feat: Add XAI SDK integration to TensorFlow models with LIT integration (#917)

taiseiak · web-flow · commit ea2b5cfbcafe · 2022-01-19T16:04:12.000Z
Add automatic addition of feature attribution for TensorFlow 2 models in the LIT integration on Vertex Notebooks. Detects for Vertex Notebooks by looking for the same environment variable to check for Vertex Notebooks as the LIT library does.

Fixes b/210943910 &#129429;

go/local-explanations-lit-xai-notebook
diff --git a/google/cloud/aiplatform/explain/lit.py b/google/cloud/aiplatform/explain/lit.py
@@ -14,10 +14,14 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
-from typing import Dict, List, Tuple, Union
+import logging
+import os
+
+from typing import Dict, List, Optional, Tuple, Union
 
 try:
     from lit_nlp.api import dataset as lit_dataset
+    from lit_nlp.api import dtypes as lit_dtypes
     from lit_nlp.api import model as lit_model
     from lit_nlp.api import types as lit_types
     from lit_nlp import notebook
@@ -82,6 +86,7 @@ def __init__(
         model: str,
         input_types: "OrderedDict[str, lit_types.LitType]",  # noqa: F821
         output_types: "OrderedDict[str, lit_types.LitType]",  # noqa: F821
+        attribution_method: str = "sampled_shapley",
     ):
         """Construct a VertexLitModel.
             Args:
@@ -94,39 +99,33 @@ def __init__(
               output_types:
                 Required. An OrderedDict of string names matching the labels of the model
                 as the key, and the associated LitType of the label.
+              attribution_method:
+                Optional. A string to choose what attribution configuration to
+                set up the explainer with. Valid options are 'sampled_shapley'
+                or 'integrated_gradients'.
         """
-        self._loaded_model = tf.saved_model.load(model)
-        serving_default = self._loaded_model.signatures[
-            tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY
-        ]
-        _, self._kwargs_signature = serving_default.structured_input_signature
-        self._output_signature = serving_default.structured_outputs
-
-        if len(self._kwargs_signature) != 1:
-            raise ValueError("Please use a model with only one input tensor.")
-
-        if len(self._output_signature) != 1:
-            raise ValueError("Please use a model with only one output tensor.")
-
+        self._load_model(model)
         self._input_types = input_types
         self._output_types = output_types
+        self._input_tensor_name = next(iter(self._kwargs_signature))
+        self._attribution_explainer = None
+        if os.environ.get("LIT_PROXY_URL"):
+            self._set_up_attribution_explainer(model, attribution_method)
+
+    @property
+    def attribution_explainer(self,) -> Optional["AttributionExplainer"]:  # noqa: F821
+        """Gets the attribution explainer property if set."""
+        return self._attribution_explainer
 
     def predict_minibatch(
         self, inputs: List[lit_types.JsonDict]
     ) -> List[lit_types.JsonDict]:
-        """Returns predictions for a single batch of examples.
-            Args:
-              inputs:
-                sequence of inputs, following model.input_spec()
-            Returns:
-                list of outputs, following model.output_spec()
-        """
         instances = []
         for input in inputs:
             instance = [input[feature] for feature in self._input_types]
             instances.append(instance)
         prediction_input_dict = {
-            next(iter(self._kwargs_signature)): tf.convert_to_tensor(instances)
+            self._input_tensor_name: tf.convert_to_tensor(instances)
         }
         prediction_dict = self._loaded_model.signatures[
             tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY
@@ -140,6 +139,15 @@ def predict_minibatch(
                     for label, value in zip(self._output_types.keys(), prediction)
                 }
             )
+        # Get feature attributions
+        if self.attribution_explainer:
+            attributions = self.attribution_explainer.explain(
+                [{self._input_tensor_name: i} for i in instances]
+            )
+            for i, attribution in enumerate(attributions):
+                outputs[i]["feature_attribution"] = lit_dtypes.FeatureSalience(
+                    attribution.feature_importance()
+                )
         return outputs
 
