Upstream internal changes

tensorflow_serving/apis/prediction_service_pb2.py

@@ -14,6 +14,8 @@
 # ==============================================================================
 # Generated by the protocol buffer compiler.  DO NOT EDIT!
 # source: tensorflow_serving/apis/prediction_service.proto
+# To regenerate run
+# python -m grpc.tools.protoc --python_out=. --grpc_python_out=. -I. tensorflow_serving/apis/prediction_service.proto
 
 import sys
 _b=sys.version_info[0]<3 and (lambda x:x) or (lambda x:x.encode('latin1'))
@@ -27,17 +29,19 @@
 _sym_db = _symbol_database.Default()
 
 
+from tensorflow_serving.apis import classification_pb2 as tensorflow__serving_dot_apis_dot_classification__pb2
 from tensorflow_serving.apis import get_model_metadata_pb2 as tensorflow__serving_dot_apis_dot_get__model__metadata__pb2
 from tensorflow_serving.apis import predict_pb2 as tensorflow__serving_dot_apis_dot_predict__pb2
+from tensorflow_serving.apis import regression_pb2 as tensorflow__serving_dot_apis_dot_regression__pb2
 
 
 DESCRIPTOR = _descriptor.FileDescriptor(
   name='tensorflow_serving/apis/prediction_service.proto',
   package='tensorflow.serving',
   syntax='proto3',
-  serialized_pb=_b('\n0tensorflow_serving/apis/prediction_service.proto\x12\x12tensorflow.serving\x1a\x30tensorflow_serving/apis/get_model_metadata.proto\x1a%tensorflow_serving/apis/predict.proto2\xd6\x01\n\x11PredictionService\x12R\n\x07Predict\x12\".tensorflow.serving.PredictRequest\x1a#.tensorflow.serving.PredictResponse\x12m\n\x10GetModelMetadata\x12+.tensorflow.serving.GetModelMetadataRequest\x1a,.tensorflow.serving.GetModelMetadataResponseB\x03\xf8\x01\x01\x62\x06proto3')
+  serialized_pb=_b('\n0tensorflow_serving/apis/prediction_service.proto\x12\x12tensorflow.serving\x1a,tensorflow_serving/apis/classification.proto\x1a\x30tensorflow_serving/apis/get_model_metadata.proto\x1a%tensorflow_serving/apis/predict.proto\x1a(tensorflow_serving/apis/regression.proto2\x93\x03\n\x11PredictionService\x12\x61\n\x08\x43lassify\x12).tensorflow.serving.ClassificationRequest\x1a*.tensorflow.serving.ClassificationResponse\x12X\n\x07Regress\x12%.tensorflow.serving.RegressionRequest\x1a&.tensorflow.serving.RegressionResponse\x12R\n\x07Predict\x12\".tensorflow.serving.PredictRequest\x1a#.tensorflow.serving.PredictResponse\x12m\n\x10GetModelMetadata\x12+.tensorflow.serving.GetModelMetadataRequest\x1a,.tensorflow.serving.GetModelMetadataResponseB\x03\xf8\x01\x01\x62\x06proto3')
   ,
-  dependencies=[tensorflow__serving_dot_apis_dot_get__model__metadata__pb2.DESCRIPTOR,tensorflow__serving_dot_apis_dot_predict__pb2.DESCRIPTOR,])
+  dependencies=[tensorflow__serving_dot_apis_dot_classification__pb2.DESCRIPTOR,tensorflow__serving_dot_apis_dot_get__model__metadata__pb2.DESCRIPTOR,tensorflow__serving_dot_apis_dot_predict__pb2.DESCRIPTOR,tensorflow__serving_dot_apis_dot_regression__pb2.DESCRIPTOR,])
 _sym_db.RegisterFileDescriptor(DESCRIPTOR)
 
 
@@ -66,6 +70,16 @@ def __init__(self, channel):
     Args:
       channel: A grpc.Channel.
     """
+    self.Classify = channel.unary_unary(
+        '/tensorflow.serving.PredictionService/Classify',
+        request_serializer=tensorflow__serving_dot_apis_dot_classification__pb2.ClassificationRequest.SerializeToString,
+        response_deserializer=tensorflow__serving_dot_apis_dot_classification__pb2.ClassificationResponse.FromString,
+        )
+    self.Regress = channel.unary_unary(
+        '/tensorflow.serving.PredictionService/Regress',
+        request_serializer=tensorflow__serving_dot_apis_dot_regression__pb2.RegressionRequest.SerializeToString,
+        response_deserializer=tensorflow__serving_dot_apis_dot_regression__pb2.RegressionResponse.FromString,
+        )
     self.Predict = channel.unary_unary(
         '/tensorflow.serving.PredictionService/Predict',
         request_serializer=tensorflow__serving_dot_apis_dot_predict__pb2.PredictRequest.SerializeToString,
@@ -84,6 +98,20 @@ class PredictionServiceServicer(object):
   model_servers.
   """
 
+  def Classify(self, request, context):
+    """Classify.
+    """
+    context.set_code(grpc.StatusCode.UNIMPLEMENTED)
+    context.set_details('Method not implemented!')
+    raise NotImplementedError('Method not implemented!')
+
+  def Regress(self, request, context):
+    """Regress.
+    """
+    context.set_code(grpc.StatusCode.UNIMPLEMENTED)
+    context.set_details('Method not implemented!')
+    raise NotImplementedError('Method not implemented!')
+
   def Predict(self, request, context):
     """Predict -- provides access to loaded TensorFlow model.
     """
@@ -101,6 +129,16 @@ def GetModelMetadata(self, request, context):
 
