Add unit-testing for savedmodel configuration.

src/core/model_config_manager.cc

@@ -66,7 +66,6 @@ ModelConfigManager::GetModelConfigPlatform(
     TF_RETURN_IF_ERROR(singleton->GetModelConfigInternal(name, &mc));
     *platform = GetPlatform(mc.platform());
     singleton->platforms_.emplace(name, *platform);
-    LOG_INFO << "Got platform for " << name;
   } else {
     *platform = itr->second;
   }

src/servables/tensorflow/base_bundle.cc

@@ -657,4 +657,107 @@ operator<<(std::ostream& out, const BaseBundle& pb)
   return out;
 }
 
+bool
+CompareDims(
+  const tensorflow::TensorShapeProto& model_shape, const DimsList& dims)
+{
+  // The first model dimension can be -1 to serve as a placeholder for
+  // batch. The batch dim doesn't appear in the configuration 'dims'.
+  const bool has_batch_dim =
+    (model_shape.dim().size() >= 1) && (model_shape.dim(0).size() == -1);
+  if (model_shape.dim().size() != (dims.size() + (has_batch_dim ? 1 : 0))) {
+    return false;
+  }
+
+  for (int i = 0; i < dims.size(); ++i) {
+    if (model_shape.dim(i + (has_batch_dim ? 1 : 0)).size() != dims[i]) {
+      return false;
+    }
+  }
+
+  return true;
+}
+
+bool
+CompareDataType(tensorflow::DataType model_dtype, DataType dtype)
+{
+  tensorflow::DataType cdtype = ConvertDataType(dtype);
+  if (cdtype == tensorflow::DT_INVALID) {
+    return false;
+  }
+
+  return model_dtype == cdtype;
+}
+
+const std::string
+DimsDebugString(const DimsList& dims)
+{
+  bool first = true;
+  std::string str;
+  str.append("[");
+  for (int i = 0; i < dims.size(); ++i) {
+    if (!first) {
+      str.append(",");
+    }
+    str.append(std::to_string(dims[i]));
+    first = false;
+  }
+  str.append("]");
+  return str;
+}
+
+const std::string
+DimsDebugString(const tensorflow::TensorShapeProto& dims)
+{
+  bool first = true;
+  std::string str;
+  str.append("[");
+  for (int i = 0; i < dims.dim().size(); ++i) {
+    if (!first) {
+      str.append(",");
+    }
+    str.append(std::to_string(dims.dim(i).size()));
+    first = false;
+  }
+  str.append("]");
+  return str;
+}
+
+tensorflow::DataType
+ConvertDataType(DataType dtype)
+{
+  switch (dtype) {
+    case DataType::TYPE_INVALID:
+      return tensorflow::DT_INVALID;
+    case DataType::TYPE_BOOL:
+      return tensorflow::DT_BOOL;
+    case DataType::TYPE_UINT8:
+      return tensorflow::DT_UINT8;
+    case DataType::TYPE_UINT16:
+      return tensorflow::DT_UINT16;
+    case DataType::TYPE_UINT32:
+      return tensorflow::DT_UINT32;
+    case DataType::TYPE_UINT64:
+      return tensorflow::DT_UINT64;
+    case DataType::TYPE_INT8:
+      return tensorflow::DT_INT8;
+    case DataType::TYPE_INT16:
+      return tensorflow::DT_INT16;
+    case DataType::TYPE_INT32:
+      return tensorflow::DT_INT32;
+    case DataType::TYPE_INT64:
+      return tensorflow::DT_INT64;
+    case DataType::TYPE_FP16:
+      return tensorflow::DT_HALF;
+    case DataType::TYPE_FP32:
+      return tensorflow::DT_FLOAT;
+    case DataType::TYPE_FP64:
+      return tensorflow::DT_DOUBLE;
+    default:
+      break;
+  }
+
+  return tensorflow::DT_INVALID;
+}
+
 }}  // namespace nvidia::inferenceserver

src/servables/tensorflow/base_bundle.h

@@ -149,4 +149,23 @@ class BaseBundle : public InferenceServable {
 
 std::ostream& operator<<(std::ostream& out, const BaseBundle& pb);
 
