3.5 Adding a new adapter

Adding a new Adapter

This guide is explaining how to add a new Adapter. It will be describe how we added h2o as a new AutoML solution to OMA-ML.
As a first step copy another Adapter and rename it for your specific solution. We used AutoKeras. It didn't seem to have a lot of special behaviour.
In the following *filename means YOUR_AUTOMLfilename.

Testing the AutoML solution

First we created a running docker container. H2O has a test run to check whether everything is working. You can just replace the code of main() in *Server.py by some test code. This way you can check easily whether all dependencies are satisfied and the container is working fine and confirm the AutoML-Solution is running at all before adjusting the controller etc.
Obviously you can also do it by using the .venv. Add the requirements to the requirements.txt.

Changes to the adapter folder

For the adapter most changes have to be made in the /AutoMLs folder. *Adapter.py is responsible for the training of your model. The *ParameterConfig is mapping the parameter names of the ontology to the parameters of the specific AutoML solution. The *Wrapper includes functionality to make predictions.
For the Wrapper there is a python script created to automatically do predictions with a model. You have to change this in /app-data/templates/predict.ji.
In /config you should update the adapter-name, job-file-name and export-zip-file-name.
In /dependancy-injection/Container.py you have to switch the used *AdapterManager.
In AutoML.py you have to switch the used *Adapter.
TODO: Tests

Setup Ports

In /adapters/YOUR_AUTOML/.vscode/launch.json you have to change the GRPC_SERVER_PORT. Add a new *_SERVICE_HOST and *_SERVICE_PORT to launch.json. For the docker environment you have to add the same values to the controller environment variables in the docker-compose.yml.
Add ":your_automl"(automl tag in the ontology): [*_SERVICE_HOST, *_SERVICE_PORT] to self.__automl_addresses in AdapterRuntimeManager.py and AdapterRuntimeExplainerDashboardManager.py. Of course you have to check that the port is not used yet.

Updating the Ontology

Our Notes - guide has to be extended
Activate in AutoML_solution: Already existing entry set supported_by_oma_ml to true
Add Configuration parameters (with version and description) in ML_Ontology and create a new sheet (H2O_config) (for the new TTL file).
Check usability of Configuration parameters
Setup Configuration parameters \

Updating the docker-compose.yml

As mentioned in Setup ports: Add a new service entry for the adapter in docker-compose.yml Copy an entry for the new autoML container and adjust it accordingly. (name, dockerfile, PORTS, training-volume)