Implement SoftmaxCrossEntropyLoss for ONNX frontend (opset 12 & 13)

src/frontends/onnx/frontend/src/op/softmax_cossentropy_loss.cpp

@@ -0,0 +1,114 @@
+// Copyright (C) 2018-2025 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "core/operator_set.hpp"
+#include "openvino/op/convert.hpp"
+#include "openvino/op/divide.hpp"
+#include "openvino/op/gather.hpp"
+#include "openvino/op/log.hpp"
+#include "openvino/op/multiply.hpp"
+#include "openvino/op/negative.hpp"
+#include "openvino/op/not_equal.hpp"
+#include "openvino/op/reduce_mean.hpp"
+#include "openvino/op/reduce_sum.hpp"
+#include "openvino/op/softmax.hpp"
+
+#include "softmax_crossentropy_loss.hpp"
+
+using namespace ov::op;
+
+namespace ov {
+namespace frontend {
+namespace onnx {
+namespace {
+OutputVector onnx_softmax_crossentropy_loss(const Node& node, int64_t axis_default) {
+    const auto inputs = node.get_ov_inputs();
+
+    const auto scores = inputs[0];
+    const auto labels = inputs[1];
+    bool has_weights = inputs.size() > 2;
+    std::shared_ptr<ov::Node> weights_gather = nullptr;
+
+    bool has_ignore_index = node.has_attribute("ignore_index");
+    int64_t ignore_index_val = 0;
+    std::shared_ptr<ov::Node> mask = nullptr;
+
+    if (has_ignore_index) {
+        ignore_index_val = node.get_attribute_value<int64_t>("ignore_index");
+        auto ignore_index_node = ::v0::Constant::create(labels.get_element_type(), {}, {ignore_index_val});
+        auto neq = std::make_shared<::v1::NotEqual>(labels, ignore_index_node);
+        mask = std::make_shared<::v0::Convert>(neq, scores.get_element_type());
+    }
+
+    if (has_weights) {
+        const auto weights = inputs[2];
+        const auto axis_for_weights = ::v0::Constant::create(element::i64, {}, {0});
+        weights_gather = std::make_shared<::v8::Gather>(weights, labels, axis_for_weights);
+
+        if (has_ignore_index) {
+            weights_gather = std::make_shared<::v1::Multiply>(weights_gather, mask);
+        }
+    } else if (has_ignore_index) {
+        weights_gather = mask;
+    }
+
+    const auto axis = node.get_attribute_value<int64_t>("axis", axis_default);
+    const auto reduction = node.get_attribute_value<std::string>("reduction", "mean");
+
+    const auto softmax = std::make_shared<::v8::Softmax>(scores, axis);
+    const auto log_softmax = std::make_shared<::v0::Log>(softmax);
+
+    const auto axis_const = ::v0::Constant::create(element::i64, {}, {axis});
+    const auto gathered = std::make_shared<::v8::Gather>(log_softmax, labels, axis_const);
+
+    std::shared_ptr<ov::Node> loss = std::make_shared<::v0::Negative>(gathered);
+
+    if (weights_gather) {
+        loss = std::make_shared<::v1::Multiply>(loss, weights_gather);
+    }
+
+    if (reduction != "none") {
+        auto loss_shape = loss->get_output_partial_shape(0);
+        if (loss_shape.rank().is_static()) {
+            size_t loss_rank = loss_shape.rank().get_length();
+            std::vector<int64_t> reduce_axes(loss_rank);
+            std::iota(reduce_axes.begin(), reduce_axes.end(), 0);
+            auto reduce_axis = ::v0::Constant::create(ov::element::i64, {reduce_axes.size()}, reduce_axes);
+
+            if (reduction == "mean") {
+                if (weights_gather) {
+                    auto loss_sum = std::make_shared<::v1::ReduceSum>(loss, reduce_axis, false);
+                    auto weight_sum = std::make_shared<::v1::ReduceSum>(weights_gather, reduce_axis, false);
+                    loss = std::make_shared<::v1::Divide>(loss_sum, weight_sum);
+                } else {
+                    loss = std::make_shared<::v1::ReduceMean>(loss, reduce_axis, false);
+                }
+            } else if (reduction == "sum") {
+                loss = std::make_shared<::v1::ReduceSum>(loss, reduce_axis, false);
+            }
+        } else {
+            OPENVINO_THROW("Dynamic rank is not supported for SoftmaxCrossEntropyLoss reduction.");
+        }
+    }
+
+    return {loss};
+}
+}  // namespace
+namespace ai_onnx {
+namespace opset_12 {
+OutputVector softmax_cross_entropy_loss(const Node& node) {
+    return onnx_softmax_crossentropy_loss(node, 1);
+}
+ONNX_OP("SoftmaxCrossEntropyLoss", OPSET_RANGE(1, 12), ai_onnx::opset_12::softmax_cross_entropy_loss);
+}  // namespace opset_12
+namespace opset_13 {
+OutputVector softmax_cross_entropy_loss(const Node& node) {
+    return onnx_softmax_crossentropy_loss(node, 1);
+}
+ONNX_OP("SoftmaxCrossEntropyLoss", OPSET_SINCE(13), ai_onnx::opset_13::softmax_cross_entropy_loss);
+}  // namespace opset_13
+}  // namespace ai_onnx
+}  // namespace onnx
+}  // namespace frontend
+}  // namespace ov

