[TF FE][TF Hub] Support DivNoNan and EnsureShape operations (#19007)

Signed-off-by: Kazantsev, Roman <roman.kazantsev@intel.com>

src/frontends/tensorflow/src/op_table.cpp

@@ -140,12 +140,14 @@ const std::map<std::string, CreatorFunction> get_supported_ops() {
         {"CTCGreedyDecoder", CreatorFunction(translate_ctc_greedy_decoder_op)},
         {"CTCLoss", CreatorFunction(translate_ctc_loss_op)},
         {"Cumsum", CreatorFunction(translate_cumsum_op)},
+        {"DivNoNan", CreatorFunction(translate_div_no_nan_op)},
         {"DepthToSpace", CreatorFunction(translate_depth_to_space_op)},
         {"DepthwiseConv2dNative", CreatorFunction(translate_depthwise_conv_2d_native_op)},
         {"DynamicPartition", CreatorFunction(translate_dynamic_partition_op)},
         {"Einsum", CreatorFunction(translate_einsum_op)},
         {"Elu", CreatorFunction(translate_elu_op)},
         {"EmptyTensorList", CreatorFunction(translate_tensor_list_reserve_op)},
+        {"EnsureShape", CreatorFunction(translate_identity_op)},
         {"ExpandDims", CreatorFunction(translate_expand_dims_op)},
         {"ExtractImagePatches", CreatorFunction(translate_extract_image_patches_op)},
         {"FakeQuantWithMinMaxVars", CreatorFunction(translate_fake_quant_op)},

src/frontends/tensorflow_common/include/common_op_table.hpp

@@ -55,6 +55,7 @@ OP_CONVERTER(translate_cumsum_op);
 OP_CONVERTER(translate_crop_and_resize_op);
 OP_CONVERTER(translate_depth_to_space_op);
 OP_CONVERTER(translate_depthwise_conv_2d_native_op);
+OP_CONVERTER(translate_div_no_nan_op);
 OP_CONVERTER(translate_dynamic_partition_op);
 OP_CONVERTER(translate_einsum_op);
 OP_CONVERTER(translate_elu_op);

src/frontends/tensorflow_common/src/op/div_no_nan.cpp

@@ -0,0 +1,48 @@
+// Copyright (C) 2018-2023 Intel Corporation
+// SPDX-License-Identifier: Apache-2.0
+//
+
+#include "common_op_table.hpp"
+#include "openvino/op/constant.hpp"
+#include "openvino/op/convert_like.hpp"
+#include "openvino/op/divide.hpp"
+#include "openvino/op/equal.hpp"
+#include "openvino/op/select.hpp"
+
+using namespace std;
+using namespace ov::op;
+
+namespace ov {
+namespace frontend {
+namespace tensorflow {
+namespace op {
+OutputVector translate_div_no_nan_op(const NodeContext& node) {
+    default_op_checks(node, 2, {"DivNoNan"});
+    auto numer = node.get_input(0);
+    auto denom = node.get_input(1);
+
+    // prepare auxiliary zero and one constants of the same type as the inputs
+    auto zero = make_shared<v0::Constant>(element::f32, Shape{}, 0.0f)->output(0);
+    auto one = make_shared<v0::Constant>(element::f32, Shape{}, 1.0f)->output(0);
+    zero = make_shared<v1::ConvertLike>(zero, denom);
+    one = make_shared<v1::ConvertLike>(one, denom);
+
+    // compute a mask to get positions of Nan values of division result
+    auto is_zero = make_shared<v1::Equal>(denom, zero);
+
+    // fix zeros in the denomimator to avoid undefined behaviour
+    auto fixed_denom = make_shared<v1::Select>(is_zero, one, denom);
+
+    // compute Division and do not afraid division by zero
+    // since all of them fixed
+    auto div = make_shared<v1::Divide>(numer, fixed_denom);
+
+    // set zero to the result where initially the denomimator is zero
+    auto div_no_nan = make_shared<v1::Select>(is_zero, zero, div);
+    set_node_name(node.get_name(), div_no_nan);
+    return div_no_nan->outputs();
+}
+}  // namespace op
+}  // namespace tensorflow
+}  // namespace frontend
+}  // namespace ov

src/frontends/tensorflow_common/src/op/identity.cpp

@@ -3,10 +3,8 @@
 //
 
 #include "common_op_table.hpp"
-#include "openvino/opsets/opset8.hpp"
 
 using namespace std;
-using namespace ov::opset8;
 
 namespace ov {
 namespace frontend {
@@ -14,7 +12,8 @@ namespace tensorflow {
 namespace op {
 