+/// \return true if a TensorFlow shape matches a model configuration
+/// shape.
+bool CompareDims(
+  const tensorflow::TensorShapeProto& model_shape, const DimsList& dims);
+
+/// \return true if a TensorFlow data-type matches a model
+/// configuration data-type.
+bool CompareDataType(tensorflow::DataType model_dtype, DataType dtype);
+
+/// \return the string representation of a model configuration shape.
+const std::string DimsDebugString(const DimsList& dims);
+
+/// \return the string representation of a TensorFlow shape.
+const std::string DimsDebugString(const tensorflow::TensorShapeProto& dims);
+
+/// \return the TensorFlow data-type that corresponds to a model
+/// configuration data-type.
+tensorflow::DataType ConvertDataType(DataType dtype);
+
 }}  // namespace nvidia::inferenceserver

src/servables/tensorflow/savedmodel_bundle.cc

@@ -38,66 +38,6 @@
 
 namespace nvidia { namespace inferenceserver {
 
-namespace {
-
-bool
-CompareDims(
-  const tensorflow::TensorShapeProto& model_shape, const DimsList& dims)
-{
-  // The first model dimension can be -1 to serve as a placeholder for
-  // batch. The batch dim doesn't appear in the configuration 'dims'.
-  const bool has_batch_dim =
-    (model_shape.dim().size() >= 1) && (model_shape.dim(0).size() == -1);
-  if (model_shape.dim().size() != (dims.size() + (has_batch_dim ? 1 : 0))) {
-    return false;
-  }
-
-  for (int i = 0; i < dims.size(); ++i) {
-    if (model_shape.dim(i + (has_batch_dim ? 1 : 0)).size() != dims[i]) {
-      return false;
-    }
-  }
-
-  return true;
-}
-
-const std::string
-DimsDebugString(const DimsList& dims)
-{
-  bool first = true;
-  std::string str;
-  str.append("[");
-  for (int i = 0; i < dims.size(); ++i) {
-    if (!first) {
-      str.append(",");
-    }
-    str.append(std::to_string(dims[i]));
-    first = false;
-  }
-  str.append("]");
-  return str;
-}
-
-const std::string
-DimsDebugString(const tensorflow::TensorShapeProto& dims)
-{
-  bool first = true;
-  std::string str;
-  str.append("[");
-  for (int i = 0; i < dims.dim().size(); ++i) {
-    if (!first) {
-      str.append(",");
-    }
-    str.append(std::to_string(dims.dim(i).size()));
-    first = false;
-  }
-  str.append("]");
-  return str;
-}
-
-
-}  // namespace
-
 tensorflow::Status
 SavedModelBundle::Init(
   const tensorflow::StringPiece& path, const ModelConfig& config)
@@ -208,6 +148,13 @@ SavedModelBundle::CreateSession(
         DimsDebugString(iitr->second.tensor_shape()),
         " don't match configuration dims ", DimsDebugString(io.dims()));
     }
+    if (!CompareDataType(iitr->second.dtype(), io.data_type())) {
+      return tensorflow::errors::InvalidArgument(
+        "unable to load model '", Name(), "', input '", io.name(),
+        "' data-type ", tensorflow::DataType_Name(iitr->second.dtype()),
+        " doesn't match configuration data-type ",
+        DataType_Name(io.data_type()));
+    }
   }
 
   for (const auto& io : Config().output()) {
@@ -225,6 +172,13 @@ SavedModelBundle::CreateSession(
         DimsDebugString(oitr->second.tensor_shape()),
         " don't match configuration dims ", DimsDebugString(io.dims()));
     }
+    if (!CompareDataType(oitr->second.dtype(), io.data_type())) {
+      return tensorflow::errors::InvalidArgument(
+        "unable to load model '", Name(), "', output '", io.name(),
+        "' data-type ", tensorflow::DataType_Name(oitr->second.dtype()),
+        " doesn't match configuration data-type ",
+        DataType_Name(io.data_type()));
+    }
   }
 
   *session = bundle->session.release();

src/servables/tensorflow/savedmodel_bundle_test.cc

@@ -41,7 +41,21 @@ TEST_F(SavedModelBundleTest, ModelConfigSanity)
       const std::string& path,
       const ModelConfig& config) -> tensorflow::Status {
     std::unique_ptr<SavedModelBundle> bundle(new SavedModelBundle());
-    return bundle->Init(path, config);
+    tensorflow::Status status = bundle->Init(path, config);
+    if (status.ok()) {
+      std::unordered_map<std::string, std::string> savedmodel_paths;
+      std::string filename = "model.savedmodel";
+      const auto savedmodel_path = tensorflow::io::JoinPath(path, filename);
+      savedmodel_paths.emplace(
+        std::piecewise_construct, std::make_tuple(filename),
+        std::make_tuple(savedmodel_path));
+
+      tensorflow::ConfigProto session_config;
+      status =
+        bundle->CreateExecutionContexts(session_config, savedmodel_paths);
+    }
+
+    return status;
   };
 
