from contextlib import contextmanager
import os
import tempfile
import yaml
import warnings
import numpy as np
import mlflow
import types
import mlflow.utils.autologging_utils
from mlflow import pyfunc
from mlflow.exceptions import MlflowException
from mlflow.utils.annotations import experimental
from mlflow.utils.uri import append_to_uri_path
from mlflow.models import Model
from mlflow.tracking.artifact_utils import _download_artifact_from_uri
from mlflow.models.model import MLMODEL_FILE_NAME
from mlflow.models.signature import ModelSignature
from mlflow.models.utils import ModelInputExample, _save_example
from mlflow.utils.environment import _mlflow_conda_env, _log_pip_requirements
from mlflow.utils.model_utils import _get_flavor_configuration
from mlflow.protos.databricks_pb2 import RESOURCE_ALREADY_EXISTS
from mlflow.tracking._model_registry import DEFAULT_AWAIT_MAX_SLEEP_SECONDS
FLAVOR_NAME = "shap"
_MAXIMUM_BACKGROUND_DATA_SIZE = 100
_DEFAULT_ARTIFACT_PATH = "model_explanations_shap"
_SUMMARY_BAR_PLOT_FILE_NAME = "summary_bar_plot.png"
_BASE_VALUES_FILE_NAME = "base_values.npy"
_SHAP_VALUES_FILE_NAME = "shap_values.npy"
_UNKNOWN_MODEL_FLAVOR = "unknown"
_UNDERLYING_MODEL_SUBPATH = "underlying_model"
[docs]def get_underlying_model_flavor(model):
"""
Find the underlying models flavor.
:param model: underlying model of the explainer.
"""
# checking if underlying model is wrapped
if hasattr(model, "inner_model"):
unwrapped_model = model.inner_model
# check if passed model is a method of object
if isinstance(unwrapped_model, types.MethodType):
model_object = unwrapped_model.__self__
# check if model object is of type sklearn
try:
import sklearn
if issubclass(type(model_object), sklearn.base.BaseEstimator):
return mlflow.sklearn.FLAVOR_NAME
except ImportError:
pass
# check if passed model is of type pytorch
try:
import torch
if issubclass(type(unwrapped_model), torch.nn.Module):
return mlflow.pytorch.FLAVOR_NAME
except ImportError:
pass
return _UNKNOWN_MODEL_FLAVOR
[docs]def get_default_conda_env():
"""
:return: The default Conda environment for
MLflow Models produced by calls to
:func:`save_explainer()` and :func:`log_explainer()`.
"""
import shap
pip_deps = ["shap=={}".format(shap.__version__)]
return _mlflow_conda_env(additional_pip_deps=pip_deps)
def _load_pyfunc(path):
"""
Load PyFunc implementation. Called by ``pyfunc.load_pyfunc``.
"""
return _SHAPWrapper(path)
@contextmanager
def _log_artifact_contextmanager(out_file, artifact_path=None):
"""
A context manager to make it easier to log an artifact.
"""
with tempfile.TemporaryDirectory() as tmp_dir:
tmp_path = os.path.join(tmp_dir, out_file)
yield tmp_path
mlflow.log_artifact(tmp_path, artifact_path)
def _log_numpy(numpy_obj, out_file, artifact_path=None):
"""
Log a numpy object.
"""
with _log_artifact_contextmanager(out_file, artifact_path) as tmp_path:
np.save(tmp_path, numpy_obj)
def _log_matplotlib_figure(fig, out_file, artifact_path=None):
"""
Log a matplotlib figure.
"""
with _log_artifact_contextmanager(out_file, artifact_path) as tmp_path:
fig.savefig(tmp_path)
[docs]@experimental
def log_explanation(predict_function, features, artifact_path=None):
r"""
Given a ``predict_function`` capable of computing ML model output on the provided ``features``,
computes and logs explanations of an ML model's output. Explanations are logged as a directory
of artifacts containing the following items generated by `SHAP`_ (SHapley Additive
exPlanations).
- Base values
- SHAP values (computed using `shap.KernelExplainer`_)
- Summary bar plot (shows the average impact of each feature on model output)
:param predict_function:
A function to compute the output of a model (e.g. ``predict_proba`` method of
scikit-learn classifiers). Must have the following signature:
.. code-block:: python
def predict_function(X) -> pred:
...
