Dataclass Overview

DIAL uses the INTERSECT-SDK as its externally-facing API; all public functionality is marked in src/dial_service/dial_service.py with @intersect_message decorators. For all decorated functions, the function has a single input parameter (excluding self); the typing of this parameter represents the structure of the input you need to provide to DIAL. The structure of the return value is the output you will be given.

There are a few ways to interact with DIAL:

  • Using the iHub web interface to construct a campaign which calls into DIAL. This is the primary option to use in production, and should be what is used by any user without an extensive programming background. The capability name should be dial (the exact value is the value of intersect_sdk_capability_name in the DIAL service). iHub is able to directly display the DIAL API (including descriptions) in the browser, and can automatically help your campaign adhere to the input/output schemas.

  • Using python scripting to create an INTERSECT-SDK Client (your own custom orchestrator) and using it to make requests to DIAL directly. See the client_guide.html for an in-depth guide on how to do this. This option is more useful for development, but it requires you are able to spin up an entire INTERSECT ecosystem yourself.

Creating a workflow

  1. You need to call initialize_workflow with your workflow parameters to create a workflow. This will return a workflow_id, which you will need to provide in all subsequent calls.

  2. To do active learning, you can call the get_next_point, get_next_points, or get_surrogate_values endpoints. This only returns _suggestions_, it does not update the workflow directly.

  3. To update the workflow, you can then call the various update_workflow endpoints, which will actually update the workflow state.

Active learning models are encapsulated from external view, but are only trained when calling initialize_workflow and the various update_workflow endpoints. This means that calling the active learning endpoints should generally be cheap, and you can call active learning multiple times before updating the workflow with your selected data.

References