[MO] Move reverse input channels before mean/scale

docs/MO_DG/prepare_model/convert_model/Converting_Model_General.md

@@ -59,11 +59,14 @@ Framework-agnostic parameters:
   --reverse_input_channels
                         Switch the input channels order from RGB to BGR (or
                         vice versa). Applied to original inputs of the model
-                        if and only if a number of channels equals 3. Applied
-                        after application of --mean_values and --scale_values
-                        options, so numbers in --mean_values and
-                        --scale_values go in the order of channels used in the
-                        original model.
+                        if and only if a number of channels equals 3.
+                        When --mean_values/--scale_values are also specified,
+                        reversing of channels will be applied to user's input
+                        data first, so that numbers in --mean_values and
+                        --scale_values go in the order of channels used in
+                        the original model. In other words, if both options are
+                        specified then the data flow in the model looks as following:
+                        Parameter -> ReverseInputChannels -> Mean/Scale apply -> the original body of the model.
   --log_level {CRITICAL,ERROR,WARN,WARNING,INFO,DEBUG,NOTSET}
                         Logger level
   --input INPUT         Quoted list of comma-separated input nodes names with

tools/mo/openvino/tools/mo/back/preprocessing.py

@@ -372,6 +372,12 @@ def apply_preprocessing(ov_function: Model, argv: argparse.Namespace):
             else:
                 prep.output(node_name).tensor().set_layout(Layout(layout_value['target_layout']))
 
+    # Apply reverse_input_channels
+    if need_reverse:
+        for name, _ in suitable_params_ric:
+            prep.input(name).preprocess().reverse_channels()
+            log.debug('reverse_input_channels pre-processing applied to {}'.format(name))
+
     for node_name, node_mean_scale_values in mean_scale_values.items():
         # Apply mean first, then scale
         if node_mean_scale_values['mean'] is not None:
@@ -380,12 +386,6 @@ def apply_preprocessing(ov_function: Model, argv: argparse.Namespace):
             prep.input(node_name).preprocess().scale(node_mean_scale_values['scale'])
         log.debug('Mean/Scale pre-processing applied to {}'.format(node_name))
 
-    # Apply reverse_input_channels
-    if need_reverse:
-        for name, _ in suitable_params_ric:
-            prep.input(name).preprocess().reverse_channels()
-            log.debug('reverse_input_channels pre-processing applied to {}'.format(name))
-
     # Apply pre-processing builder to a function
     ov_function = prep.build()
 

tools/mo/openvino/tools/mo/utils/cli_parser.py

@@ -271,10 +271,12 @@ def get_common_cli_parser(parser: argparse.ArgumentParser = None):
                                    'the original input of the model.')
     common_group.add_argument('--reverse_input_channels',
                               help='Switch the input channels order from RGB to BGR (or vice versa). Applied to '
-                                   'original inputs of the model if and only if a number of channels equals 3. Applied '
-                                   'after application of --mean_values and --scale_values options, so numbers in '
-                                   '--mean_values and --scale_values go in the order of channels used in the original '
-                                   'model.',
+                                   'original inputs of the model if and only if a number of channels equals 3. '
+                                   'When --mean_values/--scale_values are also specified, reversing of channels will '
+                                   'be applied to user\'s input data first, so that numbers in --mean_values '
+                                   'and --scale_values go in the order of channels used in the original model. '
+                                   'In other words, if both options are specified, then the data flow in the model '
+                                   'looks as following: Parameter -> ReverseInputChannels -> Mean/Scale apply -> the original body of the model.',
                               action='store_true')
     common_group.add_argument('--log_level',
                               help='Logger level',

tools/mo/unit_tests/mo/back/moc_preprocessing_test_actual.py

@@ -56,25 +56,23 @@ class TestPreprocessingMOC(unittest.TestCase):
     def setUp(self):
         pass
 
