[python] Fix classification usage via OVMS (#198)

* Fix classification usage via OVMS

* Tighten the condition to filter possible non-ov adapters usage

* Add one more comment on classificaiton postprocessing

model_api/python/model_api/models/classification.py

@@ -80,9 +80,16 @@ def __init__(self, inference_adapter, configuration=dict(), preload=False):
             if self.output_raw_scores:
                 self.out_layer_names.append(self.raw_scores_name)
         except (RuntimeError, AttributeError):
-            self.embedded_topk = False
-            self.out_layer_names = _get_non_xai_names(self.outputs.keys())
-            self.raw_scores_name = self.out_layer_names[0]
+            # exception means we have a non-ov model, or a model behind OVMS
+            # with already inserted softmax and topk
+            if self.embedded_processing and len(self.outputs) >= 2:
+                self.embedded_topk = True
+                self.out_layer_names = ["indices", "scores"]
+                self.raw_scores_name = _raw_scores_name
+            else:  # likely a non-ov model
+                self.embedded_topk = False
+                self.out_layer_names = _get_non_xai_names(self.outputs.keys())
+                self.raw_scores_name = self.out_layer_names[0]
 
         self.embedded_processing = True