[MO] Correct preprocessing block for TensorFlow OD 2.4+ models (#9944)

* Detect casting node inside preprocessing block and leave it

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

* Fix unit-test for ObjectDetectionAPI

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

tools/mo/openvino/tools/mo/front/tf/ObjectDetectionAPI.py

@@ -651,7 +651,15 @@ def get_preprocessing_ops(graph: Graph, start_node_id_suffix: str, end_node_id_s
     if len(preprocessing_nodes) == 0:
         preprocessing_nodes = get_specific_ops_with_const_inputs(end_node, allowed_ops, True)
         trailing = True
-    return preprocessing_nodes, trailing
+
+    # try to detect floating-point casting inside the body graph
+    # that also needs to stay in the resulted graph
+    casting = False
+    cast_nodes = backward_bfs_for_operation(start_node, ['Cast'])
+    if len(cast_nodes) == 1 and cast_nodes[0].dst_type == np.float32:
+        casting = True
+
+    return preprocessing_nodes, trailing, casting
 
 
 """ 
@@ -848,11 +856,11 @@ class ObjectDetectionAPIPreprocessor2Replacement(FrontReplacementFromConfigFileG
             # we stick to the nodes with ids 'map/while/Preprocessor/unstack' and 'map/while/Preprocessor/stack' as they
             # "wrap" nodes performing image resize. The scale/mean values nodes are located strictly before or after
             # them
-            pre_processing_ops, trailing = get_preprocessing_ops(body_graph,
+            pre_processing_ops, trailing, casting = get_preprocessing_ops(body_graph,
                                                                  'map/while/Preprocessor/unstack',
                                                                  'map/while/Preprocessor/stack')
         else:
-            pre_processing_ops, trailing = get_preprocessing_ops(graph, start_node.id, end_node.id)
+            pre_processing_ops, trailing, casting = get_preprocessing_ops(graph, start_node.id, end_node.id)
 
         mean_scale_kept = True
         if len(pre_processing_ops):
@@ -863,6 +871,12 @@ class ObjectDetectionAPIPreprocessor2Replacement(FrontReplacementFromConfigFileG
             if pre_processing_in_loop:  # case 4 and 5
                 # build a sub-graph containing a sequence of pre_processing_ops if they came from the Loop
                 new_preprocessing_ops = []
+
+                # cast data before start pre-processing with mean/scale values
+                if casting:
+                    cast_node = Cast(graph, {'dst_type': np.float32}).create_node()
+                    new_preprocessing_ops.append(cast_node)
+
                 ops_mapping = {'Add': Add, 'Div': Div, 'Mul': Mul, 'Sub': Sub}
                 for idx in range(len(pre_processing_ops)):
                     origin_node, const_port_ind, value = pre_processing_ops[idx]

tools/mo/unit_tests/mo/front/tf/ObjectDetectionAPI_test.py

@@ -112,7 +112,7 @@ class TestObjectDetectionAPIPreprocessor2Replacement(unittest.TestCase):
         self.sub_const = float32_array([2.0, 3.0, 4.0])
 
         self.nodes = {
-            **regular_op('input', {'type': 'Parameter'}),
+            **regular_op('input', {'op': 'Parameter', 'type': 'Parameter'}),
 
             **regular_op('mul', {'op': 'Mul', 'type': 'Multiply', 'name': 'my_mul'}),
             **regular_op('sub', {'op': 'Sub', 'type': 'Subtract', 'name': 'my_sub'}),
@@ -238,7 +238,7 @@ class TestObjectDetectionAPIPreprocessor2Replacement(unittest.TestCase):
     def build_main_graph(self, pre_processing: str):
         def build_body_graph(pre_processing: str):
             nodes = {
-                **regular_op('input', {'type': 'Parameter'}),
+                **regular_op('input', {'type': 'Parameter', 'op': 'Parameter'}),
 
                 **regular_op('mul', {'op': 'Mul', 'type': 'Multiply', 'name': 'my_body_mul'}),
                 **regular_op('sub', {'op': 'Sub', 'type': 'Subtract', 'name': 'my_body_sub'}),