     def input_spec(self) -> lit_types.Spec:
@@ -148,7 +156,70 @@ def input_spec(self) -> lit_types.Spec:
 
     def output_spec(self) -> lit_types.Spec:
         """Return a spec describing model outputs."""
-        return self._output_types
+        output_spec_dict = dict(self._output_types)
+        if self.attribution_explainer:
+            output_spec_dict["feature_attribution"] = lit_types.FeatureSalience(
+                signed=True
+            )
+        return output_spec_dict
+
+    def _load_model(self, model: str):
+        """Loads a TensorFlow saved model and populates the input and output signature attributes of the class.
+            Args:
+              model: Required. A string reference to a TensorFlow saved model directory.
+            Raises:
+                ValueError if the model has more than one input tensor or more than one output tensor.
+            """
+        self._loaded_model = tf.saved_model.load(model)
+        serving_default = self._loaded_model.signatures[
+            tf.saved_model.DEFAULT_SERVING_SIGNATURE_DEF_KEY
+        ]
+        _, self._kwargs_signature = serving_default.structured_input_signature
+        self._output_signature = serving_default.structured_outputs
+
+        if len(self._kwargs_signature) != 1:
+            raise ValueError("Please use a model with only one input tensor.")
+
+        if len(self._output_signature) != 1:
+            raise ValueError("Please use a model with only one output tensor.")
+
+    def _set_up_attribution_explainer(
+        self, model: str, attribution_method: str = "integrated_gradients"
+    ):
+        """Populates the attribution explainer attribute of the class.
+            Args:
+              model: Required. A string reference to a TensorFlow saved model directory.
+            attribution_method:
+              Optional. A string to choose what attribution configuration to
+              set up the explainer with. Valid options are 'sampled_shapley'
+              or 'integrated_gradients'.
+        """
+        try:
+            import explainable_ai_sdk
+            from explainable_ai_sdk.metadata.tf.v2 import SavedModelMetadataBuilder
+        except ImportError:
+            logging.info(
+                "Skipping explanations because the Explainable AI SDK is not installed."
+                'Please install the SDK using "pip install explainable-ai-sdk"'
+            )
+            return
+
+        builder = SavedModelMetadataBuilder(model)
+        builder.get_metadata()
+        builder.set_numeric_metadata(
+            self._input_tensor_name,
+            index_feature_mapping=list(self._input_types.keys()),
+        )
+        builder.save_metadata(model)
+        if attribution_method == "integrated_gradients":
+            explainer_config = explainable_ai_sdk.IntegratedGradientsConfig()
+        else:
+            explainer_config = explainable_ai_sdk.SampledShapleyConfig()
+
+        self._attribution_explainer = explainable_ai_sdk.load_model_from_local_path(
+            model, explainer_config
+        )
+        self._load_model(model)
 
 
 def create_lit_dataset(
@@ -172,22 +243,27 @@ def create_lit_model(
     model: str,
     input_types: "OrderedDict[str, lit_types.LitType]",  # noqa: F821
     output_types: "OrderedDict[str, lit_types.LitType]",  # noqa: F821
+    attribution_method: str = "sampled_shapley",
 ) -> lit_model.Model:
     """Creates a LIT Model object.
         Args:
           model:
-              Required. A string reference to a local TensorFlow saved model directory.
-              The model must have at most one input and one output tensor.
+            Required. A string reference to a local TensorFlow saved model directory.
+            The model must have at most one input and one output tensor.
           input_types:
-              Required. An OrderedDict of string names matching the features of the model
-              as the key, and the associated LitType of the feature.
+            Required. An OrderedDict of string names matching the features of the model
+            as the key, and the associated LitType of the feature.
           output_types:
-              Required. An OrderedDict of string names matching the labels of the model
-              as the key, and the associated LitType of the label.
+            Required. An OrderedDict of string names matching the labels of the model
+            as the key, and the associated LitType of the label.
+          attribution_method:
+            Optional. A string to choose what attribution configuration to
+            set up the explainer with. Valid options are 'sampled_shapley'
+            or 'integrated_gradients'.
         Returns:
             A LIT Model object that has the same functionality as the model provided.
     """
-    return _VertexLitModel(model, input_types, output_types)
+    return _VertexLitModel(model, input_types, output_types, attribution_method)
 
