Fix input quantization in case of embeddings (#97)

openvinotoolkit · Aug 11, 2020 · 357f96b · 357f96b
1 parent 20d9e44
commit 357f96b
Show file tree

Hide file tree

Showing 3 changed files with 455 additions and 455 deletions.
diff --git a/nncf/quantization/algo.py b/nncf/quantization/algo.py
@@ -491,7 +491,11 @@ def _quantize_inputs(self, target_model: NNCFNetwork,
 
         def traverse_function(node: NNCFNode, output) -> Tuple[bool, List[NNCFNode]]:
             module = target_model.get_module_by_scope(node.op_exec_context.scope_in_model)
-            if is_nncf_module(module) and not isinstance(module, NNCFEmbedding):  # Embeddings have integer input
+            if is_nncf_module(module):
+                if isinstance(module, NNCFEmbedding):
+                    # Embeddings have integer input and their quantization is rather controlled
+                    # by their weights.
+                    return True, output
                 current_node_scope = node.op_exec_context.scope_in_model
                 module_op_insertion_commands = []
                 for comm in prev_weight_and_activation_quantizer_insertion_commands: