demos

Getting Started with the MLOps for Good Hackathon

The Hackathon focuses on MLOps. It is important all submissions address the challenges of operationalizing the machine-learning project, and not just focus on the mode innovation. MLOps typically includes the following steps:

Business problem definition
Data ingestion and acquisition
Training pipeline
Model serving
Automated CI pipeline

This guide will help you get started.

0. Setup

Go to the Devpost hackathon page and click the “Join hackathon” button. After you login/signup to the Devpost website you will see all existing projects and you will be able to start your own project.

Collaboration, Resources + More to Help You Build

Whether or not you have started building your project, the Slack channel is the perfect place to connect with other developers!

Not only can you bounce ideas off each other and ask other developers for technical help, but you can also use the Slack Channel to ask the Iguazio team questions!

Teaming up

Looking for a team? The Slack channel can help you connect with potential teammates. You can also check out the Participants Tab here to search for other practitioners, check out their skills and connect with them.

MLRun

MLRun is an open-source end-to-end MLOps framework that will significantly help you get your data science to production for this hackathon. To get started, install and go through some of its basic examples.

Resources:

Managed Environment

Iguazio is offering several pre-configured clusters in on Azure Cloud, for free. If you’d like to access one of these clusters, send an email to hackathon@iguazio.com. Alternatively, you can set up your own cluster. To do that, check out the MLRun installation guide.

1. Business Problem Definition

This is a crucial step that will impact the success of your project. Don’t skip this step. Take the necessary time to think about the problem you would like to address in your project.

Don’t worry if you don’t come up with ideas right away or if your ideas sound far-fetched at first. It’s best to not limit yourself at the first stage, so write down every idea that comes up.

After you feel you have enough ideas to consider, evaluate each idea. Some of the criteria you should evaluate:

Alignment with the MLOps for Good social and environmental goals
Feasibility to complete within the Hackathon timeframe.
Availability of data.
Availability of existing models that address this problem.
Availability of team members with the required skillsets to complete the project.

You may find that you have more than a single idea that is worthwhile. Feel free to create more than one project or gauge the interest of other participants. You may only be a member of a single project, but it can be fun to see someone else bring your idea to reality.

2. Data Ingestion and Acquisition

Data Sources

Machine learning requires data, and therefore you should research what data sources are available for your project. Consider the size of the data: if the dataset is too small, this may limit your ability to train your model. You should also think ahead about the serving process, and whether you can obtain new data for inference.

We recommend researching publicly available datasets, such as https://github.com/awesomedata/awesome-public-datasets. Before starting to use the dataset, please verify that the dataset owner gave the proper license/permission to use this dataset. If in doubt, contact the Iguazio team.

Additional open datasets:

While it is always best to build a solution based on real data, you may find the best way to implement your project is to generate data. In some cases, the generated data is needed to enrich an existing dataset, in other cases, where data is scarce, you will need to generate the entire dataset. Data generation takes time and there are a lot of tools available for creating datasets. Do your research and try to find the tool that best fits your needs and skillset. Some tools may be generic, and others may be more specific. Make sure to include all the source code and documentation how to generate the files.

Storage

If your data is not large and just a few files, likely the simplest approach is to read the files directly.

In cases where you need to store large amounts of data, it is best to consider an object storage, such as Azure Blob Storage. If you need to access data using different patterns, check out MongoDB Atlas.

Iguazio is offering free access for hackathon participants to its multi-model data layer to store and analyze various types of data structures — such as NoSQL (“key-value”) tables, time-series databases (TSDB), data streams, binary objects, and files. This is commonly referred to as V3IO. You can access your assigned V3IO data layer by specifying the V3IO_API, V3IO_USERNAME and V3IO_ACCESS_KEY environment variables. Send an email to hackathon@iguazio.com to get your credentials.

If you have other specific data access needs, first define the requirements and it’s likely you’ll find an open-source project or a service that can be used for your project.

3. Training pipeline

Feature Engineering

This is a common step when dealing with a machine-learning project. Deep learning projects sometimes do not require special feature engineering. For example, a neural network can identify key areas in an image without special features.

If you have to perform feature engineering, remember that the training pipeline is not just the model training. You may have some data manipulation to perform. Therefore, the data processing is still key to this step. You may not know ahead of time which features your model will need, so this is an iterative process where you create a set of features, train some models, and then consider other features that may be useful. Consider not just the feature definition, but also how you would serve the features.

Finding the Right Model

It’s usually a good idea to try out several algorithms with different hyperparameters when training models. While it’s great if you come up with an innovative model, it’s a good idea to look for existing work to see if there’s a relatively good model that you can use. Remember that the primary focus of this hackathon is not the best model, but rather the best way to operationalize the workflow.

If you would like to take a more advanced approach, consider creating a model ensemble.

Resources to consider for model training:

4. Model serving

Even if you have built the most robust training pipeline, real-world scenarios require running the model as part of an application. This step is where you would gain the most benefit from using an MLOps framework early on.

Model serving requires getting input data. In many applications this is an online source where the model provides some output (e.g., prediction) based on the stream. Other applications process files in batch, so you would need to consider where to put the input files, how to trigger the pipeline, and where to store the output.

You should also think about the interaction the user can have with the model. This is usually something simple, such as a basic web page or a dashboard (e.g., Grafana).

MLRun has serving and data pipeline capability which should make it easy to deploy the model. You can deploy your pipeline in a few lines of code.

5. Automated CI pipeline

You may want to go the extra mile and create an automated pipeline. For example, using GitHub actions to trigger model training when you provide a new training file or deploy a model to serving once you train a new model or running a GitLab CI job.

See the MLRun documentation to create a CI pipeline with MLRun.

Summary

While it is tempting to start writing code right away, you’ll thank yourself later if you take some time to research. Using existing publicly available data and models will significantly reduce the amount of effort to bring your project to production.

We also recommend being active in the Slack channel. Talking about your project with other people can help everyone generate ideas together.

This site is open source. Improve this page.