Add of ReduceMax Gradient

orttraining/orttraining/core/graph/gradient_builder.cc

@@ -2253,6 +2253,66 @@ IMPLEMENT_GRADIENT_BUILDER(GetGlobalMaxPoolGradient) {
 
   result.push_back(NodeDef("Expand", {GO(0), IA("X_shape")}, {IA("expanded_dY")}));
   result.push_back(NodeDef("Mul", {IA("mask_cast"), IA("expanded_dY")}, {GI(0)}));
+  return result;
+}
+
+IMPLEMENT_GRADIENT_BUILDER(GetReduceMaxGradient) {
+  std::vector<NodeDef> result;
+  auto attributes = SrcNodeAttributes();
+  bool keepdims = true;
+
+  // Check the "keepdims" attribute
+  if (attributes.find("keepdims") != attributes.end() &&
+      attributes.at("keepdims").has_i()) {
+    keepdims = static_cast<bool>(attributes.at("keepdims").i());
+  }
+
+  ArgDef grad = GO(0);
+  ArgDef reduced_output = O(0);
+
+  if (!keepdims) {
+    size_t numInputs = GetSrcNodeInputSize();
+    ArgDef unsqueeze_axes_arg;
+    bool axes_provided = false;
+
+    // Handle "axes" as attribute or input
+    if (attributes.find("axes") != attributes.end()) {
+      axes_provided = true;
+      std::vector<int64_t> axes_values = RetrieveValues<int64_t>(attributes.at("axes"));
+      if (SrcNodeOpsetVersion() >= 13) {
+        NodeDef axes_values_node = ConstantVectorNode(axes_values, Name("axes_values"));
+        result.push_back(axes_values_node);
+        unsqueeze_axes_arg = axes_values_node.output_args[0];
+      }
+    } else if (numInputs == 2) {
+      axes_provided = true;
+      unsqueeze_axes_arg = I(1);
+    }
+
+    if (axes_provided) {
+      grad = IA("Unsqueezed_Grad");
+      reduced_output = IA("Unsqueezed_Output");
+      if (SrcNodeOpsetVersion() < 13 && attributes.find("axes") != attributes.end()) {
+        std::vector<int64_t> axes_values = RetrieveValues<int64_t>(attributes.at("axes"));
+        result.push_back(NodeDef("Unsqueeze", {GO(0)}, {grad}, {MakeAttribute("axes", axes_values)}));
+        result.push_back(NodeDef("Unsqueeze", {O(0)}, {reduced_output}, {MakeAttribute("axes", axes_values)}));
+      } else {
+        result.push_back(NodeDef(OpDef{"Unsqueeze", kOnnxDomain, 13}, {GO(0), unsqueeze_axes_arg}, {grad}));
+        result.push_back(NodeDef(OpDef{"Unsqueeze", kOnnxDomain, 13}, {O(0), unsqueeze_axes_arg}, {reduced_output}));
+      }
+    }
+  }
+
+  // Step 1: Recreate the boolean mask tensor indicating max positions
+  result.push_back(NodeDef("Shape", {I(0)}, {IA("Shaped_X")}));
+  result.push_back(NodeDef("Expand", {reduced_output, IA("Shaped_X")}, {IA("Expanded_Output")}));
+  result.push_back(NodeDef("Equal", {I(0), IA("Expanded_Output")}, {IA("Mask")}));
+  // Step 2: Convert the boolean mask to a float tensor (0.0 and 1.0)
+  result.push_back(NodeDef("Cast", {IA("Mask")}, {IA("Mask_Float")}, {MakeAttribute("to", static_cast<int64_t>(OElemType(0)))}));
+  // Step 3: Multiply the input gradient by the mask
+  result.push_back(NodeDef("Mul", {grad, IA("Mask_Float")}, {IA("Masked_Grad")}));
+  // Step 4: Ensure the output gradient has the same shape as the input
+  result.push_back(NodeDef("Expand", {IA("Masked_Grad"), IA("Shaped_X")}, {GI(0)}));
 
   return result;
 }

orttraining/orttraining/core/graph/gradient_builder.h

@@ -95,6 +95,7 @@ DECLARE_GRADIENT_BUILDER(GetConvTransposeGradient)
 DECLARE_GRADIENT_BUILDER(GetResizeGradient)
 DECLARE_GRADIENT_BUILDER(GetAtanGradient)
 DECLARE_GRADIENT_BUILDER(GetGlobalMaxPoolGradient)
+DECLARE_GRADIENT_BUILDER(GetReduceMaxGradient)
 
 DECLARE_GRADIENT_BUILDER(GetExternalGradient)
 

orttraining/orttraining/core/graph/gradient_builder_registry.cc

@@ -127,6 +127,7 @@ void GradientBuilderRegistry::RegisterGradientBuilders() {
   REGISTER_GRADIENT_BUILDER("Resize", GetResizeGradient);
   REGISTER_GRADIENT_BUILDER("Atan", GetAtanGradient);
   REGISTER_GRADIENT_BUILDER("GlobalMaxPool", GetGlobalMaxPoolGradient);
+  REGISTER_GRADIENT_BUILDER("ReduceMax", GetReduceMaxGradient);
 
   REGISTER_GRADIENT_BUILDER("ExternalGradient", GetExternalGradient);
 };

orttraining/orttraining/test/gradient/gradient_ops_test.cc

@@ -3379,6 +3379,30 @@ TEST(GradientCheckerTest, GlobalMaxPoolGrad) {
   }
 }
 
+TEST(GradientCheckerTest, ReduceMaxGrad) {
+  // Attribute axes supports negative values from opset 11.
+  OpDef op_def_11{"ReduceMax", kOnnxDomain, 11};
+
+  RunReductionTests(op_def_11, false, true);
+
+  OpDef op_def_12{"ReduceMax", kOnnxDomain, 12};
+
+  RunReductionTests(op_def_12, false, true);
+
+  OpDef op_def_13{"ReduceMax", kOnnxDomain, 13};
+
+  RunReductionTests(op_def_13, false, true);
+
+  // axes is input from opset 18.
+  OpDef op_def_18{"ReduceMax", kOnnxDomain, 18};
+
+  RunReductionTests(op_def_18, true, true);
+
+  OpDef op_def_20{"ReduceMax", kOnnxDomain, 20};
+
+  RunReductionTests(op_def_20, true, true);
+}
+
 }  // namespace test
 }  // namespace onnxruntime