Circular Convolution Function via circular padding #17240

NarineK · 2019-02-19T02:46:14Z

Summary: Added circular padding in addition to zero padding to Conv1D, Conv2D and Conv3D based on the solution suggested in: #3858

Differential Revision: D14126416

ezyang

This is a really good start. I think the big things for me are:

Getting better test coverage for padding > 1 case. I don't really care if you have a reference implementation, but we should exercise it.
Fixing the error checking in circular_pad. My suggestion is to fold it into the existing pad function.

torch/nn/modules/conv.py

test/common_nn.py

torch/nn/functional.py

apaszke

Could we just have F.pad_circular to avoid having a blow up for every single dimensionality?

ezyang · 2019-02-19T19:15:46Z

@apaszke I don't know how to write a generalized circular pad that works for arbitrary dimensionality, without writing the kernel out (which may eventually be a good idea, but I wouldn't block this PR on it.) Any ideas? :)

apaszke · 2019-02-19T21:11:04Z

I haven't actually thought about that, but it seems like the implementations for 1d and 3d look the same modulo the number of colons (which are equivalent to slice(None)), so we should be able to do this programmatically. Also, it would be good to have better examples in the docstrings, because they don't fully explain the semantics I think. Is it correct to say that in the 2D case we add the last row at the top, the first row at the bottom, the left column on the right, and the right column on the left? What happens with the indices that are not covered by those elements?

NarineK · 2019-02-20T00:05:55Z

Thank you @apaszke and @ezyang ! I can play with the slice(None) to find a more elegant implementation for the dimensionality.
With respect to What happens with the indices that are not covered by those elements? The other indices do not change or effect anything. Here is an example from #3858 for padding=1
Before circular padding
[[0 1 2]
[3 4 5]
[6 7 8]]

After circular padding
[[8 6 7 8 6]
[2 0 1 2 0]
[5 3 4 5 3]
[8 6 7 8 6]
[2 0 1 2 0]]

NarineK · 2019-02-21T02:04:21Z

@ezyang and @apaszke , I've just pushed the implementation of pad_circular which works for arbitrary dimensions.
6832149#diff-c66288b9ce36978f377a1a20d32ec53dR2993
The common_nn.py tests passed but test_jit does not like input[slice(None)] in general.
I will address other comments in a separate commit.

NarineK · 2019-02-22T17:38:59Z

After talking to jit team we came to the conclusion that currently there is no support for slice(None). They opened the following feature to work on: #17389

I'll revert back my changes and update docstring examples.

cc: @ezyang, @apaszke

test/common_nn.py

ezyang · 2019-02-25T04:44:51Z

@pytorchbot rebase this please

pytorchbot · 2019-02-25T04:44:53Z

There's nothing to do! This branch is already up to date with master (d76b939).

(To learn more about this bot, see Bot commands.)

test/common_nn.py

NarineK · 2019-02-27T00:25:05Z

torch/nn/modules/conv.py

We could also have a separate method to split left and right paddings.

NarineK · 2019-03-01T22:20:56Z

@pytorchbot rebase this please

pytorchbot · 2019-03-01T22:20:59Z

Sorry, I can't merge this because there are conflicts. To merge this yourself, run the commands below:

git fetch origin master
git fetch [email protected]:NarineK/pytorch.git export-D14126416
git checkout FETCH_HEAD
git merge origin/master
git push [email protected]:NarineK/pytorch.git HEAD:export-D14126416

(To learn more about this bot, see Bot commands.)

NarineK · 2019-03-09T02:33:39Z

Do you guys have any other comments on this PR ?
cc: @ezyang , @soumith

Summary: Pull Request resolved: pytorch#17240 Added circular padding in addition to zero padding to Conv1D, Conv2D and Conv3D based on the solution suggested in: pytorch#3858 Differential Revision: D14126416 fbshipit-source-id: 72cdb69b8c22e33b0ba9ae515c4d119f69ea45d9

fmassa · 2019-03-18T19:45:20Z

torch/nn/functional.py

+
+
+@weak_script
+def pad_circular(input, padding):


question: why didn't we add this functionality to F.pad?

Hi @fmassa! That was a suggestion to have padding functionally at one place.
cc: @soumith , @ezyang

fmassa · 2019-03-18T19:46:17Z

torch/nn/modules/conv.py

+            return F.conv1d(F.pad(input, expanded_padding, mode='circular'),
+                            self.weight, self.bias, self.stride,
+                            _single(0), self.dilation, self.groups)
        return F.conv1d(input, self.weight, self.bias, self.stride,


do we want to have the functional interface of conv also support the padding mode?

In the documentation it appears padding_mode is supported, but when I try to use it tells me it is an unexpected argument. I will do the manual padding, as in the module, but it would be convenient if the operation is supported.
In my case, my filter is a tensor, not a parameter, and I would prefer not to use a module.

