about GatherElements operator

[Like2021]

Thanks for your excellent open source work, I guess I've run into a new problem.

I have a GatherElements operator, like:

This operator has two inputs. The input shapes are 1×195 and 1×196 respectively.
When i use onnx_tool.share_infer(), I found that GatherElements node will call shape_infer directly, like:

class GatherElementsNode(Node):
    def shape_infer(self, intensors: List[Tensor], outtensors: List[Tensor]):
        outtensors[0].update_shape(intensors[0].get_shape())
        outtensors[0].update_dtype(intensors[0].dtype)

According to the code, it will directly default the output shape to be the same as the first Tensor of intensors, but for this GatherElements node, its Indices should be 1×195 input, then the output should be consistent with Indices, which is also 1×195, not 1×196.

The node attribute shape_calc will call the function shape_infer if it is False,

def shape_infer(self, inputs: {} = None):
    ...
    if node.shape_calc:
        node.value_infer(itensors, otensors)
    else:
        node.shape_infer(itensors, otensors)

so I guess shape_calc about GatherElements should be True?

Could you give some tips for this situation?

[ThanatosShinji]

It seems like a bug of GatherElements. The output shape should be outtensors[0].update_shape(intensors[1].get_shape()). Is your graph with tensor shapes (in your posted image) generated by onnx_tool?

[ThanatosShinji]

I've fixed the bug of GatherElements in this commit 1d170c7.
Thanks for your feedback!

[Like2021]

No, it is not generated by onnx_tool.

It seems like a bug of GatherElements. The output shape should be outtensors[0].update_shape(intensors[1].get_shape()). Is your graph with tensor shapes (in your posted image) generated by onnx_tool?