Confused about AveragePool op ceil_mode attribute

[blacklong28]

Ask a Question

Question

When I was study the AveragePool op, I am confused about the ceil_mode attr.
What I see at AveragePool

ceil_mode : int (default is 0)
Whether to use ceil or floor (default) to compute the output shape.

But I notice when auto_pad is set,
it only uses ceil to compute the output shape.
Is the ceil_mode attr only used when auto_pad is 'NOTSET'?

Further information

• This is the python code I find to compute the output shape, the code also is only use ceil to compute output shape. get_output_shape
• I notice pytorch's op AvgPool2d have ceil_mode( AvgPool2d, )I guess the onnx ceil_mode attr is only used when auto_pad is 'NOTSET', and when auto_pad is not 'NOTSET',the ceil_mode default value is 1 ,is true?

Who can help me? Thanks very much.

[jcwchen]

Hi @blacklong28,
I think Operator document is correct and the test script in ONNX you mentioned get_output_shape is not updated while introducing ceil_mode. You can check the inference logic here in ONNXRunitme:
https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/core/providers/cpu/nn/pool_attributes.h
By default, ceil_mode is 0 no matter what auto_pad is. Thanks.

[kraiskil]

For reference - I had a similar question about ceil mode in MaxPool #2971
@jcwchen can you confirm that the onnx test scripts are wrong? Since they seem to be the de-facto acceptance tests for backends, at least I went with the script and not the documentation as the basis for my backend implementation.

[blacklong28]

@kraiskil @jcwchen
Thank you for your attention and your answers.
I have read this code: https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/core/providers/cpu/nn/pool_attributes.h
and I find the output shape compute when auto_pad is SAME_LOWER SAME_UPPER is equal to :
ceil_or_floor(static_cast(in_size + pad_needed - dilation * (kernel - 1) - 1) / stride + 1)
pad_needed = (int((in_size + stride - 1) / stride)-1)*stride+kernel-in_size
dilation = 1 for AveragePool,
output_shape =ceil_or_floor(static_cast<float>(in_size + (int((in_size + stride - 1) / stride)-1)*stride+kernel-in_size - dilation * (kernel - 1) - 1) / stride + 1)
= ceil_or_floor(
static_cast<float>(
( int( (in_size + stride - 1) / stride
) -1)*stride
)/stride+1
)
= ceil_or_floor(int( (in_size + stride - 1)/stride-1)+1
=ceil_or_floor( int ( (in_size + stride - 1) / stride ) )
=floor ( (in_size + stride - 1) / stride )
= ceil((in_size + stride) / stride)
We know that int ( (in_size + stride - 1) / stride ) is a int number,
so no matter what the ceil_mode is, the output shape is int ( (in_size + stride - 1) / stride ),int() is a floor method.
And is equal to the python get_out_shape function:

  out_shape[i] = int(
      np.ceil(
          float(
              input_spatial_shape[i])
          / float(
              strides_spatial[i])))

from the https://github.com/microsoft/onnxruntime/blob/master/onnxruntime/core/providers/cpu/nn/pool_attributes.h,
When auto_pad is VALID,the output shape is:
ceil_or_floor( (in_size + pad_needed - dilation * (kernel - 1) - 1) / stride + 1)
= ceil_or_floor( (in_size + 0 - 1 * (kernel - 1) - 1) / stride + 1)
= ceil_or_floor( (in_size - kernel ) / stride + 1)
ceil_or_floor is depend on ceil_mode
But,the doc https://github.com/onnx/onnx/blob/master/docs/Operators.md#AveragePool

VALID: output_spatial_shape[i] = ceil((input_spatial_shape[i] - kernel_spatial_shape[i] + 1) / strides_spatial_shape[i])

The two places don't mesh well

1. Does it mean that when auto_pad is equal to SAME_LOWER and SAME_UPPER, no matter what the value of ceil_mode is, the output_shape calculation is independent of ceil_mode?
2. When auto pad is VAILD, which one we take as the standard?
3. Can we choose OnnxRuntime as our standard?

[jcwchen]

HI @kraiskil @blacklong28,
Sorry for the late reply. We are still discussing which behavior is expected while using audo_pad and ceil_mode at the same time. IMO, the implementation in ORT makes more sense to me because ceil_mode should be not silent in any condition. Even if ceil_mode does not apply to auto_pad, then I think ONNX should at least forbid users to use them simultaneously to avoid confusion.

We will update here once @onnx/sig-operators has finalized the decision. It is definitely something needs to be clear in the document/ONNX test code. Thank you for bringing this up!

[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.

[jcwchen]

Reopen it since this issue hasn't been resolved yet.