MLOps | Versioning Datasets with Git & DVC

Amit Kulkarni 19 Jun, 2021 • 6 min read

This article was published as a part of the Data Science Blogathon

“DATA IS LIKE SUNSHINE WE KEEP USING IT AND IT KEEP REGENERATING” – Ruth Porat

Introduction:

The version control is becoming an integral part of the data science project. The research-oriented work in data science makes it even more important to have a system that tracks all the changes that one makes in the data, features, machine learning models, parameters for tuning, etc. With the teams distributed across geographies and open-source collaboration being the norm, every stage of the data science lifecycle is versioned, tracked, shared, and monitored. Even if one works in isolation, having a version control immensely helps to track the changes, make updates or roll back changes. One such popular tool which enables data scientists/analysts to leverage version control is GitHub.

Github uses an application known as Git to apply version control to your code. All the files for a project are stored in a central remote location known as a repository. Its simplistic UI and ease of using commands make it the best fit for versioning the files. But, the data science projects also deal with data files along with code files, and certainly, it is not advisable to maintain let’s say a 50 GB data file and multiple versions on Github. So, is there a way or a workaround to version our data files and keep track of it? Yes, and we achieve this with DVC(Data Version Control).

DVC enables Git to handle large files and directories with the same performance that you get with small code files. The commands like git clone can still be used and it pulls not only code files but also the associated data files in our workspace. Here are the code and data flows in DVC.
Versioning Datasets | git and dvc
                                                                            Code and Data flows in DVC

 

In this blog, we will explore the below feature:

  • Look into DVC and how it works in tandem with Git
  • Create branches and switch between different branches
  • Create experiments and track both the code and data files

Setting Up Working Environment:

We will be using Kaggle’s South Africa Heart Disease dataset. Our objective is to predict chd i.e., coronary heart disease (yes=1 or no=0). A simple binary classification model. To start with, our project folder structure is as below. As we process the files in the later sections, each of these folders/files will be updated.

Please refer to the blog – MLOps | Tracking ML Experiments With Data Version Control to set up the project.

DVC setup: If you don’t have DVC Install then install as below. For more information refer install DVC

pip install dvc

 

To start with, our project folder structure is as below.  You can also clone the repository from Github

|   .dvcignore
|   .gitignore
|   cm.csv
|   dvc.lock
|   dvc.yaml
|   exp.txt
|   params.yaml
|   plots.html
|   README.md
|   requirements.txt
+---data
|   +---processed
|   |       test_SAHeart.csv
|   |       train_SAHeart.csv
|   ---raw
|           SAHeart.csv
+---data_given
|       SAHeart.csv
+---report
+---saved_models
|       model.joblib
---src
    |   get_data.py
    |   load_data.py
    |   split_data.py
    |   train_and_evaluate.py

Data & Model Versioning:

DVC lets capture the versions of your data and models in Git commits, while storing them on-premises or in cloud storage. It also provides a mechanism to switch between these different data contents.

                                                                     DVC tracks the versions of the data & models 

Lets us start with the process:

Step 1: Initiate git and DVC. This would create two folders by name .git and .dvc in your working directory.

git init
git dvc

 

Step 2: Add git remote and map the local directory to Github repo.

git remote add origin <your git repo URL>

Step 3: We will add dataset data_given/SAHeart.csv for dvc to track changes. This would generate a file SAHeart.csv.dvc. Please note that the CSV file is of size 615 KB.

dvc add data_given/SAHeart.csv

Step 4: We have let GitHub know that we are tracking our data file with DVC and here is the dvc file. We do this by add the dvc file to git and commit it. Please note that we are adding .dvc file and not .csv.

git add data_given/SAHeart.csv.dvc
git commit -m "Dataset file size 615 KB"

Step 5: We will create a remote repository for storing the dataset and let dvc know the source of the dataset. This can be on Azure, AWS or google drive, etc. In our case, we will keep it on our local machine and map it to dvc.

dvc remote add -d myremote data_given

Step 6: Once the mapping of remote is completed. We can push the dataset to the remote folder.

dvc push

Step 7: let’s check for the git branch before proceeding further

git log --oneline
# Here is the output
274c7fc (HEAD) 600 KB dataset File

Step 8: Let’s run the DVC pipeline to check if everything works as expected or not. If we execute the DVC pipeline then the corresponding 615 KB dataset file will be used for model building.

dvc repro

Let’s follow the above steps and create one more branch but this time we will use a data_given/SAHeart.csv is of files size 20 MB. I generated synthetic data to add more records to the previous file.  If we execute the DVC pipeline then the corresponding 20 MB dataset file will be used for model building.

Here are the two branches that we have each mapped to different dataset files.

# Here is the output
274c7fc (HEAD) 600 KB dataset File
278fcfb 20 MB dataset File

Key points to note:

  1. We had all our code and config files on Github.
  2. We had both datasets file the 615 KB and 20 MB files in DVC remote repo (not on Github).
  3. On executing the DVC pipeline, the entire cycle of data loading, data splitting, model building that was defined in dvc.yaml was carried out with the respective dataset which DVC is tracking.
  4. DVC is intelligent enough to identify which git version of code is being used and then maps the proper dataset for execution.

Is it possible to run a DVC pipeline with code from one branch and dataset from other?

Yes, certainly possible. As the model building is an iterative process, there may be scenarios where a given branch has some new features in it and we want to check if those new features have any impact on my current model’s performance in another branch. So instead of replicating the dataset and reproducing the new features, we could switch branches and leverage the existing data versions.

Step 1: Now, let’s do a DVC checkout from 278fcfb which points to a 20 MB dataset file.

git checkout 278fcfb data_given/SAHeart.csv.dvc

Step 2: It’s time for DVC checkout which will update the workspace with the data.

dvc checkout

Step 3: Executing the DVC pipeline.

dvc repro

Key points to note:

  1. The DVC pipeline used the dataset and code base from different branches.
  2. Our dataset is a 20 MB file with ~ 4 lakh records and a codebase built a logistic regression model as defined in the pipeline.
  3. Successfully executed DVC pipeline with combination for branches. 

We have learnt versioning of data but is it possible to create versions of ML experiments with DVC and keep track it?

 

DVC Experiments:

Data science and ML are iterative processes that require a large number of attempts to reach a certain level of a metric. Experimentation is part of the development of data features, hyperspace exploration, deep learning optimization, etc. DVC helps you codify and manage all of your experiments.

We will experiment with various parameters params.yaml file eg: test_size, solver, etc and run the experiment

dvc exp run

Once successfully executed, let’s look at the results with the below command.

dvc exp show --no-pager

 

 

Closing Note:

We have just learned about how to control different versions of your data with DVC and how an experiment can be set up and tracked. For every experiment, there is a version generated and metrics logged. I hope this article was useful. Good luck with all your data versioning & experiments !!!!

 

You can connect with me – Linkedin

You can find the code for reference – Github

References:

https://dvc.org/

https://unsplash.com/

Amit Kulkarni 19 Jun 2021

Frequently Asked Questions

Lorem ipsum dolor sit amet, consectetur adipiscing elit,

Responses From Readers

Clear