- ``X``: An array-like object whose shape should be (# samples, # features).
- ``pred``: An array-like object whose shape should be (# samples) for
a regressor or (# classes, # samples) for a classifier. For a classifier,
the values in ``pred`` should correspond to the predicted probability of each class.
Acceptable array-like object types:
- ``numpy.array``
- ``pandas.DataFrame``
- ``shap.common.DenseData``
- ``scipy.sparse matrix``
:param features:
A matrix of features to compute SHAP values with. The provided features should
have shape (# samples, # features), and can be either of the array-like object
types listed above.
.. note::
Background data for `shap.KernelExplainer`_ is generated by subsampling ``features``
with `shap.kmeans`_. The background data size is limited to 100 rows for performance
reasons.
:param artifact_path:
The run-relative artifact path to which the explanation is saved.
If unspecified, defaults to "model_explanations_shap".
:return: Artifact URI of the logged explanations.
.. _SHAP: https://github.com/slundberg/shap
.. _shap.KernelExplainer: https://shap.readthedocs.io/en/latest/generated
/shap.KernelExplainer.html#shap.KernelExplainer
.. _shap.kmeans: https://github.com/slundberg/shap/blob/v0.36.0/shap/utils/_legacy.py#L9
.. code-block:: python
:caption: Example
import os
import numpy as np
import pandas as pd
from sklearn.datasets import load_boston
from sklearn.linear_model import LinearRegression
import mlflow
# prepare training data
dataset = load_boston()
X = pd.DataFrame(dataset.data[:50, :8], columns=dataset.feature_names[:8])
y = dataset.target[:50]
# train a model
model = LinearRegression()
model.fit(X, y)
# log an explanation
with mlflow.start_run() as run:
mlflow.shap.log_explanation(model.predict, X)
# list artifacts
client = mlflow.tracking.MlflowClient()
artifact_path = "model_explanations_shap"
artifacts = [x.path for x in client.list_artifacts(run.info.run_id, artifact_path)]
print("# artifacts:")
print(artifacts)
# load back the logged explanation
dst_path = client.download_artifacts(run.info.run_id, artifact_path)
base_values = np.load(os.path.join(dst_path, "base_values.npy"))
shap_values = np.load(os.path.join(dst_path, "shap_values.npy"))
print("\n# base_values:")
print(base_values)
print("\n# shap_values:")
print(shap_values[:3])
.. code-block:: text
:caption: Output
# artifacts:
['model_explanations_shap/base_values.npy',
'model_explanations_shap/shap_values.npy',
'model_explanations_shap/summary_bar_plot.png']
# base_values:
20.502000000000002
# shap_values:
[[ 2.09975523 0.4746513 7.63759026 0. ]
[ 2.00883109 -0.18816665 -0.14419184 0. ]
[ 2.00891772 -0.18816665 -0.14419184 0. ]]
.. figure:: ../_static/images/shap-ui-screenshot.png
Logged artifacts
"""
import matplotlib.pyplot as plt
import shap
artifact_path = _DEFAULT_ARTIFACT_PATH if artifact_path is None else artifact_path
with mlflow.utils.autologging_utils.disable_autologging():
background_data = shap.kmeans(features, min(_MAXIMUM_BACKGROUND_DATA_SIZE, len(features)))
explainer = shap.KernelExplainer(predict_function, background_data)
shap_values = explainer.shap_values(features)
_log_numpy(explainer.expected_value, _BASE_VALUES_FILE_NAME, artifact_path)
_log_numpy(shap_values, _SHAP_VALUES_FILE_NAME, artifact_path)
shap.summary_plot(shap_values, features, plot_type="bar", show=False)
fig = plt.gcf()
fig.tight_layout()
_log_matplotlib_figure(fig, _SUMMARY_BAR_PLOT_FILE_NAME, artifact_path)
plt.close(fig)
return append_to_uri_path(mlflow.active_run().info.artifact_uri, artifact_path)
[docs]@experimental
def log_explainer(
explainer,
artifact_path,
serialize_model_using_mlflow=True,
conda_env=None,
registered_model_name=None,
signature: ModelSignature = None,
input_example: ModelInputExample = None,
await_registration_for=DEFAULT_AWAIT_MAX_SLEEP_SECONDS,
):
"""
Log an SHAP explainer as an MLflow artifact for the current run.
