Python Developer / Client Example#

Usage of Federated Model Aggregation service#

A developers guide to using federated model aggregation service.

Users must have a deployment of the FMA service setup and running. Directions for doing that locally can be found in this repo under Django Deployment .
As part of the system mentioned above, an admin (superuser) must be established to access the service as a developer.

To create a weights file that contains fresh weights for dataprofiler with the correct labels run python3 create_initial_model_weights.py
This will create a file with the weights you need in the form of a json file at create_initial_model_weights.json

Once you have started the service locally on your browser the service will ask you to login

Provided users use the default deployment, /api/v1/models is the default endpoint users can find the model aggregation setup. The default endpoint is http://127.0.0.1:8000/api/v1/models
- http://127.0.0.1:8000 is swappable and may differ depending on your implementation.
- You can edit pre-existing models here by clicking on the name of them in the list
This will take you to a page called “Federated Model List”
- “Initial model” field: This is where the json formatted values for your model go (usually weights from previous above step)
- “Name” field: The name of the model aggregation
- “Requirement” field: A dropdown list of additional requirements functions for aggregation that are explained below
- “Requirement args”: A list of arguments for the extra requirements function above
- “Aggregator”: A field to fill out the aggregation function you wish to use for your client models
- “Developer”: (should see your username)
- “Clients” field: a selectable list that shows all the clients subscribed to your service
  - This can be left alone if there are not specific clients with which you wish to interact for this model.
Once all relevant fields are filled out, click “POST” and your model aggregation is created.

A demonstration of the federated model aggregation service for the dataprofiler.

Make sure the FMA service is up and running via this server setup readme
Upload custom model weight extraction function to model_utils.py
Run example, specifying the desired number of clients (n=3), user (abc123), the base url the service is at (http://localhost:8000), the number of epochs to run the experiment (1), and the train entities (None) to be included in the experiment.

bash run_example.sh 3 abc123 http://localhost:8000 1 None

Generates initial model weights and schema via create_initial_model_json_files.py
POSTS model registration
Executes run_client.sh (n) times