Added specific spec tests. Added more extensions. Fix for preserve de…

…faults.
IFV · Sep 17, 2019 · e318e7b · e318e7b
1 parent 021ecad
commit e318e7b
Show file tree

Hide file tree

Showing 13 changed files with 183 additions and 69 deletions.
diff --git a/python/interpret-core/interpret/api/base.py b/python/interpret-core/interpret/api/base.py
@@ -13,7 +13,7 @@ class ExplainerMixin(ABC):
         available_explanations: A list of strings subsetting the following
             - "perf", "data", "local", "global".
         explainer_type: A string that is one of the following
-            - "blackbox", "model", "specific".
+            - "blackbox", "model", "specific", "data", "perf".
     """
 
     @property

diff --git a/python/interpret-core/interpret/blackbox/lime.py b/python/interpret-core/interpret/blackbox/lime.py
@@ -7,7 +7,6 @@
 from ..utils import gen_name_from_class, gen_local_selector
 from ..utils import perf_dict
 from ..utils import unify_data, unify_predict_fn
-from lime.lime_tabular import LimeTabularExplainer
 import warnings
 
 
@@ -27,6 +26,7 @@ def __init__(
         n_jobs=1,
         **kwargs
     ):
+        from lime.lime_tabular import LimeTabularExplainer
 
         self.data, _, self.feature_names, self.feature_types = unify_data(
             data, None, feature_names, feature_types

diff --git a/python/interpret-core/interpret/blackbox/shap.py b/python/interpret-core/interpret/blackbox/shap.py
@@ -1,14 +1,11 @@
 # Copyright (c) 2019 Microsoft Corporation
 # Distributed under the MIT software license
+from ..utils.shap import shap_explain_local
 
 from ..api.base import ExplainerMixin
-from ..api.templates import FeatureValueExplanation
 from ..utils import unify_predict_fn, unify_data
-from ..utils import perf_dict, gen_name_from_class, gen_local_selector
 import warnings
 
-import shap
-
 
 class ShapKernel(ExplainerMixin):
     available_explanations = ["local"]
@@ -26,6 +23,8 @@ def __init__(
         **kwargs
     ):
 
+        import shap
+
         self.data, _, self.feature_names, self.feature_types = unify_data(
             data, None, feature_names, feature_types
         )
@@ -41,63 +40,4 @@ def __init__(
         self.shap = shap.KernelExplainer(self.predict_fn, data, **self.kwargs)
 
     def explain_local(self, X, y=None, name=None):
-        if name is None:
-            name = gen_name_from_class(self)
-        X, y, _, _ = unify_data(X, y, self.feature_names, self.feature_types)
-
-        all_shap_values = self.shap.shap_values(X)
-        predictions = self.predict_fn(X)
-
-        data_dicts = []
-        scores_list = all_shap_values
-        perf_list = []
-        for i, instance in enumerate(X):
-            shap_values = all_shap_values[i]
-            perf_dict_obj = perf_dict(y, predictions, i)
-
-            perf_list.append(perf_dict_obj)
-
-            data_dict = {
-                "type": "univariate",
-                "names": self.feature_names,
-                "perf": perf_dict_obj,
-                "scores": shap_values,
-                "values": instance,
-                "extra": {
-                    "names": ["Base Value"],
-                    "scores": [self.shap.expected_value],
-                    "values": [1],
-                },
-            }
-            data_dicts.append(data_dict)
-
-        internal_obj = {
-            "overall": None,
-            "specific": data_dicts,
-            "mli": [
-                {
-                    "explanation_type": "local_feature_importance",
-                    "value": {
-                        "scores": scores_list,
-                        "intercept": self.shap.expected_value,
-                        "perf": perf_list,
-                    },
-                }
-            ],
-        }
-        internal_obj["mli"].append(
-            {
-                "explanation_type": "evaluation_dataset",
-                "value": {"dataset_x": X, "dataset_y": y},
-            }
-        )
-        selector = gen_local_selector(X, y, predictions)
-
-        return FeatureValueExplanation(
-            "local",
-            internal_obj,
-            feature_names=self.feature_names,
-            feature_types=self.feature_types,
-            name=name,
-            selector=selector,
-        )
+        return shap_explain_local(self, X, y=y, name=name)
diff --git a/python/interpret-core/interpret/ext/data/__init__.py b/python/interpret-core/interpret/ext/data/__init__.py
@@ -0,0 +1,10 @@
+# Copyright (c) 2019 Microsoft Corporation
+# Distributed under the MIT software license
+
+import sys
+from interpret.ext.extension_utils import load_class_extensions
+from interpret.ext.extension import DATA_EXTENSION_KEY, _is_valid_data_explainer
+
+load_class_extensions(
+    sys.modules[__name__], DATA_EXTENSION_KEY, _is_valid_data_explainer
+)
diff --git a/python/interpret-core/interpret/ext/extension.py b/python/interpret-core/interpret/ext/extension.py
@@ -7,6 +7,8 @@
 PROVIDER_EXTENSION_KEY = "interpret_ext_provider"
 BLACKBOX_EXTENSION_KEY = "interpret_ext_blackbox"
 GREYBOX_EXTENSION_KEY = "interpret_ext_greybox"
+DATA_EXTENSION_KEY = "interpret_ext_data"
+PERF_EXTENSION_KEY = "interpret_ext_perf"
 
 
 def _is_valid_explainer(target_explainer_type, proposed_explainer):
@@ -45,6 +47,14 @@ def _is_valid_greybox_explainer(proposed_explainer):
     return _is_valid_explainer("specific", proposed_explainer)
 
