[TF FE] Support SparseTensorDenseMatMul operation (#26064)

**Details:** Support SparseTensorDenseMatMul operation. Required for
customer model.

**Ticket:** 104539

---------

Signed-off-by: Kazantsev, Roman <[email protected]>

src/frontends/tensorflow/docs/supported_ops.md

@@ -1150,7 +1150,7 @@ A "supported operation" is one that TensorFlow Frontend can convert to the OpenV
 | SparseSparseMinimum                                     | NO                            |                               |
 | SparseSplit                                             | NO                            |                               |
 | SparseTensorDenseAdd                                    | NO                            |                               |
-| SparseTensorDenseMatMul                                 | NO                            |                               |
+| SparseTensorDenseMatMul                                 | YES                           |                               |
 | SparseTensorSliceDataset                                | NO                            |                               |
 | SparseTensorToCSRSparseMatrix                           | NO                            |                               |
 | SparseToDense                                           | YES                           |                               |

src/frontends/tensorflow/src/op_table.cpp

@@ -370,7 +370,6 @@ const std::map<std::string, CreatorFunction> get_supported_ops() {
         {"SaveV2", CreatorFunction(translate_no_op)},
         {"ScatterNd", CreatorFunction(translate_scatter_nd_op)},
         {"SegmentSum", CreatorFunction(translate_segment_sum_op)},
-        {"SparseToDense", CreatorFunction(translate_sparse_to_dense_op)},
         {"Select", CreatorFunction(translate_select_op)},
         {"SelectV2", CreatorFunction(translate_select_v2_op)},
         {"Shape", CreatorFunction(translate_shape_op)},
@@ -381,6 +380,8 @@ const std::map<std::string, CreatorFunction> get_supported_ops() {
         {"Softmax", CreatorFunction(translate_softmax_op)},
         {"SpaceToDepth", CreatorFunction(translate_space_to_depth_op)},
         {"SparseReshape", CreatorFunction(translate_sparse_reshape_op)},
+        {"SparseTensorDenseMatMul", CreatorFunction(translate_sparse_tensor_dense_mat_mul_op)},
+        {"SparseToDense", CreatorFunction(translate_sparse_to_dense_op)},
         {"Split", CreatorFunction(translate_split_op)},
         {"SplitV", CreatorFunction(translate_split_v_op)},
         {"StopGradient", CreatorFunction(translate_identity_op)},

src/frontends/tensorflow_common/include/common_op_table.hpp

@@ -146,6 +146,7 @@ OP_CONVERTER(translate_rsqrt_op);
 OP_CONVERTER(translate_scatter_nd_op);
 OP_CONVERTER(translate_segment_sum_op);
 OP_CONVERTER(translate_space_to_batch_nd_op);
+OP_CONVERTER(translate_sparse_tensor_dense_mat_mul_op);
 OP_CONVERTER(translate_sparse_to_dense_op);
 OP_CONVERTER(translate_select_op);
 OP_CONVERTER(translate_select_v2_op);

src/frontends/tensorflow_common/src/op/sparse_tensor_dense_mat_mul.cpp

@@ -0,0 +1,39 @@
+// Copyright (C) 2018-2024 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "common_op_table.hpp"
+#include "openvino/op/broadcast.hpp"
+#include "openvino/op/matmul.hpp"
+#include "openvino/op/scatter_nd_update.hpp"
+#include "utils.hpp"
+
+using namespace std;
+using namespace ov::op;
+
+namespace ov {
+namespace frontend {
+namespace tensorflow {
+namespace op {
+OutputVector translate_sparse_tensor_dense_mat_mul_op(const NodeContext& node) {
+    default_op_checks(node, 4, {"SparseTensorDenseMatMul"});
+    auto a_indices = node.get_input(0);
+    auto a_values = node.get_input(1);
+    auto a_shape = node.get_input(2);
+    auto b = node.get_input(3);
+    auto adjoint_a = node.get_attribute<bool>("adjoint_a", false);
+    auto adjoint_b = node.get_attribute<bool>("adjoint_b", false);
+
+    // create dense tensor
+    auto zero_const = create_same_type_const_scalar<int32_t>(a_values, 0);
+    ov::Output<ov::Node> a = make_shared<v3::Broadcast>(zero_const, a_shape);
+    a = make_shared<v15::ScatterNDUpdate>(a, a_indices, a_values);
+    auto res = make_shared<v0::MatMul>(a, b, adjoint_a, adjoint_b);
+    set_node_name(node.get_name(), res);
+    return {res};
+}
+
+}  // namespace op
+}  // namespace tensorflow
+}  // namespace frontend
+}  // namespace ov

tests/layer_tests/tensorflow_tests/test_tf_SparseTensorDenseMatMul.py

@@ -0,0 +1,100 @@
+# Copyright (C) 2018-2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+import numpy as np
+import platform
+import pytest
+import tensorflow as tf
+from common.tf_layer_test_class import CommonTFLayerTest
+
+rng = np.random.default_rng(475912)
+
+
+class TestSparseTensorDenseMatMul(CommonTFLayerTest):
+    def _prepare_input(self, inputs_info):
+        assert 'a_indices:0' in inputs_info
+        assert 'a_values:0' in inputs_info
+        assert 'b:0' in inputs_info
+
+        a_values_shape = inputs_info['a_values:0']
+        b_shape = inputs_info['b:0']
+
+        inputs_data = {}
+        if np.issubdtype(self.data_type, np.floating):
+            inputs_data['a_values:0'] = rng.uniform(-5.0, 5.0, a_values_shape).astype(self.data_type)
+            inputs_data['b:0'] = rng.uniform(-5.0, 5.0, b_shape).astype(self.data_type)
+        elif np.issubdtype(self.data_type, np.signedinteger):
+            inputs_data['a_values:0'] = rng.integers(-8, 8, a_values_shape).astype(self.data_type)
+            inputs_data['b:0'] = rng.integers(-8, 8, b_shape).astype(self.data_type)
+        else:
+            inputs_data['a_values:0'] = rng.integers(0, 8, a_values_shape).astype(self.data_type)
+            inputs_data['b:0'] = rng.integers(0, 8, b_shape).astype(self.data_type)
+
+        a_rows_num = self.a_shape[0]
+        a_cols_num = self.a_shape[1]
+
+        # generate all possible indices
+        all_indices = []
+        for row_ind in range(0, a_rows_num):
+            for col_ind in range(0, a_cols_num):
+                all_indices.append([row_ind, col_ind])
+        inputs_data['a_indices:0'] = rng.choice(all_indices, self.nnz, replace=False).astype(self.indices_type)
+
+        return inputs_data
+
+    def create_sparse_tensor_dense_mat_mul_net(self, data_type, indices_type,
+                                               adjoint_a, adjoint_b,
+                                               a_shape, b_shape, nnz):
+        a_shape = a_shape.copy()
+        b_shape = b_shape.copy()
+        if adjoint_a:
+            a_shape.reverse()
+        if adjoint_b:
+            b_shape.reverse()
+
+        self.data_type = data_type
+        self.indices_type = indices_type
+        self.a_shape = a_shape
+        self.nnz = nnz
+        tf.compat.v1.reset_default_graph()
+        with tf.compat.v1.Session() as sess:
+            a_indices = tf.compat.v1.placeholder(indices_type, [nnz, 2], 'a_indices')
+            a_values = tf.compat.v1.placeholder(data_type, [nnz], 'a_values')
+            a_shape = tf.constant(a_shape, dtype=tf.int64)
+            b = tf.compat.v1.placeholder(data_type, b_shape, 'b')
+            tf.raw_ops.SparseTensorDenseMatMul(
+                a_indices=a_indices,
+                a_values=a_values,
+                a_shape=a_shape,
+                b=b,
+                adjoint_a=adjoint_a,
+                adjoint_b=adjoint_b)
+            tf.compat.v1.global_variables_initializer()
+            tf_net = sess.graph_def
+
+        ref_net = None
+
+        return tf_net, ref_net
+
+    @pytest.mark.parametrize('data_type', [np.float32, np.float64, np.int32])
+    @pytest.mark.parametrize('indices_type', [np.int32, np.int64])
+    @pytest.mark.parametrize('adjoint_a', [True, False])
+    @pytest.mark.parametrize('adjoint_b', [True, False])
+    @pytest.mark.parametrize('a_shape, b_shape, nnz', [
+        [[4, 10], [10, 5], 8],
+        [[5, 5], [5, 5], 3],
+    ])
+    @pytest.mark.precommit
+    @pytest.mark.nightly
+    def test_sparse_tensor_dense_mat_mul(self, data_type, indices_type,
+                                         adjoint_a, adjoint_b,
+                                         a_shape, b_shape, nnz,
+                                         ie_device, precision, ir_version, temp_dir,
+                                         use_legacy_frontend):
+        if ie_device == 'GPU':
+            pytest.skip("149830: ScatterNDUpdate-15 is not supported on GPU")
+        self._test(*self.create_sparse_tensor_dense_mat_mul_net(data_type, indices_type,
+                                                                adjoint_a, adjoint_b,
+                                                                a_shape, b_shape, nnz),
+                   ie_device, precision, ir_version, temp_dir=temp_dir,
+                   use_legacy_frontend=use_legacy_frontend)