:param explainer: SHAP explainer to be saved.
:param artifact_path: Run-relative artifact path.
:param serialize_model_using_mlflow: When set to True, MLflow will extract the underlying
model and serialize it as an MLmodel, otherwise it
uses SHAP's internal serialization. Defaults to True.
Currently MLflow serialization is only supported for
models of 'sklearn' or 'pytorch' flavors.
:param conda_env: Either a dictionary representation of a Conda environment or the path to a
Conda environment yaml file. If provided, this decsribes the environment
this model should be run in. At minimum, it should specify the dependencies
contained in :func:`get_default_conda_env()`. If `None`, the default
:func:`get_default_conda_env()` environment is added to the model.
The following is an *example* dictionary representation of a Conda
environment::
{
'name': 'mlflow-env',
'channels': ['defaults'],
'dependencies': [
'python=3.6.0',
'shap=0.37.0'
]
}
:param registered_model_name: (Experimental) If given, create a model version under
``registered_model_name``, also creating a registered model if one
with the given name does not exist.
:param signature: (Experimental) :py:class:`ModelSignature <mlflow.models.ModelSignature>`
describes model input and output :py:class:`Schema <mlflow.types.Schema>`.
The model signature can be :py:func:`inferred <mlflow.models.infer_signature>`
from datasets with valid model input (e.g. the training dataset with target
column omitted) and valid model output (e.g. model predictions generated on
the training dataset), for example:
.. code-block:: python
from mlflow.models.signature import infer_signature
train = df.drop_column("target_label")
predictions = ... # compute model predictions
signature = infer_signature(train, predictions)
:param input_example: (Experimental) Input example provides one or several instances of valid
model input. The example can be used as a hint of what data to feed the
model. The given example will be converted to a Pandas DataFrame and then
serialized to json using the Pandas split-oriented format. Bytes are
base64-encoded.
:param await_registration_for: Number of seconds to wait for the model version to finish
being created and is in ``READY`` status. By default, the function
waits for five minutes. Specify 0 or None to skip waiting.
"""
Model.log(
artifact_path=artifact_path,
flavor=mlflow.shap,
explainer=explainer,
conda_env=conda_env,
serialize_model_using_mlflow=serialize_model_using_mlflow,
registered_model_name=registered_model_name,
signature=signature,
input_example=input_example,
await_registration_for=await_registration_for,
)
[docs]@experimental
def save_explainer(
explainer,
path,
serialize_model_using_mlflow=True,
conda_env=None,
mlflow_model=None,
signature: ModelSignature = None,
input_example: ModelInputExample = None,
):
"""
Save a SHAP explainer to a path on the local file system. Produces an MLflow Model
containing the following flavors:
- :py:mod:`mlflow.shap`
- :py:mod:`mlflow.pyfunc`
:param explainer: SHAP explainer to be saved.
:param path: Local path where the explainer is to be saved.
:param serialize_model_using_mlflow: When set to True, MLflow will extract the underlying
model and serialize it as an MLmodel, otherwise it
uses SHAP's internal serialization. Defaults to True.
Currently MLflow serialization is only supported for
models of 'sklearn' or 'pytorch' flavors.
:param conda_env: Either a dictionary representation of a Conda environment or the path to a
Conda environment yaml file. If provided, this decsribes the environment
this model should be run in. At minimum, it should specify the dependencies
contained in :func:`get_default_conda_env()`. If `None`, the default
:func:`get_default_conda_env()` environment is added to the model.
The following is an *example* dictionary representation of a Conda
environment::
{
'name': 'mlflow-env',
'channels': ['defaults'],
'dependencies': [
'python=3.6.0',
'shap=0.37.0'
]
}
:param mlflow_model: :py:mod:`mlflow.models.Model` this flavor is being added to.
:param signature: (Experimental) :py:class:`ModelSignature <mlflow.models.ModelSignature>`
describes model input and output :py:class:`Schema <mlflow.types.Schema>`.
