Refine ORT QDQ quantization graph (#589)

Signed-off-by: yuwenzho <yuwen.zhou@intel.com>

neural_compressor/adaptor/ox_utils/quantizer.py

@@ -220,6 +220,23 @@ def merge_dedicated_qdq_pair(self):
                 for node, old_input_name, new_input_name in self.replace_input:
                     self.model.replace_node_input(node, old_input_name, new_input_name)
                 self.model.update()
+
+        if self.mode == 'qdq':
+            for node in self.model.nodes():
+                if node.op_type in ['QuantizeLinear'] and len(self.model.get_parents(node)) > 0:
+                    if 'QuantizeLinear' in [sibling.op_type \
+                                            for sibling in self.model.get_siblings(node)]:
+                        continue
+                    for sibling in self.model.get_siblings(node):
+                        if not self.should_quantize(sibling) and sibling.op_type in OPERATORS:
+                            for inp_idx in range(len(sibling.input)):
+                                if sibling.input[inp_idx] == node.input[0]:
+                                    self.replace_input.append([sibling,
+                                        sibling.input[inp_idx],
+                                        self.model.get_children(node)[0].output[0]])
+            for node, old_input_name, new_input_name in self.replace_input:
+                self.model.replace_node_input(node, old_input_name, new_input_name)
+            self.model.update()
 
     def should_cast(self, node):
         """Check if node should be casted."""

neural_compressor/experimental/benchmark.py

@@ -604,6 +604,7 @@ def postprocess(self, user_postprocess):
             logger.warning("Override the value of `postprocess` field defined in yaml file" \
                            " as user defines the value of `postprocess` attribute by code.")
         deep_set(self.conf.usr_cfg, "evaluation.accuracy.postprocess.transform", postprocess_cfg)
+        from neural_compressor.data import TRANSFORMS
         postprocesses = TRANSFORMS(self.framework, 'postprocess')
         postprocesses.register(user_postprocess.name, user_postprocess.postprocess_cls)
 

neural_compressor/model/onnx_model.py

@@ -253,6 +253,15 @@ def _get_output_name_to_node(self, nodes):
             for output_name in node.output:
                 self._output_name_to_node[output_name] = node
 
+    def get_siblings(self, node):
+        """Get siblings nodes."""
+        siblings = []
+        for parent in self.get_parents(node):
+            for child in self.get_children(parent):
+                if child.name != node.name:
+                    siblings.append(child)
+        return siblings
+
     def get_children(self, node, input_name_to_nodes=None):
         """Get children nodes."""
         if input_name_to_nodes is None:

test/model/test_onnx_model.py

@@ -203,6 +203,16 @@ def test_output_name_to_node(self):
         opts = ['X1', 'X2', 'X3', 'X4', 'X5', 'output']
         for opt in opts:
             self.assertTrue(opt in opts_name)
+
+    def test_get_siblings(self):
+        for node in self.model.nodes():
+            if node.name == "Conv1":
+                siblings = self.model.get_siblings(node)
+        self.assertEqual(len(siblings), 1)
+        siblings_name = [sibling.name for sibling in siblings]
+        names = ["Conv3"]
+        for name in names:
+            self.assertTrue(name in siblings_name)
 
     def test_get_children(self):
         for node in self.model.nodes():