[InferDT] disable scaled-int dtype inference by default

fastmachinelearning · Jan 30, 2025 · c34bd9a · c34bd9a
1 parent 2b56701
commit c34bd9a
Show file tree

Hide file tree

Showing 2 changed files with 2 additions and 2 deletions.
diff --git a/src/qonnx/transformation/infer_datatypes.py b/src/qonnx/transformation/infer_datatypes.py
@@ -152,7 +152,7 @@ class InferDataTypes(Transformation):
     """Infer QONNX DataType info for all intermediate/output tensors based on
     inputs and node type."""
 
-    def __init__(self, allow_scaledint_dtypes=True):
+    def __init__(self, allow_scaledint_dtypes=False):
         super().__init__()
         self.allow_scaledint_dtypes = allow_scaledint_dtypes
 

diff --git a/src/qonnx/util/inference_cost.py b/src/qonnx/util/inference_cost.py
@@ -118,7 +118,7 @@ def inference_cost(
         model = model.transform(FoldConstants(exclude_op_types=[]))
         model = model.transform(RemoveUnusedTensors())
         model = model.transform(RemoveStaticGraphInputs())
-        model = model.transform(InferDataTypes())
+        model = model.transform(InferDataTypes(allow_scaledint_dtypes=True))
     model = model.transform(GiveUniqueNodeNames())
     model = model.transform(GiveReadableTensorNames())
     if output_onnx is not None: