 8 GB of memory
• For option 2: a GPU notebook node pool is required
• To create a GPU pipeline: a GPU pipeline node pool is required

caution

The exercise might not run properly if the node pool does not meet the minimum specifications.

Prepare training data

In this exercise, you build a taxi fare prediction model using a public dataset from New York City.

• Training dataset: regression model training based on the New York Taxi Fare Dataset

Source dataset information

Item	Description
Goal	Build a taxi fare prediction model
Dataset	New York City Yellow Taxi fare data from 2009–2015 - Pickup and drop-off locations, passenger count, pickup time, and more

Step 2. Create a notebook for the exercise

info

For more information about creating notebooks, see Use notebooks.

Create a CPU image-based notebook

The following example creates a notebook that uses only a CPU instance type. Configure the notebook as required for the exercise.

1. On the Kubeflow dashboard, click the Notebooks tab, and then click New Notebook.

2. On the New Notebook page, enter the required information and click LAUNCH to create a CPU image-based notebook instance.

   Create a CPU image-based notebook (option 1)

   Item	Field	Description
   Name	Name	Name used to identify the notebook instance on the Kubeflow dashboard
   	Namespace	Kubernetes namespace where the notebook instance is created
   Docker Image	Image	mlops-pipelines/jupyter-pyspark-pytorch:v1.0.1 - Docker image to use
   CPU / RAM	Requested CPUs	2 - Number of CPU cores allocated to the notebook instance
   	Requested memory in Gi	6 - Amount of memory in GiB allocated to the notebook instance
   GPUs	Number of GPUs	None - GPU resources allocated to the notebook instance
   Affinity / Tolerations	Affinity Config	CPU node pool where the notebook is created - Specifies the node on which the notebook instance runs
   	Tolerations Group	None - Configuration that permits specific node taints

caution

The Requested CPUs and Requested memory in Gi values must be less than the vCPU count and memory capacity of the node pool selected under Affinity Config.

Create a GPU image-based notebook (option 2)

The following example creates a notebook that uses only a GPU instance type. Configure the notebook as required for the exercise.

1. On the Kubeflow dashboard, click the Notebooks tab, and then click New Notebook.

2. On the New Notebook page, enter the required information and click LAUNCH to create a GPU image-based notebook instance.

   Create a GPU image-based notebook (option 2)

   Item	Field	Description
   Name	Name	Name used to identify the notebook instance on the Kubeflow dashboard
   	Namespace	Kubernetes namespace where the notebook instance is created
   Docker Image	Image	mlops-pipelines/jupyter-pyspark-pytorch-cuda:v1.0.1 - Docker image to use
   CPU / RAM	Requested CPUs	2 - Number of CPU cores allocated to the notebook instance
   	Requested memory in Gi	6 - Amount of memory in GiB allocated to the notebook instance
   GPUs	Number of GPUs	At least 1 - GPU resources allocated to the notebook instance
   	GPU Vendor	NVIDIA MIG - 1g.10gb - GPU driver and software toolkit to use
   Affinity / Tolerations	Affinity Config	GPU node pool where the notebook is created - Specifies the node on which the notebook instance runs
   	Tolerations Group	None - Configuration that permits specific node taints

Step 3. Train a predictive model in a notebook (exercise 1)

1. Download the following sample file.

   • Sample file: nyc_taxi_pytorch_run_in_notebook.ipynb

2. Access the notebook instance you created and upload the file in the browser.

   Upload a file in the Jupyter Notebook console

3. After the upload is complete, review the exercise in the right pane.

   Uploaded sample file

4. Follow the sample instructions to train the model.

Step 4. Create a pipeline and train a predictive model in a notebook (exercise 2)

1. Download the following sample files.

   • Sample files: nyc_taxi_pytorch_build_pipeline_cpu.ipynb / nyc_taxi_pytorch_build_pipeline_gpu.ipynb

   Upload a file in the Jupyter Notebook console

2. After the upload is complete, review the exercise in the right pane.

   Uploaded sample file for exercise 2

3. Follow the sample instructions to train the model.

   info

   We recommend deleting completed or unused runs as described below to conserve resources.

4. After completing the exercise, delete the completed run. Move the run to Archived, and then click Delete.

   Delete a run

5. After deleting the run, verify that its pods were also deleted.

   Verify run deletion