Thanks!

ezyang · 2019-07-09T04:23:11Z

What version of PyTorch are you running? Excerpts from Javier Zazo's message of 2019-07-05 02:41:34 -0700:

…

jzazo commented on this pull request. > @weak_script_method def forward(self, input): + if self.padding_mode == 'circular': + expanded_padding = ((self.padding[0] + 1) // 2, self.padding[0] // 2) + return F.conv1d(F.pad(input, expanded_padding, mode='circular'), + self.weight, self.bias, self.stride, + _single(0), self.dilation, self.groups) return F.conv1d(input, self.weight, self.bias, self.stride, In the documentation it appears padding_mode is supported, but when I try to use it tells me it is an unexpected argument. I will do the manual padding, as in the module, but it would be convenient if the operation is supported. In my case, my filter is a tensor, not a parameter, and I would prefer not to use a module. Thanks!

jzazo · 2019-07-10T14:21:59Z

I am running 1.1.0. Thanks!

thomasahle · 2022-09-14T15:51:59Z

If I try to run
F.pad(torch.arange(4), (0,1), mode='circular')
or
F.pad(torch.arange(9).reshape(3,3), (0,1,1,0), mode='circular')
I get the error
NotImplementedError: Only 2D, 3D, 4D, 5D padding with non-constant padding are supported for now.
(Torch v 1.11)
Am I doing this wrong?

NarineK force-pushed the export-D14126416 branch from b09b180 to a88bb61 Compare February 19, 2019 07:28

NarineK force-pushed the export-D14126416 branch from a88bb61 to eb34f09 Compare February 19, 2019 07:52

ezyang requested changes Feb 19, 2019

View reviewed changes

apaszke reviewed Feb 19, 2019

View reviewed changes

NarineK force-pushed the export-D14126416 branch from eb34f09 to 6832149 Compare February 21, 2019 01:59

NarineK commented Feb 23, 2019

View reviewed changes

test/common_nn.py Outdated Show resolved Hide resolved

NarineK force-pushed the export-D14126416 branch from 6832149 to 2243484 Compare February 24, 2019 06:12

NarineK force-pushed the export-D14126416 branch from 2243484 to 36b07f3 Compare February 24, 2019 06:19

NarineK force-pushed the export-D14126416 branch from 36b07f3 to dab9675 Compare February 24, 2019 07:36

ezyang reviewed Feb 25, 2019

View reviewed changes

test/common_nn.py Outdated Show resolved Hide resolved

NarineK force-pushed the export-D14126416 branch from dab9675 to 93d1679 Compare February 26, 2019 17:53

NarineK force-pushed the export-D14126416 branch from 93d1679 to 12f0925 Compare February 26, 2019 22:29

NarineK commented Feb 27, 2019

View reviewed changes

torch/nn/modules/conv.py Outdated

Copy link

Contributor Author

NarineK Feb 27, 2019

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

We could also have a separate method to split left and right paddings.

NarineK force-pushed the export-D14126416 branch from 12f0925 to 72f19e2 Compare March 1, 2019 06:55

NarineK force-pushed the export-D14126416 branch from 72f19e2 to b46d343 Compare March 2, 2019 02:53

NarineK force-pushed the export-D14126416 branch from b46d343 to 49d20e4 Compare March 2, 2019 07:38

ezyang approved these changes Mar 18, 2019

View reviewed changes

NarineK force-pushed the export-D14126416 branch from 49d20e4 to 0a877ca Compare March 18, 2019 18:14

facebook-github-bot closed this in 670f509 Mar 18, 2019

pytorchbot added the merged label Mar 18, 2019

fmassa reviewed Mar 18, 2019

View reviewed changes

jzazo mentioned this pull request Jul 5, 2019

[feature request] Circular Convolution Function (Convolution with circular padding) #6124

Closed

Chillee mentioned this pull request Jul 25, 2019

Wrap padding #3858

Closed

tovacinni mentioned this pull request Aug 28, 2019

Error trying to run code nv-tlabs/GSCNN#6

Closed

h3jia mentioned this pull request Jul 21, 2023

Circular padding for 2-dim images does not work as expected #105749

Closed

Circular Convolution Function via circular padding #17240

Circular Convolution Function via circular padding #17240

Uh oh!

Conversation

NarineK commented Feb 19, 2019

Uh oh!

ezyang left a comment

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

Uh oh!

apaszke left a comment

Choose a reason for hiding this comment

Uh oh!

ezyang commented Feb 19, 2019

Uh oh!

apaszke commented Feb 19, 2019

Uh oh!

NarineK commented Feb 20, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

NarineK commented Feb 21, 2019

Uh oh!

NarineK commented Feb 22, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

ezyang commented Feb 25, 2019

Uh oh!

pytorchbot commented Feb 25, 2019

Uh oh!

Uh oh!

NarineK Feb 27, 2019

Choose a reason for hiding this comment

Uh oh!

NarineK commented Mar 1, 2019

Uh oh!

pytorchbot commented Mar 1, 2019

Uh oh!

NarineK commented Mar 9, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

fmassa Mar 18, 2019

Choose a reason for hiding this comment

Uh oh!

NarineK Mar 18, 2019

Choose a reason for hiding this comment

Uh oh!

fmassa Mar 18, 2019

Choose a reason for hiding this comment

Uh oh!

jzazo Jul 5, 2019

Choose a reason for hiding this comment

Uh oh!

ezyang commented Jul 9, 2019 via email

Uh oh!

jzazo commented Jul 10, 2019

Uh oh!

thomasahle commented Sep 14, 2022

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

8 participants

NarineK commented Feb 20, 2019 •

edited

Loading

NarineK commented Feb 22, 2019 •

edited

Loading

NarineK commented Mar 9, 2019 •

edited

Loading