test_sklearn_model_export.py

from __future__ import print_function

import sys
import os
import pickle
import pytest
import yaml
import json
from collections import namedtuple

import numpy as np
import pandas as pd
import pandas.testing
import sklearn
import sklearn.datasets as datasets
import sklearn.linear_model as glm
import sklearn.neighbors as knn
from sklearn.pipeline import Pipeline as SKPipeline
from sklearn.preprocessing import FunctionTransformer as SKFunctionTransformer

import mlflow.sklearn
import mlflow.utils
import mlflow.pyfunc.scoring_server as pyfunc_scoring_server
from mlflow import pyfunc
from mlflow.exceptions import MlflowException
from mlflow.protos.databricks_pb2 import INVALID_PARAMETER_VALUE
from mlflow.models import Model
from mlflow.store.s3_artifact_repo import S3ArtifactRepository
from mlflow.tracking.artifact_utils import _download_artifact_from_uri
from mlflow.utils.environment import _mlflow_conda_env
from mlflow.utils.file_utils import TempDir
from mlflow.utils.model_utils import _get_flavor_configuration

from tests.helper_functions import set_boto_credentials  # pylint: disable=unused-import
from tests.helper_functions import mock_s3_bucket  # pylint: disable=unused-import
from tests.helper_functions import score_model_in_sagemaker_docker_container

ModelWithData = namedtuple("ModelWithData", ["model", "inference_data"])


@pytest.fixture(scope="session")
def sklearn_knn_model():
    iris = datasets.load_iris()
    X = iris.data[:, :2]  # we only take the first two features.
    y = iris.target
    knn_model = knn.KNeighborsClassifier()
    knn_model.fit(X, y)
    return ModelWithData(model=knn_model, inference_data=X)


@pytest.fixture(scope="session")
def sklearn_logreg_model():
    iris = datasets.load_iris()
    X = iris.data[:, :2]  # we only take the first two features.
    y = iris.target
    linear_lr = glm.LogisticRegression()
    linear_lr.fit(X, y)
    return ModelWithData(model=linear_lr, inference_data=X)


@pytest.fixture(scope="session")
def sklearn_custom_transformer_model(sklearn_knn_model):
    def transform(vec):
        print("Invoking custom transformer!")
        return vec + 1

    transformer = SKFunctionTransformer(transform, validate=True)
    pipeline = SKPipeline([("custom_transformer", transformer), ("knn", sklearn_knn_model.model)])
    return ModelWithData(pipeline, inference_data=datasets.load_iris().data[:, :2])


@pytest.fixture
def model_path(tmpdir):
    return os.path.join(str(tmpdir), "model")


@pytest.fixture
def sklearn_custom_env(tmpdir):
    conda_env = os.path.join(str(tmpdir), "conda_env.yml")
    _mlflow_conda_env(
            conda_env,
            additional_conda_deps=["scikit-learn", "pytest"])
    return conda_env


@pytest.mark.large
def test_model_save_load(sklearn_knn_model, model_path):
    knn_model = sklearn_knn_model.model

    mlflow.sklearn.save_model(sk_model=knn_model, path=model_path)
    reloaded_knn_model = mlflow.sklearn.load_model(model_uri=model_path)
    reloaded_knn_pyfunc = pyfunc.load_pyfunc(model_uri=model_path)

    np.testing.assert_array_equal(
            knn_model.predict(sklearn_knn_model.inference_data),
            reloaded_knn_model.predict(sklearn_knn_model.inference_data))

    np.testing.assert_array_equal(
            reloaded_knn_model.predict(sklearn_knn_model.inference_data),
            reloaded_knn_pyfunc.predict(sklearn_knn_model.inference_data))


@pytest.mark.large
def test_model_load_from_remote_uri_succeeds(sklearn_knn_model, model_path, mock_s3_bucket):
    mlflow.sklearn.save_model(sk_model=sklearn_knn_model.model, path=model_path)

    artifact_root = "s3://{bucket_name}".format(bucket_name=mock_s3_bucket)
    artifact_path = "model"
    artifact_repo = S3ArtifactRepository(artifact_root)
    artifact_repo.log_artifacts(model_path, artifact_path=artifact_path)

    model_uri = artifact_root + "/" + artifact_path
    reloaded_knn_model = mlflow.sklearn.load_model(model_uri=model_uri)
    np.testing.assert_array_equal(
            sklearn_knn_model.model.predict(sklearn_knn_model.inference_data),
            reloaded_knn_model.predict(sklearn_knn_model.inference_data))


