Merge pull request #864 from cliveseldon/explainers

Create explainer notebook and fix tfserving proxy
SeldonIO · Sep 16, 2019 · 6747a21 · 6747a21
2 parents ec41863 + 08a3a52
commit 6747a21
Show file tree

Hide file tree

Showing 16 changed files with 1,129 additions and 25 deletions.
diff --git a/doc/source/analytics/cat.png b/doc/source/analytics/cat.png
diff --git a/doc/source/analytics/cat_explanation.png b/doc/source/analytics/cat_explanation.png
diff --git a/doc/source/analytics/explainers.md b/doc/source/analytics/explainers.md
@@ -0,0 +1,10 @@
+# Model Explainers
+
+![cat](cat.png)
+![explanation](cat_explanation.png)
+
+Seldon provides model explanations using its [Alibi](https://github.com/SeldonIO/alibi)  Open Sourve library.
+
+We provide [an example notebook](../examples/explainer_examples.html) showing how to deploy an explainer for Tabular, Text and Image models.
+
+
diff --git a/doc/source/examples/explainer_examples.nblink b/doc/source/examples/explainer_examples.nblink
@@ -0,0 +1,3 @@
+{
+  "path": "../../../notebooks/explainer_examples.ipynb"
+}
diff --git a/doc/source/examples/notebooks.rst b/doc/source/examples/notebooks.rst
@@ -18,6 +18,7 @@ Notebooks
    Custom Endpoints <custom_endpoints>
    Example Helm Deployments <helm_examples>
    Explainer Alibi Anchor Tabular <alibi_anchor_tabular>
+   Tabular, Text and Image Model Explainers <explainer_examples>
    Go Model <go_example>
    GPU Tensorflow Deep MNIST <gpu_tensorflow_deep_mnist>
    H2O Java MoJo <h2o_mojo>

diff --git a/doc/source/index.rst b/doc/source/index.rst
@@ -88,6 +88,7 @@ Seldon Core is an open source platform for deploying machine learning models on
    :maxdepth: 1
    :caption: ML Compliance and Governance
 
+   Model Explanations <analytics/explainers.md>
    Outlier Detection <analytics/outlier_detection.md>
    Routers (incl. Multi Armed Bandits)  <analytics/routers.md>   
 

diff --git a/engine/src/main/java/io/seldon/engine/service/InternalPredictionService.java b/engine/src/main/java/io/seldon/engine/service/InternalPredictionService.java
@@ -247,6 +247,7 @@ else if (state.type == PredictiveUnitType.MODEL)
 
     public SeldonMessage transformInput(SeldonMessage input, PredictiveUnitState state) throws InvalidProtocolBufferException
     {
+    	logger.info("Calling grpc for transform-input");
     	final Endpoint endpoint = state.endpoint;
 		switch (endpoint.getType()){
 			case REST:
@@ -267,11 +268,20 @@ public SeldonMessage transformInput(SeldonMessage input, PredictiveUnitState sta
 							.withMaxOutboundMessageSize(grpcMaxMessageSize);
 						return genStub.transformInput(input);
 					case MODEL:
-						ModelBlockingStub modelStub = ModelGrpc.newBlockingStub(grpcChannelHandler.get(endpoint))
-							.withDeadlineAfter(grpcReadTimeout, TimeUnit.MILLISECONDS)
-							.withMaxInboundMessageSize(grpcMaxMessageSize)
-							.withMaxOutboundMessageSize(grpcMaxMessageSize);
-						return modelStub.predict(input);
+						try
+						{
+							ModelBlockingStub modelStub = ModelGrpc.newBlockingStub(grpcChannelHandler.get(endpoint))
+								.withDeadlineAfter(grpcReadTimeout, TimeUnit.MILLISECONDS)
+								.withMaxInboundMessageSize(grpcMaxMessageSize)
+								.withMaxOutboundMessageSize(grpcMaxMessageSize);
+							logger.info(modelStub.getCallOptions().toString());
+							return modelStub.predict(input);
+						}
+						catch (Exception e)
+						{
+							logger.error("grpc exception ",e);
+							throw e;
+						}
 					case TRANSFORMER:
 						TransformerBlockingStub transformerStub = TransformerGrpc.newBlockingStub(grpcChannelHandler.get(endpoint))
 						.withDeadlineAfter(grpcReadTimeout, TimeUnit.MILLISECONDS)

diff --git a/integrations/tfserving/Makefile b/integrations/tfserving/Makefile
@@ -1,15 +1,15 @@
-IMAGE_VERSION=0.6
+IMAGE_VERSION=0.7
 IMAGE_NAME = docker.io/seldonio/tfserving-proxy
 
 SELDON_CORE_DIR=../../..
 
 .PHONY: build_rest
 build_rest: 
-	s2i build -E environment_rest . seldonio/seldon-core-s2i-python3:0.11-SNAPSHOT $(IMAGE_NAME)_rest:$(IMAGE_VERSION)
+	s2i build -E environment_rest . seldonio/seldon-core-s2i-python3:0.12-SNAPSHOT $(IMAGE_NAME)_rest:$(IMAGE_VERSION)
 
 .PHONY: build_grpc
 build_grpc: 
-	s2i build -E environment_grpc . seldonio/seldon-core-s2i-python3:0.11-SNAPSHOT $(IMAGE_NAME)_grpc:$(IMAGE_VERSION)
+	s2i build -E environment_grpc . seldonio/seldon-core-s2i-python3:0.12-SNAPSHOT $(IMAGE_NAME)_grpc:$(IMAGE_VERSION)
 
 
 push_to_dockerhub_rest:

diff --git a/integrations/tfserving/TfServingProxy.py b/integrations/tfserving/TfServingProxy.py
@@ -34,7 +34,11 @@ def __init__(
         log.debug("grpc_endpoint:",grpc_endpoint)
         if not grpc_endpoint is None:
             self.grpc = True
-            channel = grpc.insecure_channel(grpc_endpoint)
+            max_msg = 1000000000
+            options = [('grpc.max_message_length', max_msg),
+                       ('grpc.max_send_message_length', max_msg),
+                       ('grpc.max_receive_message_length', max_msg)]
+            channel = grpc.insecure_channel(grpc_endpoint,options)
             self.stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)
         else:
             self.grpc = False
@@ -80,13 +84,10 @@ def predict_grpc(self,request):
                     data_arr.tolist(),
                     shape=data_arr.shape))
             result = self.stub.Predict(tfrequest)
-            result_arr = numpy.array(result.outputs[self.model_output].float_val)
-            if len(result_arr.shape) == 1:
-                result_arr = numpy.expand_dims(result_arr, axis=0)
-            class_names = []
-            data = array_to_grpc_datadef(
-                default_data_type, result_arr, class_names)
-            return prediction_pb2.SeldonMessage(data=data)
+            datadef = prediction_pb2.DefaultData(
+                tftensor=result.outputs[self.model_output]
+            )
+            return prediction_pb2.SeldonMessage(data=datadef)
 
     def predict(self, X, features_names=[]):
         """