 def add_PredictionServiceServicer_to_server(servicer, server):
   rpc_method_handlers = {
+      'Classify': grpc.unary_unary_rpc_method_handler(
+          servicer.Classify,
+          request_deserializer=tensorflow__serving_dot_apis_dot_classification__pb2.ClassificationRequest.FromString,
+          response_serializer=tensorflow__serving_dot_apis_dot_classification__pb2.ClassificationResponse.SerializeToString,
+      ),
+      'Regress': grpc.unary_unary_rpc_method_handler(
+          servicer.Regress,
+          request_deserializer=tensorflow__serving_dot_apis_dot_regression__pb2.RegressionRequest.FromString,
+          response_serializer=tensorflow__serving_dot_apis_dot_regression__pb2.RegressionResponse.SerializeToString,
+      ),
       'Predict': grpc.unary_unary_rpc_method_handler(
           servicer.Predict,
           request_deserializer=tensorflow__serving_dot_apis_dot_predict__pb2.PredictRequest.FromString,
@@ -127,6 +165,14 @@ class BetaPredictionServiceServicer(object):
   PredictionService provides access to machine-learned models loaded by
   model_servers.
   """
+  def Classify(self, request, context):
+    """Classify.
+    """
+    context.code(beta_interfaces.StatusCode.UNIMPLEMENTED)
+  def Regress(self, request, context):
+    """Regress.
+    """
+    context.code(beta_interfaces.StatusCode.UNIMPLEMENTED)
   def Predict(self, request, context):
     """Predict -- provides access to loaded TensorFlow model.
     """
@@ -147,6 +193,16 @@ class BetaPredictionServiceStub(object):
   PredictionService provides access to machine-learned models loaded by
   model_servers.
   """
+  def Classify(self, request, timeout, metadata=None, with_call=False, protocol_options=None):
+    """Classify.
+    """
+    raise NotImplementedError()
+  Classify.future = None
+  def Regress(self, request, timeout, metadata=None, with_call=False, protocol_options=None):
+    """Regress.
+    """
+    raise NotImplementedError()
+  Regress.future = None
   def Predict(self, request, timeout, metadata=None, with_call=False, protocol_options=None):
     """Predict -- provides access to loaded TensorFlow model.
     """
@@ -166,16 +222,22 @@ def beta_create_PredictionService_server(servicer, pool=None, pool_size=None, de
   file not marked beta) for all further purposes. This function was
   generated only to ease transition from grpcio<0.15.0 to grpcio>=0.15.0"""
   request_deserializers = {
+    ('tensorflow.serving.PredictionService', 'Classify'): tensorflow__serving_dot_apis_dot_classification__pb2.ClassificationRequest.FromString,
     ('tensorflow.serving.PredictionService', 'GetModelMetadata'): tensorflow__serving_dot_apis_dot_get__model__metadata__pb2.GetModelMetadataRequest.FromString,
     ('tensorflow.serving.PredictionService', 'Predict'): tensorflow__serving_dot_apis_dot_predict__pb2.PredictRequest.FromString,
+    ('tensorflow.serving.PredictionService', 'Regress'): tensorflow__serving_dot_apis_dot_regression__pb2.RegressionRequest.FromString,
   }
   response_serializers = {
+    ('tensorflow.serving.PredictionService', 'Classify'): tensorflow__serving_dot_apis_dot_classification__pb2.ClassificationResponse.SerializeToString,
     ('tensorflow.serving.PredictionService', 'GetModelMetadata'): tensorflow__serving_dot_apis_dot_get__model__metadata__pb2.GetModelMetadataResponse.SerializeToString,
     ('tensorflow.serving.PredictionService', 'Predict'): tensorflow__serving_dot_apis_dot_predict__pb2.PredictResponse.SerializeToString,
+    ('tensorflow.serving.PredictionService', 'Regress'): tensorflow__serving_dot_apis_dot_regression__pb2.RegressionResponse.SerializeToString,
   }
   method_implementations = {
+    ('tensorflow.serving.PredictionService', 'Classify'): face_utilities.unary_unary_inline(servicer.Classify),
     ('tensorflow.serving.PredictionService', 'GetModelMetadata'): face_utilities.unary_unary_inline(servicer.GetModelMetadata),
     ('tensorflow.serving.PredictionService', 'Predict'): face_utilities.unary_unary_inline(servicer.Predict),
+    ('tensorflow.serving.PredictionService', 'Regress'): face_utilities.unary_unary_inline(servicer.Regress),
   }
   server_options = beta_implementations.server_options(request_deserializers=request_deserializers, response_serializers=response_serializers, thread_pool=pool, thread_pool_size=pool_size, default_timeout=default_timeout, maximum_timeout=maximum_timeout)
   return beta_implementations.server(method_implementations, options=server_options)
@@ -188,16 +250,22 @@ def beta_create_PredictionService_stub(channel, host=None, metadata_transformer=
   file not marked beta) for all further purposes. This function was
   generated only to ease transition from grpcio<0.15.0 to grpcio>=0.15.0"""
   request_serializers = {
+    ('tensorflow.serving.PredictionService', 'Classify'): tensorflow__serving_dot_apis_dot_classification__pb2.ClassificationRequest.SerializeToString,
     ('tensorflow.serving.PredictionService', 'GetModelMetadata'): tensorflow__serving_dot_apis_dot_get__model__metadata__pb2.GetModelMetadataRequest.SerializeToString,
     ('tensorflow.serving.PredictionService', 'Predict'): tensorflow__serving_dot_apis_dot_predict__pb2.PredictRequest.SerializeToString,
+    ('tensorflow.serving.PredictionService', 'Regress'): tensorflow__serving_dot_apis_dot_regression__pb2.RegressionRequest.SerializeToString,
   }
   response_deserializers = {
+    ('tensorflow.serving.PredictionService', 'Classify'): tensorflow__serving_dot_apis_dot_classification__pb2.ClassificationResponse.FromString,
     ('tensorflow.serving.PredictionService', 'GetModelMetadata'): tensorflow__serving_dot_apis_dot_get__model__metadata__pb2.GetModelMetadataResponse.FromString,
     ('tensorflow.serving.PredictionService', 'Predict'): tensorflow__serving_dot_apis_dot_predict__pb2.PredictResponse.FromString,
+    ('tensorflow.serving.PredictionService', 'Regress'): tensorflow__serving_dot_apis_dot_regression__pb2.RegressionResponse.FromString,
   }
   cardinalities = {
+    'Classify': cardinality.Cardinality.UNARY_UNARY,
     'GetModelMetadata': cardinality.Cardinality.UNARY_UNARY,
     'Predict': cardinality.Cardinality.UNARY_UNARY,
+    'Regress': cardinality.Cardinality.UNARY_UNARY,
   }
   stub_options = beta_implementations.stub_options(host=host, metadata_transformer=metadata_transformer, request_serializers=request_serializers, response_deserializers=response_deserializers, thread_pool=pool, thread_pool_size=pool_size)
   return beta_implementations.dynamic_stub(channel, 'tensorflow.serving.PredictionService', cardinalities, options=stub_options)