The model signature can be :py:func:`inferred <mlflow.models.infer_signature>`
from datasets with valid model input (e.g. the training dataset with target
column omitted) and valid model output (e.g. model predictions generated on
the training dataset), for example:
.. code-block:: python
from mlflow.models.signature import infer_signature
train = df.drop_column("target_label")
predictions = ... # compute model predictions
signature = infer_signature(train, predictions)
:param input_example: (Experimental) Input example provides one or several instances of valid
model input. The example can be used as a hint of what data to feed the
model. The given example will be converted to a Pandas DataFrame and then
serialized to json using the Pandas split-oriented format. Bytes are
base64-encoded.
"""
import shap
if os.path.exists(path):
raise MlflowException(
message="Path '{}' already exists".format(path), error_code=RESOURCE_ALREADY_EXISTS,
)
os.makedirs(path)
if mlflow_model is None:
mlflow_model = Model()
if signature is not None:
mlflow_model.signature = signature
if input_example is not None:
_save_example(mlflow_model, input_example, path)
underlying_model_flavor = None
underlying_model_path = None
serializable_by_mlflow = False
# saving the underlying model if required
if serialize_model_using_mlflow:
underlying_model_flavor = get_underlying_model_flavor(explainer.model)
if underlying_model_flavor != _UNKNOWN_MODEL_FLAVOR:
serializable_by_mlflow = True # prevents SHAP from serializing the underlying model
underlying_model_path = os.path.join(path, _UNDERLYING_MODEL_SUBPATH)
else:
warnings.warn(
"Unable to serialize underlying model using MLflow, will use SHAP serialization"
)
if underlying_model_flavor == mlflow.sklearn.FLAVOR_NAME:
mlflow.sklearn.save_model(explainer.model.inner_model.__self__, underlying_model_path)
elif underlying_model_flavor == mlflow.pytorch.FLAVOR_NAME:
mlflow.pytorch.save_model(explainer.model.inner_model, underlying_model_path)
# saving the explainer object
explainer_data_subpath = "explainer.shap"
explainer_output_path = os.path.join(path, explainer_data_subpath)
with open(explainer_output_path, "wb") as explainer_output_file_handle:
if serialize_model_using_mlflow and serializable_by_mlflow:
explainer.save(explainer_output_file_handle, model_saver=False)
else:
explainer.save(explainer_output_file_handle)
conda_env_subpath = "conda.yaml"
if conda_env is None:
conda_env = get_default_conda_env()
elif not isinstance(conda_env, dict):
with open(conda_env, "r") as f:
conda_env = yaml.safe_load(f)
# merging the conda environment generated by serializing the underlying model
if underlying_model_path is not None:
underlying_model_conda_path = os.path.join(underlying_model_path, "conda.yaml")
with open(underlying_model_conda_path, "r") as underlying_model_conda_file:
underlying_model_conda_env = yaml.safe_load(underlying_model_conda_file)
conda_env = _merge_environments(conda_env, underlying_model_conda_env)
with open(os.path.join(path, conda_env_subpath), "w") as f:
yaml.safe_dump(conda_env, stream=f, default_flow_style=False)
_log_pip_requirements(conda_env, path)
pyfunc.add_to_model(
mlflow_model,
loader_module="mlflow.shap",
model_path=explainer_data_subpath,
underlying_model_flavor=underlying_model_flavor,
env=conda_env_subpath,
)
mlflow_model.add_flavor(
FLAVOR_NAME,
shap_version=shap.__version__,
serialized_explainer=explainer_data_subpath,
underlying_model_flavor=underlying_model_flavor,
)
mlflow_model.save(os.path.join(path, MLMODEL_FILE_NAME))
# Defining save_model (Required by Model.log) to refer to save_explainer
save_model = save_explainer
def _get_conda_and_pip_dependencies(conda_env):
"""
Extract conda and pip dependencies from conda environments
:param conda_env: Conda environment
"""
conda_deps = []
pip_deps = []
for dependency in conda_env["dependencies"]:
if isinstance(dependency, dict) and dependency["pip"]:
for pip_dependency in dependency["pip"]:
if pip_dependency != "mlflow":
pip_deps.append(pip_dependency)
else:
if dependency.split("=")[0] != "python" and dependency.split("=")[0] != "pip":
conda_deps.append(dependency)
return conda_deps, pip_deps
def _merge_environments(shap_environment, model_environment):
"""
Merge conda environments of underlying model and shap.
