[Feature request] Support 1D vector for w_scale in QLinearConv reference implementation #6428
Description
System information
1.17.0 release
What is the problem that this feature solves?
Documentation for w_scale argument to QLinearConv operator (https://github.com/onnx/onnx/blob/41cba9d512620781257163cb2ba072871456f8d3/docs/Operators.md#QLinearConv) states:
It could be a scalar or a 1-D tensor, which means a per-tensor/layer or per output channel quantization. If it's a 1-D tensor, its number of elements should be equal to the number of output channels (M).
If using QLinearConv with a 1D vector for w_scale in onnx.reference.ReferenceEvaluator, though, will get a numpy broadcast error. It can be seen by running this file:
https://gist.github.com/mcollinswisc/bccaa6730089c4221f53415a39b6835b
I see error:
Traceback (most recent call last):
File "onnx_qlinearconv.py", line 71, in <module>
ref_result = ref_eval.run(None, {"x": x})
^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "env/qlinearconv/lib/python3.12/site-packages/onnx/reference/reference_evaluator.py", line 599, in run
outputs = node.run(*inputs, **linked_attributes)
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
File "env/qlinearconv/lib/python3.12/site-packages/onnx/reference/op_run.py", line 466, in run
res = self._run(*args, **kwargs)
^^^^^^^^^^^^^^^^^^^^^^^^^^
File "env/qlinearconv/lib/python3.12/site-packages/onnx/reference/ops/op_qlinear_conv.py", line 53, in _run
R = res * (x_scale * w_scale / y_scale)
~~~~^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
ValueError: operands could not be broadcast together with shapes (1,8,24,24) (8,)
Thrown from here:
Alternatives considered
N/A
Describe the feature
For getting full reference implementation results from QLinearConv.
Will this influence the current api (Y/N)?
No
Feature Area
ReferenceEvaluator
Are you willing to contribute it (Y/N)
Yes
Notes
No response