 OutputVector translate_identity_op(const NodeContext& node) {
-    vector<string> supported_ops = {"Identity",
+    vector<string> supported_ops = {"EnsureShape",
+                                    "Identity",
                                     "PreventGradient",
                                     "Snapshot",
                                     "StopGradient",

tests/layer_tests/tensorflow_tests/test_tf_DivNoNan.py

@@ -0,0 +1,48 @@
+# Copyright (C) 2018-2023 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
+
+
+class TestDivNoNan(CommonTFLayerTest):
+    def _prepare_input(self, inputs_info):
+        assert 'x' in inputs_info
+        assert 'y' in inputs_info
+        x_shape = inputs_info['x']
+        y_shape = inputs_info['y']
+        inputs_data = {}
+        inputs_data['x'] = np.random.randint(-10, 10, x_shape).astype(self.input_type)
+        # generate y in way to have zeros
+        inputs_data['y'] = np.random.randint(-10, 10, y_shape).astype(self.input_type) * \
+                           np.random.randint(0, 2, y_shape).astype(self.input_type)
+        return inputs_data
+
+    def create_div_no_nan_net(self, input_shape, input_type):
+        self.input_type = input_type
+        tf.compat.v1.reset_default_graph()
+        # Create the graph and model
+        with tf.compat.v1.Session() as sess:
+            x = tf.compat.v1.placeholder(input_type, input_shape, 'x')
+            y = tf.compat.v1.placeholder(input_type, input_shape, 'y')
+            tf.raw_ops.DivNoNan(x=x, y=y)
+            tf.compat.v1.global_variables_initializer()
+            tf_net = sess.graph_def
+
+        return tf_net, None
+
+    test_data_basic = [
+        dict(input_shape=[10, 20], input_type=np.float32),
+        dict(input_shape=[2, 3, 4], input_type=np.float32),
+    ]
+
+    @pytest.mark.parametrize("params", test_data_basic)
+    @pytest.mark.precommit_tf_fe
+    @pytest.mark.nightly
+    def test_div_no_nan_basic(self, params, ie_device, precision, ir_version, temp_dir,
+                              use_new_frontend, use_old_api):
+        self._test(*self.create_div_no_nan_net(**params),
+                   ie_device, precision, ir_version, temp_dir=temp_dir,
+                   use_new_frontend=use_new_frontend, use_old_api=use_old_api)

tests/layer_tests/tensorflow_tests/test_tf_EnsureShape.py

@@ -0,0 +1,44 @@
+# Copyright (C) 2018-2023 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
+
+
+class TestEnsureShape(CommonTFLayerTest):
+    def _prepare_input(self, inputs_info):
+        assert 'tensor' in inputs_info
+        tensor_shape = inputs_info['tensor']
+        inputs_data = {}
+        inputs_data['tensor'] = np.random.randint(-10, 10, tensor_shape).astype(self.input_type)
+        return inputs_data
+
+    def create_ensure_shape_net(self, input_shape, input_type, target_shape):
+        self.input_type = input_type
+        tf.compat.v1.reset_default_graph()
+        # Create the graph and model
+        with tf.compat.v1.Session() as sess:
+            tensor = tf.compat.v1.placeholder(input_type, input_shape, 'tensor')
+            shape = tf.constant(target_shape, dtype=tf.int32)
+            reshape = tf.raw_ops.Reshape(tensor=tensor, shape=shape)
+            tf.raw_ops.EnsureShape(input=reshape, shape=target_shape)
+            tf.compat.v1.global_variables_initializer()
+            tf_net = sess.graph_def
+
+        return tf_net, None
+
+    test_data_basic = [
+        dict(input_shape=[2, 6], input_type=np.float32, target_shape=[2, 3, 2]),
+        dict(input_shape=[1], input_type=np.float32, target_shape=[]),
+    ]
+
+    @pytest.mark.parametrize("params", test_data_basic)
+    @pytest.mark.precommit_tf_fe
+    @pytest.mark.nightly
+    def test_ensure_shape_basic(self, params, ie_device, precision, ir_version, temp_dir,
+                                use_new_frontend, use_old_api):
+        self._test(*self.create_ensure_shape_net(**params),
+                   ie_device, precision, ir_version, temp_dir=temp_dir,
+                   use_new_frontend=use_new_frontend, use_old_api=use_old_api)