@pytest.mark.large
def test_model_log(sklearn_logreg_model, model_path):
    old_uri = mlflow.get_tracking_uri()
    with TempDir(chdr=True, remove_on_exit=True) as tmp:
        for should_start_run in [False, True]:
            try:
                mlflow.set_tracking_uri("test")
                if should_start_run:
                    mlflow.start_run()

                artifact_path = "linear"
                conda_env = os.path.join(tmp.path(), "conda_env.yaml")
                _mlflow_conda_env(conda_env, additional_pip_deps=["scikit-learn"])

                mlflow.sklearn.log_model(
                        sk_model=sklearn_logreg_model.model,
                        artifact_path=artifact_path,
                        conda_env=conda_env)
                model_uri = "runs:/{run_id}/{artifact_path}".format(
                    run_id=mlflow.active_run().info.run_id,
                    artifact_path=artifact_path)

                reloaded_logreg_model = mlflow.sklearn.load_model(model_uri=model_uri)
                np.testing.assert_array_equal(
                        sklearn_logreg_model.model.predict(sklearn_logreg_model.inference_data),
                        reloaded_logreg_model.predict(sklearn_logreg_model.inference_data))

                model_path = _download_artifact_from_uri(artifact_uri=model_uri)
                model_config = Model.load(os.path.join(model_path, "MLmodel"))
                assert pyfunc.FLAVOR_NAME in model_config.flavors
                assert pyfunc.ENV in model_config.flavors[pyfunc.FLAVOR_NAME]
                env_path = model_config.flavors[pyfunc.FLAVOR_NAME][pyfunc.ENV]
                assert os.path.exists(os.path.join(model_path, env_path))

            finally:
                mlflow.end_run()
                mlflow.set_tracking_uri(old_uri)


@pytest.mark.large
def test_custom_transformer_can_be_saved_and_loaded_with_cloudpickle_format(
        sklearn_custom_transformer_model, tmpdir):
    custom_transformer_model = sklearn_custom_transformer_model.model

    # Because the model contains a customer transformer that is not defined at the top level of the
    # current test module, we expect pickle to fail when attempting to serialize it. In contrast,
    # we expect cloudpickle to successfully locate the transformer definition and serialize the
    # model successfully.
    if sys.version_info >= (3, 0):
        expect_exception_context = pytest.raises(AttributeError)
    else:
        expect_exception_context = pytest.raises(pickle.PicklingError)
    with expect_exception_context:
        pickle_format_model_path = os.path.join(str(tmpdir), "pickle_model")
        mlflow.sklearn.save_model(sk_model=custom_transformer_model,
                                  path=pickle_format_model_path,
                                  serialization_format=mlflow.sklearn.SERIALIZATION_FORMAT_PICKLE)

    cloudpickle_format_model_path = os.path.join(str(tmpdir), "cloud_pickle_model")
    mlflow.sklearn.save_model(sk_model=custom_transformer_model,
                              path=cloudpickle_format_model_path,
                              serialization_format=mlflow.sklearn.SERIALIZATION_FORMAT_CLOUDPICKLE)

    reloaded_custom_transformer_model = mlflow.sklearn.load_model(
        model_uri=cloudpickle_format_model_path)

    np.testing.assert_array_equal(
            custom_transformer_model.predict(sklearn_custom_transformer_model.inference_data),
            reloaded_custom_transformer_model.predict(
                sklearn_custom_transformer_model.inference_data))


@pytest.mark.large
def test_model_save_persists_specified_conda_env_in_mlflow_model_directory(
        sklearn_knn_model, model_path, sklearn_custom_env):
    mlflow.sklearn.save_model(
            sk_model=sklearn_knn_model.model, path=model_path, conda_env=sklearn_custom_env)

    pyfunc_conf = _get_flavor_configuration(model_path=model_path, flavor_name=pyfunc.FLAVOR_NAME)
    saved_conda_env_path = os.path.join(model_path, pyfunc_conf[pyfunc.ENV])
    assert os.path.exists(saved_conda_env_path)
    assert saved_conda_env_path != sklearn_custom_env

    with open(sklearn_custom_env, "r") as f:
        sklearn_custom_env_parsed = yaml.safe_load(f)
    with open(saved_conda_env_path, "r") as f:
        saved_conda_env_parsed = yaml.safe_load(f)
    assert saved_conda_env_parsed == sklearn_custom_env_parsed


@pytest.mark.large
def test_model_save_accepts_conda_env_as_dict(sklearn_knn_model, model_path):
    conda_env = dict(mlflow.sklearn.get_default_conda_env())
    conda_env["dependencies"].append("pytest")
    mlflow.sklearn.save_model(
            sk_model=sklearn_knn_model.model, path=model_path, conda_env=conda_env)

    pyfunc_conf = _get_flavor_configuration(model_path=model_path, flavor_name=pyfunc.FLAVOR_NAME)
    saved_conda_env_path = os.path.join(model_path, pyfunc_conf[pyfunc.ENV])
    assert os.path.exists(saved_conda_env_path)

    with open(saved_conda_env_path, "r") as f:
        saved_conda_env_parsed = yaml.safe_load(f)
    assert saved_conda_env_parsed == conda_env


