Padding specification unclear for Convolution/AveragePool/MaxPool

[tlepley-cadence]

For Convolution, AveragePool, MaxPool operators, the definition of the 'pads' property is unclear and suggests that padding should be given for all dimensions, not just for spatial dimensions. But all the tools I've used (for instance TF2ONNX and Glow) agree that only the padding for spatial dimensions needs to be given.

I think the ONNX specification would need some clarification.

Convolution 'pads' property:

[...] pads format should be as follow [x1_begin, x2_begin...x1_end, x2_end,...], where xi_begin the number of pixels added at the beginning of axis i and xi_end, the number of pixels added at the end of axis i

For the Pad operator, ONNX tells explicitly that the padding information needs to be provided for all dimensions, so it is clear:

[...] pads rank should be double of the input's rank. pads format should be as follow [x1_begin, x2_begin...x1_end, x2_end,...], where xi_begin the number of pixels added at the beginning of axis i and xi_end, the number of pixels added at the end of axis i.

As an example, in some of the ONNX models I've used, for 4D tensors, I got a 4 element pads array for Convolution&Maxpool and a 8 element pads' array for Pad`

• Convolution op: attribute = name: "pads" ints: 1 ints: 1 ints: 1 ints: 1 type: INTS
• Maxpool op: attribute = name: "pads" ints: 0 ints: 0 ints: 0 ints: 0 type: INTS
• Pad op: attribute = name: "pads" ints: 0 ints: 0 ints: 1 ints: 1 ints: 0 ints: 0 ints: 1 ints: 1 type: INTS

[charanpool]

Hey, @tlepley-cadence did you find the padding specifications?

[stale]

Is this still relevant? If so, what is blocking it? Is there anything you can do to help move it forward?

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs.