Skip to content

Latest commit

 

History

History
81 lines (49 loc) · 2.82 KB

TUTORIAL.md

File metadata and controls

81 lines (49 loc) · 2.82 KB

Tutorial

Step by step guide to create this server based on a scikit-learn

Generate server source code from OpenAPI

There is a maven command to execute swagger code generation:

mvn -P generate generate-source

Correct generated code

At the time of writing, the openapi code generator doesn't correctly support anyOf. While waiting for the a bug fix, we suggest the following workaround:

Remove imports in openapi_server/models/parameter.py

from openapi_server.models.any_ofintegernumberstringboolean import AnyOfintegernumberstringboolean

Add instead

from openapi_server.models.check_and_get_type import check_and_get_type

Replace AnyOfintegernumberstringboolean type by check_and_get_type('Parameter.value',int,float,str,bool)

In the automatically generated file openapi_server/util.py, add Line 18:

        if klass.__name__=='_check_and_get_type':
            klass=klass(data)
    except AttributeError:
        pass

Finally copy and paste openapi_server/models/check_and_get_type.py to your model folder.

This suggested workaround has been applied in this service, please feel to read source code in doubt.

Model change

Except the workaround described above, to support anyOf, no model has been changed.

Update the info controller

Add import Capability in get_capabilities() replace return 'do some magic!' by

   return Capabilities([Capability.INFO, Capability.DISCOVER, Capability.RUN])

in get_info() replace return 'do some magic!' by

    info = dict(
        description='Open Prediction Service for Scikit Learn models based on OPSv2 API'
    )
    return ServerInfo(status='ok', info=info)

Update the discover controller

The discover controller gathers, at server initilization, the available models in the server.

This is implemented in openserver/controllers/helper.py where the supported_models list gathers all the directories with a model.pkl

The discover controller, in this example's implementation, serves the endpoints and models requests based on the data available in the file deployment_conf.json that is in the same directory as the model.pkl.

Update the run controller

The run controller has been modified, the function prediction in openapi_server/controllers/run_controller.py has been updated to take the argument body instead of the default prediction that doesn't work.

The run controller simply executes the pickle file model.pkl of the model requested by the client with the supplied arguments

Remove the manage controller

As this implementation does not manage anything, remove the file openapi_server/controllers/manage_controller.py and remove all endpoints tagged manage in the openapi_server/openapi/openapi.yaml to ensure that the generated swagger UI doesn't display them.