Port convtranspose2d #20994

xmnlab · 2019-05-27T19:40:42Z

this PR will resolve partially #18353

xmnlab · 2019-06-01T02:24:17Z

@pearu do you have any tip here?

test/test_nn.py::TestNN::test_Conv1d_dilated PASSED                      [  3%]
test/test_nn.py::TestNN::test_grad_conv1d_input PASSED                   [  6%]
test/test_nn.py::TestNN::test_conv_double_backward PASSED                [  9%]
test/test_nn.py::TestNN::test_Conv1d_pad2 PASSED                         [ 12%]
test/test_nn.py::TestNN::test_spectral_norm_dim PASSED                   [ 15%]
test/test_nn.py::TestNN::test_Conv2d_depthwise PASSED                    [ 18%]
test/test_nn.py::TestNN::test_Conv2d_pad2circular PASSED                 [ 21%]
test/test_nn.py::TestNN::test_Conv1d_stride PASSED                       [ 24%]
test/test_nn.py::TestNN::test_Conv1d_stride1_pad2circular PASSED         [ 27%]
test/test_nn.py::TestNN::test_Conv2d_depthwise_dilated PASSED            [ 30%]
test/test_nn.py::TestNN::test_Conv1d_pad2size1 PASSED                    [ 33%]
test/test_nn.py::TestNN::test_Conv2d_depthwise_padded PASSED             [ 36%]
test/test_nn.py::TestNN::test_Conv2d_no_bias PASSED                      [ 39%]
test/test_nn.py::TestNN::test_conv_noncontig_weights PASSED              [ 42%]
test/test_nn.py::TestNN::test_grad_conv2d_input PASSED                   [ 45%]
test/test_nn.py::TestNN::test_ConvTranspose1d FAILED                     [ 48%]
test/test_nn.py::TestNN::test_ConvTranspose2d FAILED                     [ 51%]
test/test_nn.py::TestNN::test_Conv2d PASSED                              [ 54%]
test/test_nn.py::TestNN::test_Conv2d_padding PASSED                      [ 57%]
test/test_nn.py::TestNN::test_ConvTranspose1d_groups FAILED              [ 60%]
test/test_nn.py::TestNN::test_Conv2d_depthwise_strided PASSED            [ 63%]
test/test_nn.py::TestNN::test_Conv2d_groups PASSED                       [ 66%]
test/test_nn.py::TestNN::test_ConvTranspose1d_no_bias FAILED             [ 69%]
test/test_nn.py::TestNN::test_Conv2d_dilated PASSED                      [ 72%]
test/test_nn.py::TestNN::test_ConvTranspose2d_dilated FAILED             [ 75%]
test/test_nn.py::TestNN::test_conv_double_backward_groups PASSED         [ 78%]
test/test_nn.py::TestNN::test_ConvTranspose2d_groups FAILED              [ 81%]
test/test_nn.py::TestNN::test_ConvTranspose2d_output_size PASSED         [ 84%]
test/test_nn.py::TestNN::test_conv_double_backward_no_bias PASSED        [ 87%]
test/test_nn.py::TestNN::test_Conv2d_groups_thnn PASSED                  [ 90%]
test/test_nn.py::TestNN::test_conv_double_backward_stride PASSED         [ 93%]
test/test_nn.py::TestNN::test_Conv2d_depthwise_with_multiplier PASSED    [ 96%]
test/test_nn.py::TestNN::test_ConvTranspose2d_no_bias FAILED             [100%]

skrah · 2019-06-03T21:46:53Z

aten/src/ATen/native/native_functions.yaml

One of the CI errors is '_thnn_conv_transpose2d_backward' is not a member of 'at::native::legacy::cuda'.

I'm pretty sure that if you do CPU/CUDA separately, you cannot remove thnn_conv_transpose2d_backward from native_functions.yaml until CUDA is ported as well.

This is why I recommend doing CPU/CUDA together.

I agree with @skrah .
However, to just test on CPU, use USE_CUDA=0 env variable when building. This allows finishing CPU porting before starting the CUDA porting.