src/frontends/onnx/frontend/src/op/softmax_crossentropy_loss.hpp

@@ -0,0 +1,22 @@
+// Copyright (C) 2018-2025 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#pragma once
+
+#include "openvino/op/op.hpp"
+
+namespace ov {
+namespace frontend {
+namespace onnx {
+namespace ai_onnx {
+namespace opset_12 {
+OutputVector softmax_cross_entropy_loss(const Node& node);
+}  // namespace opset_12
+namespace opset_13 {
+OutputVector softmax_cross_entropy_loss(const Node& node);
+}  // namespace opset_13
+}  // namespace ai_onnx
+}  // namespace onnx
+}  // namespace frontend
+}  // namespace ov

src/frontends/onnx/tests/models/softmax_crossentropy_loss_sum.prototxt

@@ -0,0 +1,51 @@
+ir_version: 7
+producer_name: "OpenVINO ONNX Frontend"
+graph {
+  node {
+    input: "x"
+    input: "y"
+    output: "z"
+    op_type: "SoftmaxCrossEntropyLoss"
+    attribute {
+      name: "reduction"
+      s: "sum"
+      type: STRING
+    }
+  }
+  name: "test_sce_sum"
+  input {
+    name: "x"
+    type {
+      tensor_type {
+        elem_type: 1
+        shape {
+          dim { dim_value: 2 }
+          dim { dim_value: 3 }
+        }
+      }
+    }
+  }
+  input {
+    name: "y"
+    type {
+      tensor_type {
+        elem_type: 7
+        shape {
+          dim { dim_value: 2 }
+        }
+      }
+    }
+  }
+  output {
+    name: "z"
+    type {
+      tensor_type {
+        elem_type: 1
+        shape {}
+      }
+    }
+  }
+}
+opset_import {
+  version: 13
+}

src/frontends/onnx/tests/onnx_import.in.cpp

@@ -6945,3 +6945,13 @@ OPENVINO_TEST(${BACKEND_NAME}, onnx_float8e4m3fn_constant) {
 
     test_case.run();
 }
+
+OPENVINO_TEST(${BACKEND_NAME}, onnx_model_softmax_crossentropy_loss_sum) {
+    auto model = convert_model("softmax_crossentropy_loss_sum.onnx");
+    auto test_case = ov::test::TestCase(model, s_device);
+    test_case.add_input<float>({0.5f, 1.5f, 2.5f, 3.5f, 4.5f, 5.5f});
+    test_case.add_input<int64_t>({1, 2});
+    test_case.add_expected_output<float>(Shape{}, {1.81521f});
+
+    test_case.run_with_tolerance_as_fp(0.0015f);
+}