tensorflow_serving/batching/batching_session.cc

@@ -126,8 +126,14 @@ class BatchingSession : public ServingSession {
              const std::vector<string>& target_node_names,
              std::vector<Tensor>* outputs) override;
 
-  // TODO(b/34971139): at the moment this method ignores run_options and
-  // run_metadata and behaves exactly like Run.
+  // RunOptions handling:
+  // Since multiple of these Run() calls get backed into a single call to the
+  // underlying Session's Run(), we select an arbitrary 'run_options' (typically
+  // they are the same across calls). The exception is the timeout; we take the
+  // largest value (after subtracting time spent in the batching queue).
+  //
+  // RunMetadata:
+  // We copy the batched call's RunMetadata to each non-batched call's output.
   Status Run(const RunOptions& run_options,
              const std::vector<std::pair<string, Tensor>>& inputs,
              const std::vector<string>& output_tensor_names,
@@ -210,21 +216,21 @@ Status BatchingSession::Create(
 }
 
 Status BatchingSession::Run(
-    const RunOptions& run_options,
     const std::vector<std::pair<string, Tensor>>& inputs,
     const std::vector<string>& output_tensor_names,
-    const std::vector<string>& target_node_names, std::vector<Tensor>* outputs,
-    RunMetadata* run_metadata) {
-  LOG(WARNING) << "Currently both run_options and run_metadata are ignored, "
-               << "see b/34971139";
-  return Run(inputs, output_tensor_names, target_node_names, outputs);
+    const std::vector<string>& target_node_names,
+    std::vector<Tensor>* outputs) {
+  RunMetadata run_metadata;
+  return Run(RunOptions(), inputs, output_tensor_names, target_node_names,
+             outputs, &run_metadata);
 }
 
 Status BatchingSession::Run(
+    const RunOptions& run_options,
     const std::vector<std::pair<string, Tensor>>& inputs,
     const std::vector<string>& output_tensor_names,
-    const std::vector<string>& target_node_names,
-    std::vector<Tensor>* outputs) {
+    const std::vector<string>& target_node_names, std::vector<Tensor>* outputs,
+    RunMetadata* run_metadata) {
   if (!target_node_names.empty()) {
     return errors::PermissionDenied(
         "BatchingSession does not support target nodes");
@@ -239,8 +245,8 @@ Status BatchingSession::Run(
     LOG(WARNING) << "Request doesn't match any declared signature. Bypassing "
                     "batcher. Request signature is: "
                  << TensorSignatureDebugString(signature);
-    return wrapped_->Run(inputs, output_tensor_names, target_node_names,
-                         outputs);
+    return wrapped_->Run(run_options, inputs, output_tensor_names,
+                         target_node_names, outputs, run_metadata);
   }
   BatchScheduler<BatchingSessionTask>* batch_scheduler =
       batch_scheduler_it->second.get();
@@ -250,12 +256,15 @@ Status BatchingSession::Run(
   Notification done;
   Status status;
   auto task = std::unique_ptr<BatchingSessionTask>(new BatchingSessionTask);
+  task->enqueue_time_micros = Env::Default()->NowMicros();
+  task->run_options = run_options;
   TF_RETURN_IF_ERROR(ComputeInputSize(inputs, &task->zeroth_dim_size));
   task->inputs = &inputs;
   task->output_tensor_names = &output_tensor_names;
   task->done = &done;
   task->status = &status;
   task->outputs = outputs;
+  task->run_metadata = run_metadata;
 
