Data Versioning

Version control is an important component when working with Machine Learning. It also increase the speed of development and reduce mistakes

• Increase the speed of development
  • Reproduciblity
  • Data Sharing
  • Change Comparison
• Reduce errors
  • Data Provenance - How the data derived
  • Revert to the correct version, when you accidentally change something
  • Debugging

Tools

There several ways to implement data versioning ranging from full duplication, to space-efficience approaches, to more advance tools (# Best 7 Data Version Control Tools - Neptune Blog, # Comparing Data Version Control Tools - DagsHub, # Top 14 Data Versioning Tools - StartupStash)

Example data tools comparison from DagsHub

One of them is DVC which is open-source, lightweight, storage-agnostic, and Git-compatible.

DVC

What is DVC?

As mention in their website. DVC is an open-source version control system for Machine Learning Projects which is Git-compatible, Storage agnostic, Lightweight, and others.

When to use DVC?

Refer to DVC Use Cases

How?

Refer to DVC Documentation which is very clear and easy to follow.

Personally, I'll split the How section into 3 main parts.

• Data Version Control - Including the concepts when dealing with data, versioning data with DVC, sharing data across person, or projects, and Data Registries
• Machine Learning Pipeline - Connect ML steps with dvc stage, reproduce pipeline with dvc repro, compare versioning with dvc params, or dvc metrics, and visualize (plot) the time-series data with dvc plots
• Experimentation - with DVC 2.0, we can easily track experiments, compare, checkout, apply, and also sharing experiments using dvc exp

1. Data Version Control

Commands

• dvc init
• dvc add
• dvc remove
• dvc commit
• dvc checkout
• dvc diff
• dvc remote
• dvc fetch
• dvc push
• dvc pull
• dvc status
• dvc list
• dvc get
• dvc get-url
• dvc import
• dvc import-url
• dvc update

Reference or Related

• Get Started with Data and Model Versioning
• Get Started with Data and Model Access
• Set up a Google Drive DVC Remote
• External Dependencies
• .dvc Files
• .dvcignore Files

2. Machine Learning Pipeline

Commands

• dvc run
• dvc stage
• dvc dag
• dvc repro
• dvc destroy
• dvc metrics
• dvc params
• dvc plots

Reference or Related

• Get Started with Data Pipelines
• Get Started with Metrics, Parameters, and Plots
• Pipeline Files (dvc.yaml)
• How to Add Dependencie or Outputs

3. Experimentation

Commands

• dvc exp

Reference or Related

• Get Started with Experiment
• Experiment Management
• Machine Learning Experiment Tracking

Interesting Resources

• Data Versioning and Reproducible ML with DVC and MLflow shares how to use both tools to gether, so we can track data using DVC and track experiment using MLflow
• Reproducible Machine Learning & Experiment Tracking Pipeline with Python and DVC shares a step by step on how to use DVC pipeline by using Kaggle dataset (Note starting around 28:20)
• Experience report: Data Version Control (DVC) for Machine Learning Projects shares pros and cons when using DVC with their workflow. Interestingly, they also create their own tools (e.g., paired with pre-commit), and also use DVC to versioning entire Jupyter Notebook.