 
 def open_lit(
@@ -198,11 +274,11 @@ def open_lit(
     """Open LIT from the provided models and datasets.
         Args:
           models:
-              Required. A list of LIT models to open LIT with.
+            Required. A list of LIT models to open LIT with.
           input_types:
-              Required. A lit of LIT datasets to open LIT with.
+            Required. A lit of LIT datasets to open LIT with.
           open_in_new_tab:
-              Optional. A boolean to choose if LIT open in a new tab or not.
+            Optional. A boolean to choose if LIT open in a new tab or not.
         Raises:
             ImportError if LIT is not installed.
     """
@@ -216,24 +292,31 @@ def set_up_and_open_lit(
     model: Union[str, lit_model.Model],
     input_types: Union[List[str], Dict[str, lit_types.LitType]],
     output_types: Union[str, List[str], Dict[str, lit_types.LitType]],
+    attribution_method: str = "sampled_shapley",
     open_in_new_tab: bool = True,
 ) -> Tuple[lit_dataset.Dataset, lit_model.Model]:
     """Creates a LIT dataset and model and opens LIT.
         Args:
-        dataset:
+          dataset:
             Required. A Pandas DataFrame that includes feature column names and data.
-        column_types:
+          column_types:
             Required. An OrderedDict of string names matching the columns of the dataset
             as the key, and the associated LitType of the column.
-        model:
+          model:
             Required. A string reference to a TensorFlow saved model directory.
             The model must have at most one input and one output tensor.
-        input_types:
+          input_types:
             Required. An OrderedDict of string names matching the features of the model
             as the key, and the associated LitType of the feature.
-        output_types:
+          output_types:
             Required. An OrderedDict of string names matching the labels of the model
             as the key, and the associated LitType of the label.
+          attribution_method:
+            Optional. A string to choose what attribution configuration to
+            set up the explainer with. Valid options are 'sampled_shapley'
+            or 'integrated_gradients'.
+          open_in_new_tab:
+            Optional. A boolean to choose if LIT open in a new tab or not.
         Returns:
             A Tuple of the LIT dataset and model created.
         Raises:
@@ -244,8 +327,12 @@ def set_up_and_open_lit(
         dataset = create_lit_dataset(dataset, column_types)
 
     if not isinstance(model, lit_model.Model):
-        model = create_lit_model(model, input_types, output_types)
+        model = create_lit_model(
+            model, input_types, output_types, attribution_method=attribution_method
+        )
 
-    open_lit({"model": model}, {"dataset": dataset}, open_in_new_tab=open_in_new_tab)
+    open_lit(
+        {"model": model}, {"dataset": dataset}, open_in_new_tab=open_in_new_tab,
+    )
 
     return dataset, model
diff --git a/setup.py b/setup.py
@@ -36,7 +36,12 @@
 tensorboard_extra_require = ["tensorflow >=2.3.0, <=2.7.0"]
 metadata_extra_require = ["pandas >= 1.0.0"]
 xai_extra_require = ["tensorflow >=2.3.0, <=2.5.0"]
-lit_extra_require = ["tensorflow >= 2.3.0", "pandas >= 1.0.0", "lit-nlp >= 0.4.0"]
+lit_extra_require = [
+    "tensorflow >= 2.3.0",
+    "pandas >= 1.0.0",
+    "lit-nlp >= 0.4.0",
+    "explainable-ai-sdk >= 1.0.0",
+]
 profiler_extra_require = [
     "tensorboard-plugin-profile >= 2.4.0",
     "werkzeug >= 2.0.0",
diff --git a/tests/unit/aiplatform/test_explain_lit.py b/tests/unit/aiplatform/test_explain_lit.py
