Each Aim is has one serving API, which you can use to get predictions. For this aim, the API is
https://....../Day2Example?features=[...]
https://....../Day2Example?batch=[...]
The API will return the predictions of all deployed models:
[{ model1: "yes", model2: "no", ...}...]
Deployed models are a collection of models that are used for each call to the serving API.
All deployed models are stored in the branch deployed_models.
All Deployed Models are choosen from a set of deployment candidates,
which are stored in the branch deployment_candidates.
A user can manually choose to commit deployment candidates to deployed models.
A user can call Model picker, which takes as input the current deployment candidates
and deployed models, and produce a new set of deployed models. This creates a new
commit in the branch deployed_models.
Each Trial is identified by one unique training set uploaded by the user.
Each trail has its own branch. Each commit in the trial branch contains:
- training set: The state of the training set
- search space: The search space of AutoML
- models: The results of all models
- other meta data, e.g., labeling functions
Every call to "What's Next" will create a set of suggested transformations. Each transformation takes as input one commit in a trail and bring it to a new commit:
- new training set: after data cleaning, labeling etc.
- new search space
- new models
- new meta data
The user might choose to move some models in each trial to the deployment
candidate branch. This creates a new commit in the deployment_candidates branch
and is where CI/CD gets triggered.
All trials in an Aim shares the same validation dataset, which is stored in
its own branch validation_set.
Each call to the serving APIs will be logged and use as the fresh production set. By defaulted they are not logged in git.