What’s a revision?
Revisions in Slai are similar to branches in git. For each model, you can have several revisions, each with a unique name. Each revision has its own sandbox, model handler, and model artifacts. When you’re working on a particular model, you’re always working on a specific revision of that model. The default starting revision is named initial. You can see and create revision from the drop-down next to the model name at the top of your sandbox:
It is important to note that each revision can only have one deployment running at any given time.
Why might you want to revision? Here are a few use cases to get you started:
- A/B testing model performance: Create a revision and train a variant of the original. Deploy each revision, divide traffic between the two models, and see how they perform against each other.
- Collaborate with your team: Signal what you’re working on with a descriptive branch name. For example
ignore-outliersorbroaden-dataset - Creating a set of similar, but slightly different models: You might start with a strong general image classifier like ImageNet and concurrently deploy several variations that make different tradeoffs.