Project-MONAI
Provide a mechanism to run a batch of jobs for a single App via App Runner
User Story As an application developer I would like to run a batch of jobs for a single application via the application runner so that I can verify the behavior of the app with a diverse set of inputs
Background MONAI App SDK provides a Python API to develop apps with AI models. As part of the app development, the developer often likes to verify how the app would function given a diverse range of input datasets. Ability to perform it in a straight forward way in a development workstation accelerates development and provides a tight feedback loop to the developer
Success Criteria The App Runner enables the developer to do the following
- Allow the user to specify a set of directories each containing data for a single job based on a single app
- Start the app runner in a batch mode with the inputs as set of directories
- The app runner then would start the jobs in a batch mode and generate output
The runner may not know the semantics of "a set of folders" as it has to generically run all possible MAP's. The best way is for the application, or rather, the function/module which will also serve as the entry point to the MAP container image, to parse and process a set of data inputs per its own expectation.
The mechanism to support batch processing has been discussed in the App SDK "Show and tell" topic, Techniques to run MONAI Deploy App in batch mode or in a long running service, and the built-in application and operator classes have been designed and implemented to support this, i.e. internal state is not affected by the data being processed (arguably stateless).