-    def check_scale_constant(self, node, expected, shape=None):
-        const_node = node.input(1).get_source_output().get_node()
+    def check_constant(self, const_node, expected, shape=None):
         self.assertEqual(const_node.get_type_name(), 'Constant')
-        if node.get_type_name() == 'Divide':
-            self.assertTrue(np.allclose(const_node.get_vector(), expected))
-        else:
-            self.assertTrue(np.allclose(const_node.get_vector(), 1. / expected))
-        if shape:
+        self.assertTrue(np.allclose(const_node.get_vector(), expected))
+        if shape is not None:
             assert const_node.shape == PartialShape(shape)
 
+    def check_scale_constant(self, node, expected, shape=None):
+        const_node = node.input(1).get_source_output().get_node()
+        if node.get_type_name() != 'Divide':
+            expected = 1. / expected
+        self.check_constant(const_node, expected, shape)
+
     def check_mean_constant(self, node, expected, shape=None):
         const_node = node.input(1).get_source_output().get_node()
-        self.assertEqual(const_node.get_type_name(), 'Constant')
-        if node.get_type_name() == 'Subtract':
-            self.assertTrue(np.allclose(const_node.get_vector(), expected))
-        else:
-            self.assertTrue(np.allclose(const_node.get_vector(), -expected.toList()))
-        if shape:
-            self.assertEqual(const_node.shape, PartialShape(shape))
+        if node.get_type_name() != 'Subtract':
+            expected = -expected.toList()
+        self.check_constant(const_node, expected, shape)
 
     def test_scale_single_value(self):
         argv = Namespace(mean_scale_values=None, scale=2.0)
@@ -615,3 +613,41 @@ def test_no_reverse_channels_even_with_layout(self):
         self.assertTrue(op_node0.get_type_name() == 'Relu')
         op_node1 = list(function.get_parameters()[1].output(0).get_target_inputs())[0].get_node()
         self.assertTrue(op_node1.get_type_name() == 'Relu')
+
+    def test_reverse_channels_and_mean_scale(self):
+        argv = Namespace(reverse_input_channels=True, mean_scale_values={
+                                                        'input2a': {
+                                                           'mean': np.array([1., 2., 3.]),
+                                                           'scale': np.array([2., 4., 8.])}},
+                         scale=None)
+        function = create_function2(shape2=[1, 3, 224, 224])
+        process_function(ov_function=function, argv=argv)
+
+        # Verify that first is gather, then subtract 'mean', then 'scale'
+        gather = list(function.get_parameters()[1].output(0).get_target_inputs())[0].get_node()
+        self.assertTrue(gather.get_type_name() == 'Gather')
+        range_node = gather.input(1).get_source_output().get_node()
+        self.assertTrue(range_node.get_type_name() == 'Range')
+        start = range_node.input(0).get_source_output().get_node()
+        end = range_node.input(1).get_source_output().get_node()
+        step = range_node.input(2).get_source_output().get_node()
+        self.check_constant(start, expected=[2], shape=[])
+        self.check_constant(end, expected=[-1], shape=[])
+        self.check_constant(step, expected=[-1], shape=[])
+        axes = gather.input(2).get_source_output().get_node()
+        self.check_constant(axes, expected=[1], shape=[1])
+
+        op_node = list(gather.output(0).get_target_inputs())[0].get_node()
+        self.assertTrue(op_node.get_type_name() == 'Subtract' or op_node.get_type_name() == 'Add')
+        self.check_mean_constant(op_node, expected=[1., 2., 3.], shape=[1, 3, 1, 1])
+
+        op_node = list(op_node.output(0).get_target_inputs())[0].get_node()
+        self.assertTrue(op_node.get_type_name() == 'Divide' or op_node.get_type_name() == 'Multiply')
+        self.check_scale_constant(op_node, expected=[2., 4., 8.], shape=[1, 3, 1, 1])
+
+        # Verify that input1 is not affected
+        op_node = list(function.get_parameters()[0].output(0).get_target_inputs())[0].get_node()
+        self.assertEqual(op_node.get_type_name(), 'Relu')
+
+        # Verify that guessed layout (?C??) is not appeared in input2
+        self.assertEqual(function.get_parameters()[1].layout, Layout())