Kubeflow Pipelines

This guide will take you step-by-step through the process of fine-tuning BERT on the SQuAD dataset, leveraging Kubeflow Pipelines to scale and automate the experiment in a Kubeflow cluster.

There are several advantages to this approach compared to running the experiment locally:

  • 🚀 Scale your runs by leveraging more resources and more powerful machines.

  • 🎏 Parallelize steps that can run independently.

  • 🫙 Cache steps, such as data processing, to avoid repeating them on each run.

  • 📅 Schedule recurring runs to retrain your model periodically.

  • 📈 Track and visualize the experiment’s configuration.

  • ✨ Automate model deployment by integrating KFP with CI/CD pipelines once.

Here’s how a complete pipeline looks like:

run

What you’ll need

Before you start, make sure you have the following:

  • A working Kubeflow deployment. Visit the VirtML project page to find out how you can create a local Kubeflow deployment.

  • A basic understanding of Kubeflow Pipelines. If you’re new to KFP, check the official documentation.

Procedure

  1. Create a new Jupyter Notebook in your Kubeflow Notebook server. Make sure that the server can submit pipelines to the Kubeflow cluster.

  2. Connect to the Jupyter Notebook server.

  3. Launch a temrinal window and clone the repository:

    user:~$ git clone https://github.com/dpoulopoulos/bert-qa-finetuning.git

  4. Navigate to the project directory:

    user:~$ cd bert-qa-finetuning

  5. Install the required packages:

    user:~/bert-qa-finetuning$ pip install -r requirements.txt

  6. Open the pipeline.ipynb notebook.

  7. Follow the instructions in the notebook to create, compile, and submit a Kubeflow Pipeline for fine-tuning BERT on the SQuAD dataset.

  8. Create a TensorBoard instance to monitor the training process. You can submit the following YAML manifest, or use the UI to create one:

    apiVersion: tensorboard.kubeflow.org/v1alpha1
kind: Tensorboard
metadata:
  name: bert-squad-logs
  namespace: kubeflow-user-example-com
spec:
  logspath: pvc://bert-squad/logs

  9. Access the TensorBoard instance to monitor the training process.

tensorboard

Next steps

Congratulations! You’ve successfully created and submitted a Kubeflow Pipeline to fine-tune BERT on the SQuAD dataset. You can now scale and automate the experiment in your Kubeflow cluster.