Padding type definition is swapped in the documentation.

[cesarsalgado]

In http://tensorflow.org/api_docs/python/nn.md#pooling and http://tensorflow.org/api_docs/python/nn.md#convolution

The definition of the padding types should be swapped.

Now it is like this:

padding = 'SAME': Round down (only full size windows are considered).
padding = 'VALID': Round up (partial windows are included).

But it should be like this:

padding = 'SAME': Round up (partial windows are included).
padding = 'VALID': Round down (only full size windows are considered).

I also think it should be made clear that the padding is already included in the shape(value) of the formula "shape(output) = (shape(value) - ksize + 1) / strides".

Or maybe the formula should be changed to the following simpler one (taken from http://cs231n.github.io/convolutional-networks/#conv):

shape(output) = ((shape(value)+2*pad - ksize) / strides) + 1

where pad = 0 if padding='VALID' or pad = floor(ksize/2) if padding='SAME'.

[Yangqing]

Thanks - the padding documentation definitely needs modification. The padding scheme is not as simple as you mentioned though - it follows some legacy reason which involved some non-trivial padding computations. I'll update the documentation to add the details.

For the time being, if you are interested in the the exact math, it is described here:

https://github.com/Yangqing/caffe2/blob/master/caffe2/proto/caffe2_legacy.proto#L8

[Yangqing]

Thanks again for reporting this. We have fixed this in our internal codebase and will be pushing it soon. Closing this for now.

[NianzuMa]

Propose to enrich the document of convolution. It is still quite confusing.
http://tensorflow.org/api_docs/python/nn.md#convolution

I think what cesarsalgado mentioned is correct.

[jiankang1991]

Could anyone help me to calculate each step output of cifar10 example of tensorflow?
The input image size is 24*24. But in max pooling step, the kernel size is 3 and stride is 2, so (24-3)/2+1 is not integer. What is wrong here? I am new in deep learning. Thank you for helping

[sesse]

Hi,

Beginner questions are best suited for Stack Overflow, not random semi-related bugs. Please post your question there instead.