literal-scatter-gather-tutorial.md

Adapt our sample "literal" code to your needs

IMPORTANT: the "literal" code available in this repo has been intentionally designed to:

• provide an effortless setup to get started.
• rely only on features that are currently generally available in AzureML SDK v2.

This tutorial addresses the following scenarios:

• To add/remove a silo:

  • You just need to make the changes in the "federated_learning/silos" section of the examples/pipelines/fl_cross_silo_literal/config.yaml file.

• To change training hyper-parameters:

  • Adjust the parameters in the "training_parameters" section of the examples/pipelines/fl_cross_silo_literal/config.yaml file.

Tutorial on how to adapt the "scatter-gather" code

Please read the following points to have a better understanding of the "scatter-gather" code:

• It has a set_orchestrator method that you can leverage to add an orchestrator to your pipeline.
• The add_silo method lets you add n number of silos to the pipeline and you don't have to worry about the configuration.
• It has a soft validation component that ensures that the appropriate permissions are granted for your assets. That being said, a compute a should not have access to dataset b and so on.
• You can bypass the validation if you have your own custom rules.
• Enabling type-check, ensures that no data is being saved and only model weights are kept in the datastore.

This tutorial addresses the following scenarios:

• To add/remove a silo:

  • You just need to make the changes in the "strategy/horizontal" section of the examples/pipelines/fl_cross_silo_scatter_gather/config.yaml file.

• To change the training hyper-parameters:

  • Adjust the parameters in the "inputs" section of the examples/pipelines/fl_cross_silo_scatter_gather/config.yaml file.

• To edit the flow of your training pipeline:

  • Pass your custom subgraph as a parameter to the scatter_gather method in the examples/pipelines/fl_cross_silo_scatter_gather/submit.py file.

• To bypass the soft validation:

  • Use --ignore_validation argument while executing the examples/pipelines/fl_cross_silo_scatter_gather/submit.py file.

• To enable multiple computes(CPU for preprocessing & GPU for training):

  • Set the compute2 parameter to true while provisioning the resources.(No further changes are required)