Introduction to Data Versioning with DVC
Learn to version machine learning data and models with DVC, share them through remotes, and run training pipelines that track metrics and plots.
Datasets and models in machine learning projects are often too big for Git. Data Version Control (DVC) versions them beside a Git repository, using a cache of MD5-hashed file versions and separate remote storage. You see why data versioning differs from code versioning, store and share versions of datasets, and automate model training as a pipeline whose metrics and plots DVC tracks.
What You Will Achieve
- Start versioning a machine learning project with DVC inside an existing Git repository
- Check out an earlier version of a dataset kept in a DVC remote
- Push versioned data to a DVC remote, then pull it or fetch it back into the cache
- Rerun model training through a DVC pipeline and compare model performance across its stages
Program Overview
Data Versioning and DVC Foundations
Why machine learning data needs its own versioning, shown on a simple classification problem, with a review of the Git and DVC command line.
- Data versioning compared with code versioning
- Model anatomy and hyperparameters
- Git CLI and DVC CLI basics
- DVC use cases, including experiment tracking and CI/CD for machine learning
DVC Configuration, Cache and Remotes
Setting up a repository and managing how DVC stores, stages and shares file versions.
- Installation, initialization and .dvcignore patterns
- Cache, staging and .dvc files with MD5 hashes
- DVC remotes compared with Git remotes
- Pushing, pulling, fetching and checking out data
Pipelines, Metrics and Plots in DVC
Refactoring training code into a pipeline described as a directed acyclic graph, then running it and evaluating the model.
- Code organization and parameter files
- Pipeline stages with dependencies, commands and outputs
- Local execution and Git tracking of DVC metadata
- Metrics tracking and plots in dvc.yaml
Certificate Awarded (Paid)
Job Outlook
DVC skills apply in machine learning work where datasets, models and hyperparameters are versioned alongside the code that uses them.
Good for: Machine learning practitioners who want their datasets, models and training pipelines versioned alongside their code.
Keep in mind: Pipelines run locally in the lessons; CI/CD for machine learning appears only as one of the use cases surveyed in the opening chapter.
Bottom line: A focused course on the DVC cache, remotes and pipelines that ends with tracking model metrics and plots inside a training pipeline.
Frequently Asked Questions
How does DVC keep large data files out of Git?
DVC identifies each file version by an MD5 hash and stores the data in its own cache and in remote storage. Git tracks the .dvc files and DVC metadata instead of the data itself.
How are DVC pipelines described?
The pipeline chapter treats a pipeline as a directed acyclic graph of stages, each defined by its dependencies, command and outputs. Settings and hyperparameters live in a separate parameter file, and plot definitions are added to dvc.yaml.
What kind of machine learning work do the exercises use?
The opening chapter uses a simple classification problem to show why data needs versioning. The pipeline chapter executes a model training pipeline locally, then prints metrics and builds plot files to evaluate the model.


There are no reviews yet.