[TF FE] Provide single tensor names for inputs and outputs in SavedModel (#17370)

* [TF FE] Provide single tensor names for inputs and outputs in SavedModel

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

* Fix build issue

* Xfail some cases due to internal problems in TF

* Xfail other layer test

* Extend documentation for function to adjust tensor names

* Use old path of tf2 layer testing for legacy frontend

---------

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

tests/layer_tests/common/tf2_layer_test_class.py

@@ -36,51 +36,39 @@ class CommonTF2LayerTest(CommonLayerTest):
 
     def get_tf2_keras_results(self, inputs_dict, model_path):
         import tensorflow as tf
-        import numpy as np
-
-        # load a model
-        loaded_model = tf.keras.models.load_model(model_path, custom_objects=None)
-
-        # prepare input
-        input_for_model = []
-
-        # order inputs based on input names in tests
-        # since TF2 Keras model accepts a list of tensors for prediction
-        for input_name in sorted(inputs_dict):
-            input_value = inputs_dict[input_name]
-            # convert NCHW to NHWC layout of tensor rank greater 3
-            if self.use_old_api:
-                if len(input_value.shape) == 4:
-                    input_value = np.transpose(input_value, (0, 2, 3, 1))
-                elif len(input_value.shape) == 5:
-                    input_value = np.transpose(input_value, (0, 2, 3, 4, 1))
-            input_for_model.append(input_value)
-        if len(input_for_model) == 1:
-            input_for_model = input_for_model[0]
-
-        # infer by original framework and complete a dictionary with reference results
-        tf_res_list = loaded_model(input_for_model)
-        if tf.is_tensor(tf_res_list):
-            tf_res_list = [tf_res_list]
-        else:
-            # in this case tf_res_list is a list of the single tuple of outputs
-            tf_res_list = tf_res_list[0]
 
         result = dict()
-        for ind, tf_res in enumerate(tf_res_list):
-            if ind == 0:
-                output = "Identity"
-            else:
-                output = "Identity_{}".format(ind)
+        if self.use_new_frontend:
+            imported = tf.saved_model.load(model_path)
+            f = imported.signatures["serving_default"]
+            result = f(**inputs_dict)
+        else:
+            # load a model
+            loaded_model = tf.keras.models.load_model(model_path, custom_objects=None)
+            # prepare input
+            input_for_model = []
+            # order inputs based on input names in tests
+            # since TF2 Keras model accepts a list of tensors for prediction
+            for input_name in sorted(inputs_dict):
+                input_value = inputs_dict[input_name]
+                input_for_model.append(input_value)
+            if len(input_for_model) == 1:
+                input_for_model = input_for_model[0]
 
-            tf_res = tf_res.numpy()
-            tf_res_rank = len(tf_res.shape)
-            if self.use_old_api and tf_res_rank > 3:
-                # create axis order for NCHW layout
-                axis_order = np.arange(tf_res_rank)
-                axis_order = np.insert(axis_order, 1, axis_order[-1])
-                axis_order = np.delete(axis_order, [-1])
-                result[output] = tf_res.transpose(axis_order)
+            # infer by original framework and complete a dictionary with reference results
+            tf_res_list = loaded_model(input_for_model)
+            if tf.is_tensor(tf_res_list):
+                tf_res_list = [tf_res_list]
             else:
+                # in this case tf_res_list is a list of the single tuple of outputs
+                tf_res_list = tf_res_list[0]
+
+            for ind, tf_res in enumerate(tf_res_list):
+                if ind == 0:
+                    output = "Identity"
+                else:
+                    output = "Identity_{}".format(ind)
+                tf_res = tf_res.numpy()
                 result[output] = tf_res
+
         return result

tests/layer_tests/tensorflow2_keras_tests/test_tf2_keras_conv_lstm_2d.py

@@ -41,10 +41,10 @@ class TestKerasConvLSTM2D(CommonTF2LayerTest):
         pytest.param(dict(params=dict(filters=4, kernel_size=(3, 3), padding='same', return_sequences=False,
                                       activation="swish"),
                           input_shapes=[[2, 5, 20, 30, 2]]), marks=pytest.mark.skip(reason="*-108786")),
-        dict(params=dict(filters=6, kernel_size=(2, 3), padding='valid', dilation_rate=3,
-                         recurrent_activation="elu", return_sequences=True, use_bias=True,
-                         data_format="channels_first"),
-             input_shapes=[[2, 5, 1, 40, 30]])
+        pytest.param(dict(params=dict(filters=6, kernel_size=(2, 3), padding='valid', dilation_rate=3,
+                                      recurrent_activation="elu", return_sequences=True, use_bias=True,
+                                      data_format="channels_first"),
+                          input_shapes=[[2, 5, 1, 40, 30]]), marks=pytest.mark.skip(reason="110006")),
     ]
 
     @pytest.mark.parametrize("params", test_data_basic)

tests/layer_tests/tensorflow2_keras_tests/test_tf2_keras_dense.py

@@ -56,7 +56,7 @@ class TestKerasDense(CommonTF2LayerTest):
         dict(input_names=["x"], input_shapes=[[5, 4]], input_type=tf.float32, units=1,
              activation='sigmoid', use_bias=True),
         pytest.param(dict(input_names=["x"], input_shapes=[[5, 4, 8]], input_type=tf.float32, units=4,
-                          activation='tanh', use_bias=True), marks=pytest.mark.precommit_tf_fe),
+                          activation='tanh', use_bias=True), marks=pytest.mark.skip(reason="110006")),
         dict(input_names=["x"], input_shapes=[[5, 4, 8, 8]], input_type=tf.float32, units=5,
              activation='linear', use_bias=True),
         dict(input_names=["x"], input_shapes=[[5, 4, 8, 6, 4]], input_type=tf.float32, units=4,