Input and output order Keras tests. (#20902)

* Input/output order Keras tests.

* Added precommit mark.

* Added xfail.

* Small correction.

* Check input/outputs by names in FW.

* Moved output order tests to Python API group.

* Corrected comments.

tests/layer_tests/ovc_python_api_tests/test_tf_output_order.py

@@ -0,0 +1,99 @@
+# Copyright (C) 2018-2023 Intel Corporation
+# SPDX-License-Identifier: Apache-2.0
+
+import tempfile
+from pathlib import Path
+
+import numpy as np
+import pytest
+import tensorflow as tf
+
+from common import constants
+
+
+def create_net_list(input_names, input_shapes):
+    tf.keras.backend.clear_session()
+
+    # Create TensorFlow 2 model with multiple outputs.
+    # Outputs are list.
+
+    inputs = []
+    outputs = []
+    for ind in range(len(input_names)):
+        input = tf.keras.Input(shape=input_shapes[ind][1:], name=input_names[ind])
+        inputs.append(input)
+        outputs.append(tf.keras.layers.Activation(tf.nn.sigmoid)(input))
+
+    return tf.keras.Model(inputs=inputs, outputs=outputs)
+
+
+def create_net_dict(input_names, input_shapes):
+    tf.keras.backend.clear_session()
+
+    # Create TensorFlow 2 model with multiple outputs.
+    # Outputs are dictionary.
+
+    inputs = []
+    outputs = {}
+    for ind in range(len(input_names)):
+        input = tf.keras.Input(shape=input_shapes[ind][1:], name=input_names[ind])
+        inputs.append(input)
+        outputs["name" + str(ind)] = tf.keras.layers.Activation(tf.nn.sigmoid)(input)
+
+    return tf.keras.Model(inputs=inputs, outputs=outputs)
+
+
+def check_outputs_by_order(fw_output, ov_output, eps):
+    # Compare outputs by indices
+    for idx, output in enumerate(fw_output):
+        fw_out = output.numpy()
+        ov_out = ov_output[idx]
+        assert fw_out.shape == ov_out.shape, "Output with index {} has shape different from original FW.".format(idx)
+        diff = np.max(np.abs(fw_out - ov_out))
+        assert diff < eps, "Output with index {} has inference result different from original FW.".format(idx)
+
+
+def check_outputs_by_names(fw_output, ov_output, eps):
+    # Compare outputs by names
+    for name, output in fw_output.items():
+        fw_out = output.numpy()
+        ov_out = ov_output[name]
+        assert fw_out.shape == ov_out.shape, "Output with name {} has shape different from original FW.".format(name)
+        diff = np.max(np.abs(fw_out - ov_out))
+        assert diff < eps, "Output with name {} has inference result different from original FW.".format(name)
+
+
+class TestTFInputOutputOrder():
+    def setup_method(self):
+        Path(constants.out_path).mkdir(parents=True, exist_ok=True)
+        self.tmp_dir = tempfile.TemporaryDirectory(dir=constants.out_path).name
+
+    @pytest.mark.parametrize("save_to_file, create_model_method, compare_model_method", [
+        (False, create_net_list, check_outputs_by_order),
+        (False, create_net_dict, check_outputs_by_names),
+        pytest.param(True, create_net_list, check_outputs_by_order, marks=pytest.mark.xfail(reason='124436')),
+        pytest.param(True, create_net_dict, check_outputs_by_names, marks=pytest.mark.xfail(reason='124436')),
+    ])
+    def test_order(self, ie_device, precision, save_to_file, create_model_method, compare_model_method):
+        from openvino import convert_model, compile_model
+        input_names = ["k", "b", "m", "c", "x"]
+        input_shapes = [[1, 1], [1, 3], [1, 2], [1, 5], [1, 4]]
+        epsilon = 0.001
+
+        fw_model = create_model_method(input_names, input_shapes)
+
+        if save_to_file:
+            tf.keras.models.save_model(fw_model, self.tmp_dir + "./model")
+            ov_model = convert_model(self.tmp_dir + "./model")
+        else:
+            ov_model = convert_model(fw_model)
+
+        cmp_model = compile_model(ov_model, ie_device)
+        test_inputs = []
+        for shape in input_shapes:
+            test_inputs.append(np.random.rand(*shape))
+
+        fw_output = fw_model(test_inputs)
+        ov_output = cmp_model(test_inputs)
+
+        compare_model_method(fw_output, ov_output, epsilon)