@skrah @pearu thanks so much for the feedback! I will try to find a workaround for that ... not sure if it would be a good approach to port cpu and gpu in the same PR. I needed to move some steps back and rewrote again my code ... locally, the problematic tests passed. I will try to fix this GPU problem to move forward this PR.

I meant either leave thnn_conv_transpose2d_backward exactly as before for CUDA and use the new name for CPU only or port CPU/CUDA together.

I've never tried leaving the old name for CUDA though, I always took the second approach.

I see .. thanks for the explanation @skrah ! I move steps back and worked on my first code ... and just realized that I already started the porting for cuda ... so I will probably port for gpu for 2d also im2col and col2im .. but I think the comparison between files will be easy ... I am not changing the structure here.

xmnlab · 2019-06-07T21:51:01Z

aten/src/ATen/cuda/CUDABlas.h

from https://github.com/pytorch/pytorch/pull/20983/files#diff-bf2bf1d6148eb27d3e3e29031655b1deR1

thanks @pearu

OK, I guess if we merge this first, @pearu will handle the merge conflicts in his diff

ezyang · 2019-06-18T14:48:02Z

A remark that applies to the whole cluster of ports that are happening here: as we are porting, we are renaming the kernels from thnn_conv_transpose2d to conv_transpose2d. I'm not sure this is actually a renaming we should do. The name conv_transpose2d sounds very general: it is "THE" 2d transposed convolution. But actually it gives you a very specific (somewhat slow) implementation. It would be good if this was made clear. We shouldn't call them thnn because that implies legacy, but it would be good if we had some name that made it clear that it was a specific implementation, much in the same way cudnn_conv_transpose2d makes it clear we are using the cuDNN implementation.

ezyang · 2019-06-18T14:53:45Z

tools/autograd/derivatives.yaml

- name: thnn_conv_transpose2d_forward(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias, int[2] stride, int[2] padding, int[2] output_padding, int[2] dilation) -> (Tensor output, Tensor columns, Tensor ones)
-  self, weight, bias: thnn_conv_transpose2d_backward(grad, self, weight, kernel_size, stride, padding, output_padding, dilation, columns, ones, grad_input_mask)
+- name: conv_transpose2d(Tensor self, Tensor weight, int[2] kernel_size, Tensor? bias=None, int[2] stride=1, int[2] padding=0, int[2] output_padding=0, int[2] dilation=1) -> Tensor
+  self, weight, bias: conv_transpose2d_backward(grad, self, weight, kernel_size, stride, padding, output_padding, dilation, empty_like(grad), empty_like(grad), grad_input_mask)


I won't block land on this, but even better would be just to axe these arguments from the function entirely and have them be created inside the function itself.

facebook-github-bot

@ezyang is landing this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

xmnlab · 2019-06-18T15:04:03Z

@ezyang I will work on a PR for the 3d so I can change the name there. do you have any suggestion for the name? does at_conv_transpose2d looks good to you?

ezyang · 2019-06-18T15:11:20Z

at_ or aten_ seems as good a name as any. Other possibilities are slow_ (cuz it's slow) or naive_. Up to you! Excerpts from Ivan Ogasawara's message of 2019-06-18 08:05:18 -0700:

…

@ezyang I will work on a PR for the 3d so I can change the name there. do you have any suggestion for the name? does `at_conv_transpose2d` looks good to you?

xmnlab · 2019-06-19T13:09:29Z

@ezyang just to sure, is there anything missing in this PR? Or could I apply the recommendations you pointed in a follow up PR?

ezyang · 2019-06-19T13:42:00Z

I apparently need to adjust our internal build system to successfully land this

xmnlab · 2019-06-19T14:48:37Z

oh I see .. let me know if I can help in anyway

ezyang · 2019-06-19T20:12:06Z

For the record, it looks like the error was we needed to delete aten/src/THCUNN/SpatialFullConvolution.cu

xmnlab · 2019-06-19T20:24:12Z

Ok thanks @ezyang for the feedback in the next PR, I will ensure that the corresponded file for 3d will be deleted. thanks!

xmnlab · 2019-06-19T23:15:07Z

@ezyang just to check, should I remove this file for 2d in this PR?

ezyang · 2019-06-20T01:02:24Z

Nope, please don't. Excerpts from Ivan Ogasawara's message of 2019-06-19 16:16:09 -0700:

…

@ezyang just to check, should I remove this file for 2d in this PR?