:param shap_environment: SHAP conda environment.
:param model_environment: Underlying model conda environment.
"""
# merge the channels from the two environments and remove the default conda
# channels if present since its added later in `_mlflow_conda_env`
merged_conda_channels = list(
set(shap_environment["channels"] + model_environment["channels"]) - set(["conda-forge"])
)
shap_conda_deps, shap_pip_deps = _get_conda_and_pip_dependencies(shap_environment)
model_conda_deps, model_pip_deps = _get_conda_and_pip_dependencies(model_environment)
merged_conda_deps = list(set(shap_conda_deps + model_conda_deps))
merged_pip_deps = list(set(shap_pip_deps + model_pip_deps))
return _mlflow_conda_env(
additional_conda_deps=merged_conda_deps,
additional_pip_deps=merged_pip_deps,
additional_conda_channels=merged_conda_channels,
)
[docs]@experimental
def load_explainer(model_uri):
"""
Load a SHAP explainer from a local file or a run.
:param model_uri: The location, in URI format, of the MLflow model, for example:
- ``/Users/me/path/to/local/model``
- ``relative/path/to/local/model``
- ``s3://my_bucket/path/to/model``
- ``runs:/<mlflow_run_id>/run-relative/path/to/model``
- ``models:/<model_name>/<model_version>``
- ``models:/<model_name>/<stage>``
For more information about supported URI schemes, see
`Referencing Artifacts <https://www.mlflow.org/docs/latest/concepts.html#
artifact-locations>`_.
:return: A SHAP explainer.
"""
explainer_path = _download_artifact_from_uri(artifact_uri=model_uri)
flavor_conf = _get_flavor_configuration(model_path=explainer_path, flavor_name=FLAVOR_NAME)
explainer_artifacts_path = os.path.join(explainer_path, flavor_conf["serialized_explainer"])
underlying_model_flavor = flavor_conf["underlying_model_flavor"]
model = None
if underlying_model_flavor != _UNKNOWN_MODEL_FLAVOR:
underlying_model_path = os.path.join(explainer_path, _UNDERLYING_MODEL_SUBPATH)
if underlying_model_flavor == mlflow.sklearn.FLAVOR_NAME:
model = mlflow.sklearn._load_pyfunc(underlying_model_path).predict
elif underlying_model_flavor == mlflow.pytorch.FLAVOR_NAME:
model = mlflow.pytorch._load_model(os.path.join(underlying_model_path, "data"))
return _load_explainer(explainer_file=explainer_artifacts_path, model=model)
@experimental
def _load_explainer(explainer_file, model=None):
"""
Load a SHAP explainer saved as an MLflow artifact on the local file system.
:param explainer_file: Local filesystem path to the MLflow Model saved with the ``shap`` flavor
:param model: model to override underlying explainer model.
"""
import shap
def inject_model_loader(_in_file):
return model
with open(explainer_file, "rb") as explainer:
if model is None:
explainer = shap.Explainer.load(explainer)
else:
explainer = shap.Explainer.load(explainer, model_loader=inject_model_loader)
return explainer
class _SHAPWrapper:
def __init__(self, path):
flavor_conf = _get_flavor_configuration(model_path=path, flavor_name=FLAVOR_NAME)
shap_explainer_artifacts_path = os.path.join(path, flavor_conf["serialized_explainer"])
underlying_model_flavor = flavor_conf["underlying_model_flavor"]
model = None
if underlying_model_flavor != _UNKNOWN_MODEL_FLAVOR:
underlying_model_path = os.path.join(path, _UNDERLYING_MODEL_SUBPATH)
if underlying_model_flavor == mlflow.sklearn.FLAVOR_NAME:
model = mlflow.sklearn._load_pyfunc(underlying_model_path).predict
elif underlying_model_flavor == mlflow.pytorch.FLAVOR_NAME:
model = mlflow.pytorch._load_model(os.path.join(underlying_model_path, "data"))
self.explainer = _load_explainer(explainer_file=shap_explainer_artifacts_path, model=model)
def predict(self, dataframe):
return self.explainer(dataframe.values).values