sklearn.py

import joblib

from typing import List
from sklearn.pipeline import Pipeline

from mlserver.codecs import NumpyCodec, NumpyRequestCodec, PandasCodec
from mlserver.errors import InferenceError
from mlserver.model import MLModel
from mlserver.types import (
    InferenceRequest,
    InferenceResponse,
    ResponseOutput,
    RequestOutput,
)
from mlserver.utils import get_model_uri

PREDICT_FN_KEY = "predict_fn"
PREDICT_OUTPUT = "predict"
PREDICT_PROBA_OUTPUT = "predict_proba"
PREDICT_TRANSFORM = "transform"
VALID_OUTPUTS = [PREDICT_OUTPUT, PREDICT_PROBA_OUTPUT, PREDICT_TRANSFORM]

WELLKNOWN_MODEL_FILENAMES = ["model.joblib", "model.pickle", "model.pkl"]


class SKLearnModel(MLModel):
    """
    Implementation of the MLModel interface to load and serve `scikit-learn`
    models persisted with `joblib`.
    """

    async def load(self) -> bool:
        # TODO: Log info message
        model_uri = await get_model_uri(
            self._settings, wellknown_filenames=WELLKNOWN_MODEL_FILENAMES
        )
        self._model = joblib.load(model_uri)

        return True

    async def predict(self, payload: InferenceRequest) -> InferenceResponse:
        payload = self._check_request(payload)

        outputs = self._get_model_outputs(payload)

        return InferenceResponse(
            model_name=self.name,
            model_version=self.version,
            outputs=outputs,
        )

    def _check_request(self, payload: InferenceRequest) -> InferenceRequest:
        if not payload.outputs:
            found_predict_fn = False
            if self.settings.parameters:
                if self.settings.parameters.extra:
                    if PREDICT_FN_KEY in self.settings.parameters.extra:
                        payload.outputs = [
                            RequestOutput(
                                name=self.settings.parameters.extra[PREDICT_FN_KEY]
                            )
                        ]
                        found_predict_fn = True
            # By default, only return the result of `predict()`
            if not found_predict_fn:
                payload.outputs = [RequestOutput(name=PREDICT_OUTPUT)]
        else:
            for request_output in payload.outputs:
                if request_output.name not in VALID_OUTPUTS:
                    raise InferenceError(
                        f"SKLearnModel only supports '{PREDICT_OUTPUT}', "
                        f"'{PREDICT_PROBA_OUTPUT}' and "
                        f"'{PREDICT_TRANSFORM}' as outputs "
                        f"({request_output.name} was received)"
                    )

        # Regression models do not support `predict_proba`
        output_names = [o.name for o in payload.outputs]  # type: ignore
        if PREDICT_PROBA_OUTPUT in output_names:
            # Ensure model supports it
            maybe_regressor = self._model
            if isinstance(self._model, Pipeline):
                maybe_regressor = maybe_regressor.steps[-1][-1]

            if not hasattr(maybe_regressor, PREDICT_PROBA_OUTPUT):
                raise InferenceError(
                    f"{type(maybe_regressor)} models do not support "
                    f"'{PREDICT_PROBA_OUTPUT}"
                )

        return payload

    def _get_model_outputs(self, payload: InferenceRequest) -> List[ResponseOutput]:
        decoded_request = self.decode_request(payload, default_codec=NumpyRequestCodec)

        outputs = []
        for request_output in payload.outputs:  # type: ignore
            predict_fn = getattr(self._model, request_output.name)
            y = predict_fn(decoded_request)

            # If output is a Pandas DataFrame, encode it manually
            if PandasCodec.can_encode(y):
                if len(payload.outputs) > 1:  # type: ignore
                    all_output_names = [o.name for o in payload.outputs]  # type: ignore
                    raise InferenceError(
                        f"{request_output.name} returned columnar data of type"
                        f" {type(y)} and {all_output_names} were"
                        f" requested. Cannot encode multiple columnar data responses"
                        f" into one response."
                    )

                # Use PandasCodec manually, to obtain the list of outputs
                return PandasCodec.encode_outputs(y)

            output = self.encode(y, request_output, default_codec=NumpyCodec)
            outputs.append(output)

        return outputs