 
+def _is_valid_data_explainer(proposed_explainer):
+    return _is_valid_explainer("data", proposed_explainer)
+
+
+def _is_valid_perf_explainer(proposed_explainer):
+    return _is_valid_explainer("perf", proposed_explainer)
+
+
 def _is_valid_provider(proposed_provider):
     try:
         has_render_method = hasattr(proposed_provider, "render")

diff --git a/python/interpret-core/interpret/ext/greybox/__init__.py b/python/interpret-core/interpret/ext/greybox/__init__.py
@@ -3,8 +3,8 @@
 
 import sys
 from interpret.ext.extension_utils import load_class_extensions
-from interpret.ext.extension import BLACKBOX_EXTENSION_KEY, _is_valid_blackbox_explainer
+from interpret.ext.extension import GREYBOX_EXTENSION_KEY, _is_valid_greybox_explainer
 
 load_class_extensions(
-    sys.modules[__name__], BLACKBOX_EXTENSION_KEY, _is_valid_blackbox_explainer
+    sys.modules[__name__], GREYBOX_EXTENSION_KEY, _is_valid_greybox_explainer
 )
diff --git a/python/interpret-core/interpret/ext/perf/__init__.py b/python/interpret-core/interpret/ext/perf/__init__.py
@@ -0,0 +1,10 @@
+# Copyright (c) 2019 Microsoft Corporation
+# Distributed under the MIT software license
+
+import sys
+from interpret.ext.extension_utils import load_class_extensions
+from interpret.ext.extension import PERF_EXTENSION_KEY, _is_valid_perf_explainer
+
+load_class_extensions(
+    sys.modules[__name__], PERF_EXTENSION_KEY, _is_valid_perf_explainer
+)
diff --git a/python/interpret-core/interpret/greybox/__init__.py b/python/interpret-core/interpret/greybox/__init__.py
@@ -2,3 +2,4 @@
 # Distributed under the MIT software license
 
 from .treeinterpreter import TreeInterpreter  # noqa: F401
+from .shaptree import ShapTree  # noqa: F401
diff --git a/python/interpret-core/interpret/greybox/shaptree.py b/python/interpret-core/interpret/greybox/shaptree.py
@@ -0,0 +1,56 @@
+# Copyright (c) 2019 Microsoft Corporation
+# Distributed under the MIT software license
+
+from ..utils.shap import shap_explain_local
+from sklearn.base import is_classifier
+
+from ..api.base import ExplainerMixin
+from ..utils import unify_predict_fn, unify_data
+
+
+class ShapTree(ExplainerMixin):
+    available_explanations = ["local"]
+    explainer_type = "specific"
+
+    def __init__(
+            self,
+            model,
+            data,
+            feature_names=None,
+            feature_types=None,
+            explain_kwargs=None,
+            n_jobs=1,
+            **kwargs
+    ):
+        import shap
+
+        self.model = model
+        if is_classifier(self):
+            predict_fn = lambda x: self.model.predict_proba(x)[:, 1]
+        else:
+            predict_fn = self.model.predict
+        self.predict_fn = unify_predict_fn(predict_fn, self.data)
+
+        self.data, _, self.feature_names, self.feature_types = unify_data(
+            data, None, feature_names, feature_types
+        )
+        self.n_jobs = n_jobs
+
+        self.explain_kwargs = explain_kwargs
+        self.kwargs = kwargs
+
+        self.shap = shap.TreeExplainer(model, data, **self.kwargs)
+
+    def explain_local(self, X, y=None, name=None):
+        """ Provides local explanations for provided instances.
+
+        Args:
+            X: Numpy array for X to explain.
+            y: Numpy vector for y to explain.
+            name: User-defined explanation name.
+
+        Returns:
+            An explanation object, visualizing feature-value pairs
+            for each instance as horizontal bar charts.
+        """
+        return shap_explain_local(self, X, y=y, name=name)
diff --git a/python/interpret-core/interpret/provider/visualize.py b/python/interpret-core/interpret/provider/visualize.py
@@ -30,6 +30,11 @@ def render(self, explanation, key=-1, **kwargs):
         selector_key = kwargs.pop('selector_key', None)
         file_name = kwargs.pop('file_name', None)
 
+        # NOTE: Preserve didn't support returning everything. If key is -1 default to key is None.
+        # This is for backward-compatibility. All of this will be deprecated shortly anyway.
+        if key == -1:
+            key = None
+
         # Get visual object
         visual = explanation.visualize(key=key)
 

diff --git a/python/interpret-core/interpret/test/test_explainers.py b/python/interpret-core/interpret/test/test_explainers.py
@@ -1,6 +1,7 @@
 # Copyright (c) 2019 Microsoft Corporation
 # Distributed under the MIT software license
 
