[TF FE] MatrixBandPart operation for TensorFlow Hub models

src/frontends/tensorflow/docs/supported_ops.md

@@ -620,7 +620,7 @@ A "supported operation" is one that TensorFlow Frontend can convert to the OpenV
 | MatMul                                                  | YES                           |                               |
 | MatchingFiles                                           | NO                            |                               |
 | MatchingFilesDataset                                    | NO                            |                               |
-| MatrixBandPart                                          | NO                            |                               |
+| MatrixBandPart                                          | YES                            |                               |
 | MatrixDeterminant                                       | NO                            |                               |
 | MatrixDiag                                              | YES                           |                               |
 | MatrixDiagPart                                          | NO                            |                               |

src/frontends/tensorflow/src/op_table.cpp

@@ -298,6 +298,7 @@ const std::map<std::string, CreatorFunction> get_supported_ops() {
         {"LookupTableInsertV2", CreatorFunction(translate_no_op)},
         {"LRN", CreatorFunction(translate_lrn_op)},
         {"MatMul", CreatorFunction(translate_mat_mul_op)},
+        {"MatrixBandPart", CreatorFunction(translate_matrix_band_part_op)},
         {"MatrixDiag", CreatorFunction(translate_matrix_diag_op)},
         {"MaxPool", CreatorFunction(translate_max_pool_op)},
         {"MaxPoolV2", CreatorFunction(translate_max_pool_op)},

src/frontends/tensorflow_common/include/common_op_table.hpp

@@ -102,6 +102,7 @@ OP_CONVERTER(translate_log_1p_op);
 OP_CONVERTER(translate_lrn_op);
 OP_CONVERTER(translate_mat_mul_op);
 OP_CONVERTER(translate_matrix_diag_op);
+OP_CONVERTER(translate_matrix_band_part_op);
 OP_CONVERTER(translate_max_pool_op);
 OP_CONVERTER_NAMED(translate_max_pool_with_argmax);
 OP_CONVERTER(translate_mirror_pad_op);

src/frontends/tensorflow_common/src/op/matrix_band_part.cpp

@@ -0,0 +1,90 @@
+// Copyright (C) 2024 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+
+#include "openvino/op/constant.hpp"
+#include "openvino/op/convert.hpp"
+#include "openvino/op/less.hpp"
+#include "openvino/op/less_eq.hpp"
+#include "openvino/op/logical_and.hpp"
+#include "openvino/op/logical_or.hpp"
+#include "openvino/op/range.hpp"
+#include "openvino/op/reshape.hpp"
+#include "openvino/op/select.hpp"
+#include "openvino/op/shape_of.hpp"
+#include "openvino/op/slice.hpp"
+#include "openvino/op/subtract.hpp"
+#include "openvino/op/unsqueeze.hpp"
+#include "utils.hpp"
+
+using namespace std;
+using namespace ov::op;
+
+namespace ov {
+namespace frontend {
+namespace tensorflow {
+namespace op {
+
+OutputVector translate_matrix_band_part_op(const NodeContext& node) {
+    default_op_checks(node, 3, {"MatrixBandPart"});
+
+    // Input tensor and parameters
+    auto input = node.get_input(0);
+    auto num_lower = node.get_input(1);
+    auto num_upper = node.get_input(2);
+
+    // create scalar auxiliary constants
+    auto const_zero = make_shared<v0::Constant>(element::i64, Shape{}, 0);
+    auto const_one = make_shared<v0::Constant>(element::i64, Shape{}, 1);
+    auto const_two = make_shared<v0::Constant>(element::i64, Shape{}, 2);
+
+    // input has a shape [I, J, K, ..., M, N]
+    // compute sizes of two last dimensions of M and N
+    auto input_shape = make_shared<v3::ShapeOf>(input, element::i64);
+    auto input_rank = make_shared<v3::ShapeOf>(input_shape, element::i64);
+    auto input_rank_minus_one = make_shared<v1::Subtract>(input_rank, const_one);
+    auto input_rank_minus_two = make_shared<v1::Subtract>(input_rank, const_two);
+    auto slice_step = make_shared<v0::Constant>(element::i64, Shape{1}, 1);
+    auto slice_axis = make_shared<v0::Constant>(element::i64, Shape{1}, 0);
+    auto m = make_shared<v8::Slice>(input_shape, input_rank_minus_two, input_rank_minus_one, slice_step, slice_axis)
+                 ->output(0);
+    auto n = make_shared<v8::Slice>(input_shape, input_rank_minus_one, input_rank, slice_step, slice_axis)->output(0);
+
+    // generate ranges [0, M) and [0, N)
+    auto scalar_shape = make_shared<v0::Constant>(element::i64, Shape{0}, vector<int64_t>{});
+    m = make_shared<v1::Reshape>(m, scalar_shape, false);
+    n = make_shared<v1::Reshape>(n, scalar_shape, false);
+    auto range_m = make_shared<v4::Range>(const_zero, m, const_one, element::i64)->output(0);
+    auto range_n = make_shared<v4::Range>(const_zero, n, const_one, element::i64)->output(0);
+    range_m = make_shared<v0::Unsqueeze>(range_m, const_one);
+    range_n = make_shared<v0::Unsqueeze>(range_n, const_zero);
+
+    // adjust num_lower and num_upper to have them of type i64
+    // the same as M and N
+    // it is needed for in_band computation
+    num_lower = make_shared<v0::Convert>(num_lower, element::i64);
+    num_upper = make_shared<v0::Convert>(num_upper, element::i64);
+
+    // compute in_band(m, n) = (num_lower < 0 || (m-n) <= num_lower)) && (num_upper < 0 || (n-m) <= num_upper)
+    auto num_lower_less_zero = make_shared<v1::Less>(num_lower, const_zero);
+    auto i_minus_j = make_shared<v1::Subtract>(range_m, range_n);
+    auto i_minus_j_less_eq_num_lower = make_shared<v1::LessEqual>(i_minus_j, num_lower);
+    auto num_upper_less_zero = make_shared<v1::Less>(num_upper, const_zero);
+    auto j_minus_i = make_shared<v1::Subtract>(range_n, range_m);
+    auto j_minus_i_less_eq_num_upper = make_shared<v1::LessEqual>(j_minus_i, num_upper);
+    auto in_band1 = make_shared<v1::LogicalOr>(num_lower_less_zero, i_minus_j_less_eq_num_lower);
+    auto in_band2 = make_shared<v1::LogicalOr>(num_upper_less_zero, j_minus_i_less_eq_num_upper);
+    auto in_band = make_shared<v1::LogicalAnd>(in_band1, in_band2);
+
+    // create zero constant of the same type as input
+    auto zero = create_same_type_const_scalar<int32_t>(input, 0);
+
+    auto result = make_shared<v1::Select>(in_band, input, zero);
+
+    set_node_name(node.get_name(), result);
+    return {result};
+}
+
+}  // namespace op
+}  // namespace tensorflow
+}  // namespace frontend
+}  // namespace ov

