ConvTranspose auto_pad and explicit pads combination should be invalid

[vuzelac-cadence]

X = helper.make_tensor_value_info('X', TensorProto.FLOAT, [1, 1, 2, 2])
W = helper.make_tensor_value_info('W', TensorProto.FLOAT, [1, 1, 3, 3])
B = helper.make_tensor_value_info('B', TensorProto.FLOAT, [1])
Y = helper.make_tensor_value_info('Y', TensorProto.FLOAT, [1, 1, 3, 3])
node_def = helper.make_node("ConvTranspose", ["X", "W", "B"], ["Y"], auto_pad="SAME_UPPER", strides=[1, 1], pads=[0, 1, 1, 0])

The above excerpt doesn't generate error although the spec for ConvTranspose says it should:

pads : list of ints
_...This attribute cannot be used simultaneously with auto_pad attribute...

If they can be used together, auto_pad description hints pads are valid only if auto_pad=NONSET. However, I see pads attribute applies even when auto_pad=SAME_UPPER/SAME_LOWER
`

[TMVector]

Hi @vuzelac-cadence, thanks for the report. The onnx.helper.make_node function has no operator-specific checks. It is the role of the ONNX checker to catch problems like this, but it is not exhaustive by any means. The specification has the final say on whether a node is a valid instantiation of an operator, and ONNX backends/compilers/tools should work with such nodes.
If you would like to add this check to the ONNX checker (if it isn't already there), then please submit a PR :)

[vuzelac-cadence]

@TMVector , I failed to write the whole code but the checker is invoked and it passes, and that was the reason to submit this issue, not the helper code. Sorry about that. So, the check is not there.
Hope you can find time to describe better what this means " it is not exhaustive by any means".

IMO, it should run against everything in the spec, otherwise it's incomplete (or can simply say that it has bugs). Is it really that you intended it to be non-exhaustive ?

[TMVector]

@vuzelac-cadence thanks for clarifying. Yes, the ONNX project is dependent on people and companies dedicating time to work on it, and unfortunately the checker is not complete, but even with lots of people working full time on it I'm sure it wouldn't be perfect :)
I believe it tries to cover the majority of issues by checking machine-readable properties of operators, such as accepted input types and attributes.

[faxu]

CC @jcwchen

[jcwchen]

Solved by #3000. This error can be caught by onnx.checker.check_model(model, full_check=True). @vuzelac-cadence Thank you for reporting this.