   // Standard testing...

src/servables/tensorflow/testdata/savedmodel_autofill_sanity/bad_input_dims/config.pbtxt

@@ -0,0 +1,25 @@
+max_batch_size: 1
+input [
+  {
+    name: "INPUT0"
+    data_type: TYPE_INT32
+    dims: [ 16, 1 ]
+  },
+  {
+    name: "INPUT1"
+    data_type: TYPE_INT32
+    dims: [ 16 ]
+  }
+]
+output [
+  {
+    name: "OUTPUT0"
+    data_type: TYPE_INT8
+    dims: [ 16 ]
+  },
+  {
+    name: "OUTPUT1"
+    data_type: TYPE_INT8
+    dims: [ 16 ]
+  }
+]

src/servables/tensorflow/testdata/savedmodel_autofill_sanity/bad_input_dims/expected

@@ -0,0 +1 @@
+Invalid argument: unable to load model 'bad_input_dims', input 'INPUT0' dims [-1,16] don't match configuration dims [16,1]

src/servables/tensorflow/testdata/savedmodel_autofill_sanity/bad_input_type/config.pbtxt

@@ -0,0 +1,25 @@
+max_batch_size: 1
+input [
+  {
+    name: "INPUT0"
+    data_type: TYPE_INT32
+    dims: [ 16 ]
+  },
+  {
+    name: "INPUT1"
+    data_type: TYPE_FP32
+    dims: [ 16 ]
+  }
+]
+output [
+  {
+    name: "OUTPUT0"
+    data_type: TYPE_INT8
+    dims: [ 16 ]
+  },
+  {
+    name: "OUTPUT1"
+    data_type: TYPE_INT8
+    dims: [ 16 ]
+  }
+]

src/servables/tensorflow/testdata/savedmodel_autofill_sanity/bad_input_type/expected

@@ -0,0 +1 @@
+Invalid argument: unable to load model 'bad_input_type', input 'INPUT1' data-type DT_INT32 doesn't match configuration data-type TYPE_FP32

src/servables/tensorflow/testdata/savedmodel_autofill_sanity/bad_output_dims/config.pbtxt

@@ -0,0 +1,25 @@
+max_batch_size: 1
+input [
+  {
+    name: "INPUT0"
+    data_type: TYPE_INT32
+    dims: [ 16 ]
+  },
+  {
+    name: "INPUT1"
+    data_type: TYPE_INT32
+    dims: [ 16 ]
+  }
+]
+output [
+  {
+    name: "OUTPUT0"
+    data_type: TYPE_INT8
+    dims: [ 16 ]
+  },
+  {
+    name: "OUTPUT1"
+    data_type: TYPE_INT8
+    dims: [ 1 ]
+  }
+]

src/servables/tensorflow/testdata/savedmodel_autofill_sanity/bad_output_dims/expected

@@ -0,0 +1 @@
+Invalid argument: unable to load model 'bad_output_dims', output 'OUTPUT1' dims [-1,16] don't match configuration dims [1]

src/servables/tensorflow/testdata/savedmodel_autofill_sanity/bad_output_type/config.pbtxt

@@ -0,0 +1,25 @@
+max_batch_size: 1
+input [
+  {
+    name: "INPUT0"
+    data_type: TYPE_INT32
+    dims: [ 16 ]
+  },
+  {
+    name: "INPUT1"
+    data_type: TYPE_INT32
+    dims: [ 16 ]
+  }
+]
+output [
+  {
+    name: "OUTPUT0"
+    data_type: TYPE_INT16
+    dims: [ 16 ]
+  },
+  {
+    name: "OUTPUT1"
+    data_type: TYPE_INT8
+    dims: [ 16 ]
+  }
+]

src/servables/tensorflow/testdata/savedmodel_autofill_sanity/bad_output_type/expected

@@ -0,0 +1 @@
+Invalid argument: unable to load model 'bad_output_type', output 'OUTPUT0' data-type DT_INT8 doesn't match configuration data-type TYPE_INT16