mxnet conv only support zero padding?

[xlvector]

I find in torch, there are different types of padding like

SpatialReflectionPadding
SpatialReplicationPadding

In mxnet, are there similar padding methods?

[piiswrong]

No. Does cudnn support these?

[sbodenstein]

@xlvector @piiswrong: Torch and TensorFlow have standalone padding layers that support reflective padding etc and padding that isn't applied to all sides. This has many uses, particularly for the likes of style transfer where zero padding is more likely to introduce edge artifacts. It obviously comes at a greater memory cost (as its not merged as part of pooling/convolution), but sometimes you would rather have more control.

I'm thinking of adding a mx.pad layer based on tf.pad. Let me know if you have thoughts on this layer, otherwise will just make a PR in the next few days.

[piiswrong]

@sbodenstein Sounds good! Thanks.
Lets try to be consistent with the other packages

[xlvector]

@sbodenstein Thanks very much!

[sbodenstein]

@piiswrong, @xlvector: its actually a larger project than I thought at first to write an efficient pad layer for n-dimensional tensors (and won't have time to do it).

The cases of interest are almost always for images (in which case only two of the four dimensions are padded), so I propose a layer mx.image_pad. This takes a tensor (n, c, h, w), and parameters:

• mode: padding mode, can be "constant" (pad with 0), "reflect", "replicate" (same as Torch)
• padding: a matrix [[pad_top, pad_bottom], [pad_left, pad_right]]

In the future, a layer mx.pad could be written that deals with the general case (and will require a more verbose syntax for the common image case). This layer still replicates the three torch layers SpatialZeroPadding, SpatialReflectionPadding and SpatialReplicationPadding.

What are your thoughts?

[piiswrong]

Let's stick with mx.pad and make the interface support general padding. Then you can check if you actually only need to pad the last 2 dims and fail if otherwise

[sbodenstein]

@piiswrong: ok. One more thing: can arguments to mxnet symbols be tuples of tuples? If not, will just use a single tuple to represent padding.

[piiswrong]

No. Only support tuple of unsigned. You can add support to your own type by writing parsers though

[sbodenstein]

Would you recommend that I just use a tuple of unsigned for the padding? Or add a parser? (if so, in which file are parsers defined?)

[piiswrong]

Please use tuple if it's enough. just to be consistent

[sbodenstein]

@piiswrong: I'm having some difficulties. I need to use atomicAdd, but it doesn't work for double + half. I can easily overload the double case, but having difficulties with the half case.

Is it better just to avoid using DType, and use floats instead?

[piiswrong]

You can use DType but only instantiate for float and double when creating
layers

On Tue, Oct 25, 2016 at 10:54 PM, Sebastian Bodenstein <
notifications@github.com> wrote:

@piiswrong https://github.com/piiswrong: I'm having some difficulties.
I need to use atomicAdd, but it doesn't work for double + half. I can
easily overload the double case, but having difficulties with the half
case.

Is it better just to avoid using DType, and use floats instead?

—
You are receiving this because you were mentioned.
Reply to this email directly, view it on GitHub
#3556 (comment), or mute
the thread
https://github.com/notifications/unsubscribe-auth/AAiudDCl0r92ipma3jXg5vJCs_oL89HWks5q3ur6gaJpZM4KZXSA
.

[erikbeall]

@sbodenstein: did you make any further progress on adding reflect and replicate to padding? I'd like to help if that would be useful.

[sbodenstein]

I haven't added reflection padding. Replicate is working working on the dims supported. If you would like to help, there are two projects:

• Currently, only padding for the spatial dimensions of 2d + 3d images is supported (when images are represented as [c,h,w] format). It is good having a fast code path for these, but it would be nice to support padding for arbitrary tensor shapes (like tensorflow supports)
• add reflection padding support

[erikbeall]

@sbodenstein: so replicate is working, that is wonderful wonderful :) I'd like to help with adding reflection padding support first. Please let me know where your replicate code is, and I can work on implementing reflection in a consistent way, test/compare with tensorflow. Are you working from https://github.com/sbodenstein/mxnet/tree/master/src/operator?