@pytest.mark.large
def test_model_log_persists_specified_conda_env_in_mlflow_model_directory(
        sklearn_knn_model, sklearn_custom_env):
    artifact_path = "model"
    with mlflow.start_run():
        mlflow.sklearn.log_model(sk_model=sklearn_knn_model.model,
                                 artifact_path=artifact_path,
                                 conda_env=sklearn_custom_env)
        model_uri = "runs:/{run_id}/{artifact_path}".format(
            run_id=mlflow.active_run().info.run_id,
            artifact_path=artifact_path)

    model_path = _download_artifact_from_uri(artifact_uri=model_uri)
    pyfunc_conf = _get_flavor_configuration(model_path=model_path, flavor_name=pyfunc.FLAVOR_NAME)
    saved_conda_env_path = os.path.join(model_path, pyfunc_conf[pyfunc.ENV])
    assert os.path.exists(saved_conda_env_path)
    assert saved_conda_env_path != sklearn_custom_env

    with open(sklearn_custom_env, "r") as f:
        sklearn_custom_env_parsed = yaml.safe_load(f)
    with open(saved_conda_env_path, "r") as f:
        saved_conda_env_parsed = yaml.safe_load(f)
    assert saved_conda_env_parsed == sklearn_custom_env_parsed


@pytest.mark.large
def test_model_save_throws_exception_if_serialization_format_is_unrecognized(
        sklearn_knn_model, model_path):
    with pytest.raises(MlflowException) as exc:
        mlflow.sklearn.save_model(sk_model=sklearn_knn_model.model, path=model_path,
                                  serialization_format="not a valid format")
        assert exc.error_code == INVALID_PARAMETER_VALUE

    # The unsupported serialization format should have been detected prior to the execution of
    # any directory creation or state-mutating persistence logic that would prevent a second
    # serialization call with the same model path from succeeding
    assert not os.path.exists(model_path)
    mlflow.sklearn.save_model(sk_model=sklearn_knn_model.model, path=model_path)


@pytest.mark.large
def test_model_save_without_specified_conda_env_uses_default_env_with_expected_dependencies(
        sklearn_knn_model, model_path):
    knn_model = sklearn_knn_model.model
    mlflow.sklearn.save_model(sk_model=knn_model, path=model_path, conda_env=None,
                              serialization_format=mlflow.sklearn.SERIALIZATION_FORMAT_PICKLE)

    pyfunc_conf = _get_flavor_configuration(model_path=model_path, flavor_name=pyfunc.FLAVOR_NAME)
    conda_env_path = os.path.join(model_path, pyfunc_conf[pyfunc.ENV])
    with open(conda_env_path, "r") as f:
        conda_env = yaml.safe_load(f)

    assert conda_env == mlflow.sklearn.get_default_conda_env()


@pytest.mark.large
def test_model_log_without_specified_conda_env_uses_default_env_with_expected_dependencies(
        sklearn_knn_model):
    artifact_path = "model"
    knn_model = sklearn_knn_model.model
    with mlflow.start_run():
        mlflow.sklearn.log_model(sk_model=knn_model, artifact_path=artifact_path, conda_env=None,
                                 serialization_format=mlflow.sklearn.SERIALIZATION_FORMAT_PICKLE)
        model_uri = "runs:/{run_id}/{artifact_path}".format(
            run_id=mlflow.active_run().info.run_id,
            artifact_path=artifact_path)

    model_path = _download_artifact_from_uri(artifact_uri=model_uri)
    pyfunc_conf = _get_flavor_configuration(model_path=model_path, flavor_name=pyfunc.FLAVOR_NAME)
    conda_env_path = os.path.join(model_path, pyfunc_conf[pyfunc.ENV])
    with open(conda_env_path, "r") as f:
        conda_env = yaml.safe_load(f)

    assert conda_env == mlflow.sklearn.get_default_conda_env()


@pytest.mark.large
def test_model_save_uses_cloudpickle_serialization_format_by_default(sklearn_knn_model, model_path):
    mlflow.sklearn.save_model(sk_model=sklearn_knn_model.model, path=model_path, conda_env=None)

    sklearn_conf = _get_flavor_configuration(
            model_path=model_path, flavor_name=mlflow.sklearn.FLAVOR_NAME)
    assert "serialization_format" in sklearn_conf
    assert sklearn_conf["serialization_format"] == mlflow.sklearn.SERIALIZATION_FORMAT_CLOUDPICKLE


