Allow access to artifacts in the service __init__(self) during pack process

[shihgianlee]

Is your feature request related to a problem? Please describe.
In BentoML service class, we don't have access to artifacts during pack process, e.g. self.artifacts.my_artifact, in the __init__(self) method. The reason to have access to the artifacts in the init is so that I can pass the artifact to my class initialization, e.g. self._provider = Provider(self.artifacts.my_artifact). Later in the service predict method, I can access my class this way, e.g. self._provider.fetch_predicted_value, where the predicted value is computed value from the prediction produced by my_artifact.

Describe the solution you'd like
In slack discussion message thread, a service_init callback is proposed to allow the above service initialization.

Describe alternatives you've considered
In slack discussion message thread, a couple solutions were proposed. I ended up using @property decorator for my use case. It is a bit more code but it achieved the end result described above and made my code simpler.

@property
def provider(self):
    if not self._provider:
          self._provider = Foo(self.artifacts.my_csv, self.artifacts.my_json) 
    return self._provider

[shihgianlee]

We encountered the same problem again where we need to load a json config file within the same package but couldn't. We had to use the same workaround for our config, which is awkward. I see help wanted label in this issue. Is there anything I can do to help with this issue?

[stale]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[parano]

This is now possible in BentoML 1.0, where models are first saved in local model store before the Bento build process. User can also create a custom Runner class, and define the model initialization process:

import bentoml

class CustomRunner(bentoml.Runner):

    def _setup(self):
        # load model from model store
        model_info = bentoml.model.get('...')
        self.model = load_model(model_info.path)

   def _run_batch(self, df: pd.DataFrame):
        # run inference on a batch input
        ...

runner = CustomRunner('my_runner')
svc = bentoml.Service('my_service', runners=[runner])

...