   TF_RETURN_IF_ERROR(batch_scheduler->Schedule(&task));
   done.WaitForNotification();
@@ -457,18 +466,55 @@ void BatchingSession::ProcessBatch(
     return;
   }
 
-  Status status;
+  const uint64 dequeue_time_micros = Env::Default()->NowMicros();
 
   // Regardless of the outcome, we need to propagate the status to the
   // individual tasks and signal that they are done. We use MakeCleanup() to
   // ensure that this happens no matter how we exit the method below.
+  Status status;
   auto finally = MakeCleanup([&status, &batch] {
     for (int i = 0; i < batch->num_tasks(); ++i) {
       *batch->mutable_task(i)->status = status;
       batch->mutable_task(i)->done->Notify();
     }
   });
 
+  // Make sure we have at least one task that hasn't exceeded its timeout from
+  // queue time alone, and find the latest task deadline which we'll use for the
+  // overall batch.
+  bool all_tasks_timeout_exceeded = true;
+  uint64 batch_deadline_micros = 0;
+  for (int i = 0; i < batch->num_tasks(); ++i) {
+    const BatchingSessionTask& task = batch->task(i);
+    // If the caller doesn't populate RunOptions, the timeout is 0 by default.
+    // Interpret that as "no timeout" i.e. infinity.
+    const int64 task_timeout_micros =
+        task.run_options.timeout_in_ms() <= 0
+            ? INT_MAX
+            : task.run_options.timeout_in_ms() * 1000;
+    const uint64 task_deadline_micros =
+        task.enqueue_time_micros + task_timeout_micros;
+    if (task_deadline_micros > dequeue_time_micros) {
+      all_tasks_timeout_exceeded = false;
+      if (task_deadline_micros > batch_deadline_micros) {
+        batch_deadline_micros = task_deadline_micros;
+      }
+    }
+  }
+  if (all_tasks_timeout_exceeded) {
+    status = Status(error::RESOURCE_EXHAUSTED,
+                    "Run() timeout exceeded while waiting in batching queue");
+    return;
+  }
+
+  RunOptions run_options = batch->task(0).run_options;
+  if (batch_deadline_micros == INT_MAX) {
+    run_options.set_timeout_in_ms(0);
+  } else {
+    run_options.set_timeout_in_ms(
+        (batch_deadline_micros - dequeue_time_micros) / 1000);
+  }
+
   std::vector<std::pair<string, Tensor>> merged_inputs;
   status = MergeInputTensors(signature, *batch, &merged_inputs);
   if (!status.ok()) {
@@ -478,8 +524,13 @@ void BatchingSession::ProcessBatch(
   const std::vector<string> output_tensor_names(
       signature.output_tensors.begin(), signature.output_tensors.end());
   std::vector<Tensor> combined_outputs;
-  status = wrapped_->Run(merged_inputs, output_tensor_names,
-                         {} /* target node names */, &combined_outputs);
+  RunMetadata run_metadata;
+  status = wrapped_->Run(run_options, merged_inputs, output_tensor_names,
+                         {} /* target node names */, &combined_outputs,
+                         &run_metadata);
+  for (int i = 0; i < batch->num_tasks(); ++i) {
+    *(batch->mutable_task(i)->run_metadata) = run_metadata;
+  }
   if (!status.ok()) {
     return;
   }

tensorflow_serving/batching/batching_session.h

@@ -167,6 +167,8 @@ struct BatchingSessionTask : public BatchTask {
   size_t size() const override { return zeroth_dim_size; }
 
   // Fields populated when a task is received.
+  uint64 enqueue_time_micros;
+  RunOptions run_options;
   size_t zeroth_dim_size;
   const std::vector<std::pair<string, Tensor>>* inputs;
   const std::vector<string>* output_tensor_names;
@@ -175,6 +177,7 @@ struct BatchingSessionTask : public BatchTask {
   Notification* done;
   Status* status;
   std::vector<Tensor>* outputs;
+  RunMetadata* run_metadata;
 };
 
 }  // namespace serving