@pytest.mark.large
def test_model_log_uses_cloudpickle_serialization_format_by_default(sklearn_knn_model):
    artifact_path = "model"
    with mlflow.start_run():
        mlflow.sklearn.log_model(
                sk_model=sklearn_knn_model.model, artifact_path=artifact_path, conda_env=None)
        model_uri = "runs:/{run_id}/{artifact_path}".format(
            run_id=mlflow.active_run().info.run_id,
            artifact_path=artifact_path)

    model_path = _download_artifact_from_uri(artifact_uri=model_uri)
    sklearn_conf = _get_flavor_configuration(
            model_path=model_path, flavor_name=mlflow.sklearn.FLAVOR_NAME)
    assert "serialization_format" in sklearn_conf
    assert sklearn_conf["serialization_format"] == mlflow.sklearn.SERIALIZATION_FORMAT_CLOUDPICKLE


@pytest.mark.large
def test_model_save_with_cloudpickle_format_adds_cloudpickle_to_conda_environment(
        sklearn_knn_model, model_path):
    mlflow.sklearn.save_model(
            sk_model=sklearn_knn_model.model,
            path=model_path,
            conda_env=None,
            serialization_format=mlflow.sklearn.SERIALIZATION_FORMAT_CLOUDPICKLE)

    sklearn_conf = _get_flavor_configuration(
            model_path=model_path, flavor_name=mlflow.sklearn.FLAVOR_NAME)
    assert "serialization_format" in sklearn_conf
    assert sklearn_conf["serialization_format"] == mlflow.sklearn.SERIALIZATION_FORMAT_CLOUDPICKLE

    pyfunc_conf = _get_flavor_configuration(model_path=model_path, flavor_name=pyfunc.FLAVOR_NAME)
    saved_conda_env_path = os.path.join(model_path, pyfunc_conf[pyfunc.ENV])
    assert os.path.exists(saved_conda_env_path)
    with open(saved_conda_env_path, "r") as f:
        saved_conda_env_parsed = yaml.safe_load(f)

    pip_deps = [dependency for dependency in saved_conda_env_parsed["dependencies"]
                if type(dependency) == dict and "pip" in dependency]
    assert len(pip_deps) == 1
    assert any(["cloudpickle" in pip_dep for pip_dep in pip_deps[0]["pip"]])


@pytest.mark.large
def test_model_save_without_cloudpickle_format_does_not_add_cloudpickle_to_conda_environment(
        sklearn_knn_model, model_path):
    non_cloudpickle_serialization_formats = list(mlflow.sklearn.SUPPORTED_SERIALIZATION_FORMATS)
    non_cloudpickle_serialization_formats.remove(mlflow.sklearn.SERIALIZATION_FORMAT_CLOUDPICKLE)

    for serialization_format in non_cloudpickle_serialization_formats:
        mlflow.sklearn.save_model(
                sk_model=sklearn_knn_model.model,
                path=model_path,
                conda_env=None,
                serialization_format=serialization_format)

        sklearn_conf = _get_flavor_configuration(
                model_path=model_path, flavor_name=mlflow.sklearn.FLAVOR_NAME)
        assert "serialization_format" in sklearn_conf
        assert sklearn_conf["serialization_format"] == serialization_format

        pyfunc_conf = _get_flavor_configuration(
                model_path=model_path, flavor_name=pyfunc.FLAVOR_NAME)
        saved_conda_env_path = os.path.join(model_path, pyfunc_conf[pyfunc.ENV])
        assert os.path.exists(saved_conda_env_path)
        with open(saved_conda_env_path, "r") as f:
            saved_conda_env_parsed = yaml.safe_load(f)
        assert all(["cloudpickle" not in dependency
                    for dependency in saved_conda_env_parsed["dependencies"]])


@pytest.mark.release
def test_sagemaker_docker_model_scoring_with_default_conda_env(sklearn_knn_model, model_path):
    mlflow.sklearn.save_model(sk_model=sklearn_knn_model.model, path=model_path, conda_env=None)
    reloaded_knn_pyfunc = pyfunc.load_pyfunc(model_uri=model_path)

    inference_df = pd.DataFrame(sklearn_knn_model.inference_data)
    scoring_response = score_model_in_sagemaker_docker_container(
            model_uri=model_path,
            data=inference_df,
            content_type=pyfunc_scoring_server.CONTENT_TYPE_JSON_SPLIT_ORIENTED,
            flavor=mlflow.pyfunc.FLAVOR_NAME)
    deployed_model_preds = pd.DataFrame(json.loads(scoring_response.content))

    pandas.testing.assert_frame_equal(
        deployed_model_preds,
        pd.DataFrame(reloaded_knn_pyfunc.predict(inference_df)),
        check_dtype=False,
        check_less_precise=6)