tests/layer_tests/tensorflow_tests/test_tf_MatrixBandPart.py

@@ -0,0 +1,41 @@
+# Copyright (C) 2024 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+import numpy as np
+import pytest
+import tensorflow as tf
+from common.tf_layer_test_class import CommonTFLayerTest
+
+rng = np.random.default_rng()
+
+
+class TestMatrixBandPart(CommonTFLayerTest):
+    def _prepare_input(self, inputs_info):
+        assert 'input:0' in inputs_info
+        input_shape = inputs_info['input:0']
+        inputs_data = {}
+        inputs_data['input:0'] = rng.integers(-8, 8, input_shape).astype(self.input_type)
+        return inputs_data
+
+    def create_matrix_band_part_net(self, input_shape, input_type, num_lower, num_upper):
+        self.input_type = input_type
+        tf.compat.v1.reset_default_graph()
+        with tf.compat.v1.Session() as sess:
+            input_tensor = tf.compat.v1.placeholder(input_type, input_shape, 'input')
+            tf.raw_ops.MatrixBandPart(input=input_tensor, num_lower=num_lower, num_upper=num_upper)
+            tf.compat.v1.global_variables_initializer()
+            tf_net = sess.graph_def
+        return tf_net, None
+
+    @pytest.mark.parametrize('input_shape', [[5, 5], [3, 4, 4], [1, 2, 5, 5], [3, 5, 4]])
+    @pytest.mark.parametrize('input_type', [np.float32, np.int32])
+    @pytest.mark.parametrize('num_lower', [-4, -1, 0, 1, 4])
+    @pytest.mark.parametrize('num_upper', [-4, -1, 0, 1, 4])
+    @pytest.mark.precommit_tf_fe
+    @pytest.mark.nightly
+    def test_matrix_band_part_basic(self, input_shape, input_type, num_lower, num_upper,
+                                    ie_device, precision, ir_version, temp_dir,
+                                    use_legacy_frontend):
+        self._test(*self.create_matrix_band_part_net(input_shape, input_type, num_lower, num_upper),
+                   ie_device, precision, ir_version, temp_dir=temp_dir,
+                   use_legacy_frontend=use_legacy_frontend)