Collaboration in ML made easy with W&B Teams
How to use W&B Teams on your team's next machine learning project
Created on March 17|Last edited on March 17
Comment
Everyone on your team cheering as your model converges
You know what's better than watching your model converge? Everyone else on your team watching your models converge with you.
Bask in that glory when you finally beat your predecessor's best performing model. Shout it from the rooftops and pin a link to your W&B project in every Slack channel!
A few of the nice things about W&B Teams:
- You have a shared workspace for all of your teams projects.
- Keep all of your teamwork in one place instead of sprinkled around presentations in a shared drive or in Slack messages.
- You can spread the word about your latest findings using W&B Reports and get comments from your team.
- It's easy to keep up to date and see what others are working on.
Run set
21
Case Study: Collaborative Training of DALL·E mini
The dalle-mini team has worked on reproducing the results of OpenAI's DALL·E with a smaller architecture and used W&B pretty extensively throughout the process.
Run set
4
They needed to make sure their experiments were reproducible and that they could easily collaborate.
Here are a few things they used:
A Shared Workspace
This is where they watched their models train and checked whether it was finished training or had crashed. You can see each time they trained models, all of the configurations of each run and compare metrics across runs.
Sharing their latest model predictions
The team put the predictions their latest model produced in a W&B Table so they could also evaluate how it was performing:
Run: 2wo1vjkg-clip16
1
Sharing results, thoughts and next steps
They used Reports to explain some of the ideas that went into making this project a success and some areas for improvement. This was aimed at helping people reproduce their work and pick up from where they've left off because they've captured all of the key concepts needed to understand their work.
Here's their writeup of their findings in a W&B Report:
Squeezing all of juice out of their models
By using W&B Sweeps, they were able to find the optimal learning rate their model needed. Sweeps allowed them to efficiently search across possible learning rates, rather than wasting time trying all of the different values manually.
Sweep: 5igzdwfv 1
6
Run set 2
6
Create a project for your model and add some runs!
- Click Create Project on your team page.
- View the results in your project workspace.
- Create a report to share progress with your team.
Already using W&B?
Send new runs to the team
In your script, set the entity to your team name. "Entity" just means the owner of the project. You can make the team your default for new runs on your Settings page.
wandb.init(entity="example-team")
Move runs from one of your personal projects to your team.
Conclusion
Summing up, if you care about tracking your work and collaborating closely with other ML practitioners, you should give W&B Teams a try.
W&B has everything you need to collaborate with your ML team:
- Experiment Tracking for viewing models training in a shared workspace
- Reports for sharing results and planning next steps
- Artifacts to save and version models and data
- Tables to see latest model predictions
- Sweeps for hyperparameter optimization
Good luck and thanks for reading!
Add a comment