+from sklearn.ensemble import RandomForestClassifier
 
 from .utils import synthetic_classification, get_all_explainers
 from .utils import assert_valid_explanation, assert_valid_model_explainer
@@ -10,12 +11,16 @@
 import pytest
 
 
+# TODO: Generalize specific models (currently only testing trees)
 @pytest.mark.slow
 def test_spec_synthetic():
     all_explainers = get_all_explainers()
     data = synthetic_classification()
+
     blackbox = LogisticRegression()
     blackbox.fit(data["train"]["X"], data["train"]["y"])
+    tree = RandomForestClassifier()
+    tree.fit(data["train"]["X"], data["train"]["y"])
 
     predict_fn = lambda x: blackbox.predict_proba(x)  # noqa: E731
 
@@ -24,9 +29,10 @@ def test_spec_synthetic():
             explainer = explainer_class(predict_fn, data["train"]["X"])
         elif explainer_class.explainer_type == "model":
             explainer = explainer_class()
-
             explainer.fit(data["train"]["X"], data["train"]["y"])
             assert_valid_model_explainer(explainer, data["test"]["X"].head())
+        elif explainer_class.explainer_type == "specific":
+            explainer = explainer_class(tree, data["train"]["X"])
         elif explainer_class.explainer_type == "data":
             explainer = explainer_class()
         elif explainer_class.explainer_type == "perf":

diff --git a/python/interpret-core/interpret/test/utils.py b/python/interpret-core/interpret/test/utils.py
@@ -9,6 +9,9 @@
 from ..blackbox import MorrisSensitivity
 from ..blackbox import PartialDependence
 
+from ..greybox import TreeInterpreter
+from ..greybox import ShapTree
+
 from ..glassbox import LogisticRegression, LinearRegression
 from ..glassbox import ClassificationTree, RegressionTree
 from ..glassbox import DecisionListClassifier
@@ -35,6 +38,10 @@ def get_all_explainers():
         LinearRegression,
         ExplainableBoostingRegressor,
     ]
+    specific_explainer_classes = [
+        TreeInterpreter,
+        ShapTree,
+    ]
     blackbox_explainer_classes = [
         LimeTabular,
         ShapKernel,
@@ -43,6 +50,7 @@ def get_all_explainers():
     ]
     all_explainers = []
     all_explainers.extend(model_explainer_classes)
+    all_explainers.extend(specific_explainer_classes)
     all_explainers.extend(blackbox_explainer_classes)
     all_explainers.extend(data_explainer_classes)
     all_explainers.extend(perf_explainer_classes)

diff --git a/python/interpret-core/interpret/utils/shap.py b/python/interpret-core/interpret/utils/shap.py
@@ -0,0 +1,68 @@
+# Copyright (c) 2019 Microsoft Corporation
+# Distributed under the MIT software license
+
+from ..api.templates import FeatureValueExplanation
+from . import gen_name_from_class, unify_data, perf_dict, gen_local_selector
+
+
+def shap_explain_local(explainer, X, y=None, name=None):
+    if name is None:
+        name = gen_name_from_class(explainer)
+    X, y, _, _ = unify_data(X, y, explainer.feature_names, explainer.feature_types)
+
+    all_shap_values = explainer.shap.shap_values(X)
+    predictions = explainer.predict_fn(X)
+
+    data_dicts = []
+    scores_list = all_shap_values
+    perf_list = []
+    for i, instance in enumerate(X):
+        shap_values = all_shap_values[i]
+        perf_dict_obj = perf_dict(y, predictions, i)
+
+        perf_list.append(perf_dict_obj)
+
+        data_dict = {
+            "type": "univariate",
+            "names": explainer.feature_names,
+            "perf": perf_dict_obj,
+            "scores": shap_values,
+            "values": instance,
+            "extra": {
+                "names": ["Base Value"],
+                "scores": [explainer.shap.expected_value],
+                "values": [1],
+            },
+        }
+        data_dicts.append(data_dict)
+
+    internal_obj = {
+        "overall": None,
+        "specific": data_dicts,
+        "mli": [
+            {
+                "explanation_type": "local_feature_importance",
+                "value": {
+                    "scores": scores_list,
+                    "intercept": explainer.shap.expected_value,
+                    "perf": perf_list,
+                },
+            }
+        ],
+    }
+    internal_obj["mli"].append(
+        {
+            "explanation_type": "evaluation_dataset",
+            "value": {"dataset_x": X, "dataset_y": y},
+        }
+    )
+    selector = gen_local_selector(X, y, predictions)
+
+    return FeatureValueExplanation(
+        "local",
+        internal_obj,
+        feature_names=explainer.feature_names,
+        feature_types=explainer.feature_types,
+        name=name,
+        selector=selector,
+    )
Original file line number	Diff line number	Diff line change
Expand Up		@@ -2,3 +2,4 @@
		# Distributed under the MIT software license

		from .treeinterpreter import TreeInterpreter # noqa: F401
		from .shaptree import ShapTree # noqa: F401