bug fix for when there are shared initializers across multiple pads in an ONNX

veerbhan · veerbhan · commit 2b04dd0bdd86 · 2023-08-23T13:10:23.000-07:00
When there are multiple Pad's in a network which use the same initializer, the subsequent Pad quantization
doesn't find this initializer since it was deleted. It ends up creating a dangling QuantizeLinear node for the input[2] to the Pad which is a constant.
diff --git a/onnxruntime/python/tools/quantization/operators/pad.py b/onnxruntime/python/tools/quantization/operators/pad.py
@@ -56,8 +56,6 @@ def quantize(self):
                         quantized_padding_constant_array,
                         quantized_padding_constant_name,
                     )
-                    # Suppose this padding constant initializer only used by the node
-                    self.quantizer.model.remove_initializer(padding_constant_initializer)
                     self.quantizer.model.add_initializer(quantized_padding_constant_initializer)
                     node.input[2] = quantized_padding_constant_name
                 else:

Original file line number	Diff line number	Diff line change
`@@ -56,8 +56,6 @@ def quantize(self):`
`56`	`56`	`quantized_padding_constant_array,`
`57`	`57`	`quantized_padding_constant_name,`
`58`	`58`	`)`
`59`		`- # Suppose this padding constant initializer only used by the node`
`60`		`- self.quantizer.model.remove_initializer(padding_constant_initializer)`
`61`	`59`	`self.quantizer.model.add_initializer(quantized_padding_constant_initializer)`
`62`	`60`	`node.input[2] = quantized_padding_constant_name`
`63`	`61`	`else:`