Check that specific inputs are constants in Conv(Add|Mul|BN)Fusion rules (#1270)

* Check for non-existent initializers while fusing conv and add.

* Fix other places where initializer can be null

* Add check if initializer is an input

* update the models to comply with the new ONNX spec.

In new ONNX spec, the initializers should not be in inputs.

* Fix previous temporary code

* Add negative test

* Revert changes to conv_bn_fusion and conv_mul_fusion

* making helper IsNodeArgConstant a little more general; updating remaining Conv*Fusion rules

* minor comment

* AllNodeIputsAreConstant to use new function

onnxruntime/core/graph/graph_utils.cc

@@ -399,16 +399,20 @@ bool IsGraphInput(const Graph& graph, const NodeArg* input) {
   return std::find(graph_inputs.begin(), graph_inputs.end(), input) != graph_inputs.end();
 }
 
+bool NodeArgIsConstant(const Graph& graph, const NodeArg& node_arg) {
+  const onnx::TensorProto* initializer = nullptr;
+  return graph.GetInitializedTensor(node_arg.Name(), initializer) && !IsGraphInput(graph, &node_arg);
+}
+
 bool AllNodeInputsAreConstant(const Graph& graph, const Node& node) {
   if (node.GetInputEdgesCount() > 0) {
     return false;
   }
-  const onnx::TensorProto* initializer = nullptr;
   for (const auto* input_def : node.InputDefs()) {
     // Important note: when an initializer appears in the graph's input, this input will not be considered constant,
     // because it can be overriden by the user at runtime. For constant folding to be applied, the initializer should
     // not appear in the graph's inputs (that is the only way to guarantee it will always be constant).
-    if (!graph.GetInitializedTensor(input_def->Name(), initializer) || IsGraphInput(graph, input_def)) {
+    if (!NodeArgIsConstant(graph, *input_def)) {
       return false;
     }
   }

onnxruntime/core/graph/graph_utils.h

@@ -41,6 +41,9 @@ bool IsGraphInput(const Graph& graph, const NodeArg* input);
 /** Checks if the given node has only constant inputs (initializers). */
 bool AllNodeInputsAreConstant(const Graph& graph, const Node& node);
 
+/** Checks if the given NodeArg is constant, i.e., it appears in the graph's initializers but not in its inputs. */
+bool NodeArgIsConstant(const Graph& graph, const NodeArg& node_arg);
+
 /** Gets the name of the incoming NodeArg with the specified index for the given node. */
 const std::string& GetNodeInputName(const Node& node, int index);
 

onnxruntime/core/optimizer/conv_add_fusion.cc

@@ -16,10 +16,14 @@ Status ConvAddFusion::Apply(Graph& graph, Node& node, RewriteRuleEffect& modifie
   const auto& add_inputs = add_node.InputDefs();
 
   const ONNX_NAMESPACE::TensorProto* conv_W_tensor_proto = nullptr;
-  graph.GetInitializedTensor(conv_inputs[1]->Name(), conv_W_tensor_proto);
+  if (!graph.GetInitializedTensor(conv_inputs[1]->Name(), conv_W_tensor_proto)) {
+    return Status::OK();
+  }
 
   const ONNX_NAMESPACE::TensorProto* add_B_tensor_proto = nullptr;
-  graph.GetInitializedTensor(add_inputs[1]->Name(), add_B_tensor_proto);
+  if (!graph.GetInitializedTensor(add_inputs[1]->Name(), add_B_tensor_proto)) {
+    return Status::OK();
+  }
 
   // Currently, fusion is only supported for float or double data type.
   if (!Initializer::IsSupportedDataType(add_B_tensor_proto) ||
@@ -49,7 +53,9 @@ Status ConvAddFusion::Apply(Graph& graph, Node& node, RewriteRuleEffect& modifie
 
   const ONNX_NAMESPACE::TensorProto* conv_B_tensor_proto = nullptr;
   if (conv_inputs.size() == 3) {
-    graph.GetInitializedTensor(conv_inputs[2]->Name(), conv_B_tensor_proto);
+    if (!graph.GetInitializedTensor(conv_inputs[2]->Name(), conv_B_tensor_proto)) {
+      return Status::OK();
+    }
 
     if (!Initializer::IsSupportedDataType(conv_B_tensor_proto) ||
         conv_B_tensor_proto->data_type() != add_B_tensor_proto->data_type() ||
@@ -114,9 +120,21 @@ bool ConvAddFusion::SatisfyCondition(const Graph& graph, const Node& node) const
   }
 
   const auto& next_node = *node.OutputNodesBegin();
-  return !(!graph_utils::IsSupportedOptypeVersionAndDomain(next_node, "Add", {7}) ||
-           next_node.GetExecutionProviderType() != node.GetExecutionProviderType() ||
-           next_node.GetInputEdgesCount() != 1 || graph.IsNodeOutputsInGraphOutputs(next_node));
+  if (!graph_utils::IsSupportedOptypeVersionAndDomain(next_node, "Add", {7}) ||
+      next_node.GetInputEdgesCount() != 1 || graph.IsNodeOutputsInGraphOutputs(next_node) ||
+      // Make sure the two nodes do not span execution providers.
+      next_node.GetExecutionProviderType() != node.GetExecutionProviderType()) {
+    return false;
+  }
+
+  // Check that the appropriate inputs to the Conv and Add nodes are constants.
+  if (!graph_utils::NodeArgIsConstant(graph, *node.InputDefs()[1]) ||
+      (node.InputDefs().size() == 3 && !graph_utils::NodeArgIsConstant(graph, *node.InputDefs()[2])) ||
+      !graph_utils::NodeArgIsConstant(graph, *next_node.InputDefs()[1])) {
+    return false;
+  }
+
+  return true;
 }
 
 }  // namespace onnxruntime

onnxruntime/core/optimizer/conv_bn_fusion.cc

@@ -24,20 +24,30 @@ Status ConvBNFusion::Apply(Graph& graph, Node& node, RewriteRuleEffect& rule_eff
   // Get initializers of BatchNormalization
   const auto& bn_inputs = bn_node.InputDefs();
   const ONNX_NAMESPACE::TensorProto* bn_scale_tensor_proto = nullptr;
-  graph.GetInitializedTensor(bn_inputs[1]->Name(), bn_scale_tensor_proto);
+  if (!graph.GetInitializedTensor(bn_inputs[1]->Name(), bn_scale_tensor_proto)) {
+    return Status::OK();
+  }
 
   const ONNX_NAMESPACE::TensorProto* bn_B_tensor_proto = nullptr;
-  graph.GetInitializedTensor(bn_inputs[2]->Name(), bn_B_tensor_proto);
+  if (!graph.GetInitializedTensor(bn_inputs[2]->Name(), bn_B_tensor_proto)) {
+    return Status::OK();
+  }
 
   const ONNX_NAMESPACE::TensorProto* bn_mean_tensor_proto = nullptr;
-  graph.GetInitializedTensor(bn_inputs[3]->Name(), bn_mean_tensor_proto);
+  if (!graph.GetInitializedTensor(bn_inputs[3]->Name(), bn_mean_tensor_proto)) {
+    return Status::OK();
+  }
 
   const ONNX_NAMESPACE::TensorProto* bn_var_tensor_proto = nullptr;
-  graph.GetInitializedTensor(bn_inputs[4]->Name(), bn_var_tensor_proto);
+  if (!graph.GetInitializedTensor(bn_inputs[4]->Name(), bn_var_tensor_proto)) {
+    return Status::OK();
+  }
 
   const auto& conv_inputs = conv_node.InputDefs();
   const ONNX_NAMESPACE::TensorProto* conv_W_tensor_proto = nullptr;
-  graph.GetInitializedTensor(conv_inputs[1]->Name(), conv_W_tensor_proto);
+  if (!graph.GetInitializedTensor(conv_inputs[1]->Name(), conv_W_tensor_proto)) {
+    return Status::OK();
+  }
 
   // Currently, fusion is only supported for float or double data type.
   if (!Initializer::IsSupportedDataType(bn_scale_tensor_proto) ||
@@ -149,9 +159,24 @@ bool ConvBNFusion::SatisfyCondition(const Graph& graph, const Node& node) const
   }
 
   const auto& next_node = *node.OutputNodesBegin();
-  return !(!graph_utils::IsSupportedOptypeVersionAndDomain(next_node, "BatchNormalization", {7, 9}) ||
-           next_node.GetInputEdgesCount() != 1 || graph.IsNodeOutputsInGraphOutputs(next_node) ||
-           next_node.GetExecutionProviderType() != node.GetExecutionProviderType());
+  if (!graph_utils::IsSupportedOptypeVersionAndDomain(next_node, "BatchNormalization", {7, 9}) ||
+      next_node.GetInputEdgesCount() != 1 || graph.IsNodeOutputsInGraphOutputs(next_node) ||
+      // Make sure the two nodes do not span execution providers.
+      next_node.GetExecutionProviderType() != node.GetExecutionProviderType()) {
+    return false;
+  }
+
+  // Check that the appropriate inputs to the Conv and BN nodes are constants.
+  if (!graph_utils::NodeArgIsConstant(graph, *node.InputDefs()[1]) ||
+      (node.InputDefs().size() == 3 && !graph_utils::NodeArgIsConstant(graph, *node.InputDefs()[2])) ||
+      !graph_utils::NodeArgIsConstant(graph, *next_node.InputDefs()[1]) ||
+      !graph_utils::NodeArgIsConstant(graph, *next_node.InputDefs()[2]) ||
+      !graph_utils::NodeArgIsConstant(graph, *next_node.InputDefs()[3]) ||
+      !graph_utils::NodeArgIsConstant(graph, *next_node.InputDefs()[4])) {
+    return false;
+  }
+
+  return true;
 }
 
 }  // namespace onnxruntime

onnxruntime/core/optimizer/conv_mul_fusion.cc

@@ -16,10 +16,14 @@ Status ConvMulFusion::Apply(Graph& graph, Node& node, RewriteRuleEffect& rule_ef
   const auto& mul_inputs = mul_node.InputDefs();
 
   const ONNX_NAMESPACE::TensorProto* conv_W_tensor_proto = nullptr;
-  graph.GetInitializedTensor(conv_inputs[1]->Name(), conv_W_tensor_proto);
+  if (!graph.GetInitializedTensor(conv_inputs[1]->Name(), conv_W_tensor_proto)) {
+    return Status::OK();
+  }
 
   const ONNX_NAMESPACE::TensorProto* mul_B_tensor_proto = nullptr;
-  graph.GetInitializedTensor(mul_inputs[1]->Name(), mul_B_tensor_proto);
+  if (!graph.GetInitializedTensor(mul_inputs[1]->Name(), mul_B_tensor_proto)) {
+    return Status::OK();
+  }
 
   if (!Initializer::IsSupportedDataType(conv_W_tensor_proto) ||
       !Initializer::IsSupportedDataType(mul_B_tensor_proto) ||
@@ -112,9 +116,21 @@ bool ConvMulFusion::SatisfyCondition(const Graph& graph, const Node& node) const
   }
 
   const auto& next_node = *node.OutputNodesBegin();
-  return !(!graph_utils::IsSupportedOptypeVersionAndDomain(next_node, "Mul", {7}) ||
-           next_node.GetInputEdgesCount() != 1 || graph.IsNodeOutputsInGraphOutputs(next_node) ||
-           next_node.GetExecutionProviderType() != node.GetExecutionProviderType());
+  if (!graph_utils::IsSupportedOptypeVersionAndDomain(next_node, "Mul", {7}) ||
+      next_node.GetInputEdgesCount() != 1 || graph.IsNodeOutputsInGraphOutputs(next_node) ||
+      // Make sure the two nodes do not span execution providers.
+      next_node.GetExecutionProviderType() != node.GetExecutionProviderType()) {
+    return false;
+  }
+
+  // Check that the appropriate inputs to the Conv and Mul nodels are constants.
+  if (!graph_utils::NodeArgIsConstant(graph, *node.InputDefs()[1]) ||
+      (node.InputDefs().size() == 3 && !graph_utils::NodeArgIsConstant(graph, *node.InputDefs()[2])) ||
+      !graph_utils::NodeArgIsConstant(graph, *next_node.InputDefs()[1])) {
+    return false;
+  }
+
+  return true;
 }
 
 }  // namespace onnxruntime

onnxruntime/core/session/inference_session.h

@@ -434,7 +434,7 @@ class InferenceSession {
   //So its lifetime should be same as its constituents. This vector is to extend the lifetime of the owner.
   std::vector<std::shared_ptr<CustomRegistry>> custom_registries_;
 
-#ifdef ENABLE_LANGUAGE_INTEROP_OPS 
+#ifdef ENABLE_LANGUAGE_INTEROP_OPS
   InterOpDomains interop_domains_;
 #endif
 };

onnxruntime/test/onnx/main.cc

@@ -164,7 +164,7 @@ int real_main(int argc, char* argv[], Ort::Env& env) {
         case 'o':
           graph_optimization_level = static_cast<uint32_t>(OrtStrtol<PATH_CHAR_TYPE>(optarg, nullptr));
           if (graph_optimization_level > 2) {
-            fprintf(stderr, "See usage for valid values of graph optimization level");
+            fprintf(stderr, "See usage for valid values of graph optimization level\n");
             usage();
             return -1;
           }

onnxruntime/test/optimizer/graph_transform_test.cc

@@ -286,6 +286,31 @@ TEST(GraphTransformationTests, FuseConvAddNoBias) {
   ASSERT_TRUE(op_to_count["Unsqueeze"] == 0);
 }
 
+TEST(GraphTransformationTests, NegativeFuseConvAddNoBias) {
+  string model_uri = MODEL_FOLDER + "fusion/negative-fuse-conv-add-no-bias.onnx";
+
+  std::shared_ptr<Model> p_model;
+  ASSERT_TRUE(Model::Load(model_uri, p_model).IsOK());
+  Graph& graph = p_model->MainGraph();
+
+  onnxruntime::GraphTransformerManager graph_transformation_mgr{5};
+  auto rule_transformer_L1 = std::make_unique<RuleBasedGraphTransformer>("RuleTransformer1");
+  rule_transformer_L1->Register(std::make_unique<UnsqueezeElimination>());
+  graph_transformation_mgr.Register(std::move(rule_transformer_L1), TransformerLevel::Level1);
+
+  auto rule_transformer_L2 = std::make_unique<RuleBasedGraphTransformer>("RuleTransformerL2");
+  rule_transformer_L2->Register(std::make_unique<ConvAddFusion>());
+  graph_transformation_mgr.Register(std::move(rule_transformer_L2), TransformerLevel::Level2);
+
+  ASSERT_TRUE(graph_transformation_mgr.ApplyTransformers(graph, TransformerLevel::Level1).IsOK());
+  ASSERT_TRUE(graph_transformation_mgr.ApplyTransformers(graph, TransformerLevel::Level2).IsOK());
+
+  // Nodes are not fused because the weights to conv/add are not constants (they appear in the graph inputs).
+  std::map<std::string, int> op_to_count = CountOpsInGraph(graph);
+  ASSERT_TRUE(op_to_count["Add"] != 0);
+  ASSERT_TRUE(op_to_count["Unsqueeze"] == 0);
+}
+
 TEST(GraphTransformationTests, FuseConvAddMul3D) {
   string model_uri = MODEL_FOLDER + "fusion/fuse-conv-add-mul-3d.onnx";