Summary: this PR will resolve partially pytorch/pytorch#18353 Pull Request resolved: pytorch/pytorch#20994 Differential Revision: D15876052 Pulled By: ezyang fbshipit-source-id: 5896e0cbb656d0530e39fd681808adc685841b37

facebook-github-bot · 2019-06-20T15:39:55Z

@ezyang merged this pull request in f8f583c.

pytorchbot added module: cpu CPU specific problem (e.g., perf, algorithm) module: internals Related to internal abstractions in c10 and ATen module: nccl Problems related to nccl support module: onnx Related to torch.onnx module: operators module: third_party labels May 27, 2019

xmnlab force-pushed the 18353-convtranspose2d branch from 5cdbfd9 to a83dfda Compare May 29, 2019 00:18

pytorchbot added the caffe2 label May 29, 2019

xmnlab force-pushed the 18353-convtranspose2d branch 2 times, most recently from 4742369 to ba14ded Compare May 31, 2019 04:56

xmnlab requested review from pearu and skrah June 3, 2019 14:17

skrah reviewed Jun 3, 2019

View reviewed changes

xmnlab force-pushed the 18353-convtranspose2d branch from 3107b28 to 06670f7 Compare June 4, 2019 23:10

pytorchbot added module: build Build system issues module: cuda Related to torch.cuda, and CUDA support in general labels Jun 4, 2019

ezyang added the open source label Jun 5, 2019

xmnlab force-pushed the 18353-convtranspose2d branch 2 times, most recently from 0e28c4a to a23b8fb Compare June 7, 2019 00:56

xmnlab changed the title ~~[wip] 18353 convtranspose2d~~ 18353 convtranspose2d Jun 7, 2019

xmnlab commented Jun 7, 2019

View reviewed changes

xmnlab marked this pull request as ready for review June 8, 2019 00:40

xmnlab force-pushed the 18353-convtranspose2d branch 2 times, most recently from 2ade5a4 to 233df28 Compare June 8, 2019 17:37

pytorchbot added the module: cublas Problem related to cublas support label Jun 8, 2019

xmnlab force-pushed the 18353-convtranspose2d branch from 233df28 to 57e2312 Compare June 10, 2019 03:35

jeffreyksmithjr added the triaged This issue has been looked at a team member, and triaged and prioritized into an appropriate module label Jun 10, 2019

xmnlab force-pushed the 18353-convtranspose2d branch 2 times, most recently from 9a75e13 to 9e8ad74 Compare June 12, 2019 14:53

xmnlab added 2 commits June 17, 2019 15:25

Port SpatialFullDilatedConvolution

08c0e11

Fixed some issues with im2col col2im

c1622eb

xmnlab force-pushed the 18353-convtranspose2d branch from 9e8ad74 to c1622eb Compare June 17, 2019 22:10

xmnlab requested a review from ezyang June 18, 2019 14:38

ezyang changed the title ~~18353 convtranspose2d~~ Port convtranspose2d Jun 18, 2019

ezyang reviewed Jun 18, 2019

View reviewed changes

ezyang approved these changes Jun 18, 2019

View reviewed changes

facebook-github-bot reviewed Jun 18, 2019

View reviewed changes

facebook-github-bot closed this in f8f583c Jun 20, 2019

xmnlab deleted the 18353-convtranspose2d branch June 20, 2019 14:16

facebook-github-bot added the merged label Jun 20, 2019

ailzhang mentioned this pull request Jun 20, 2019

fix breakage PT#20994 pytorch/xla#766

Merged

xmnlab mentioned this pull request Jun 23, 2019

porting convtranspose3d to ATen #22019

Closed

ngimel mentioned this pull request Jul 30, 2019

THCUNN convolutions fail with half inputs/weights #23545

Closed

mruberry added the Merged label Oct 28, 2020

Port convtranspose2d #20994

Port convtranspose2d #20994

Uh oh!

Conversation

xmnlab commented May 27, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

xmnlab commented Jun 1, 2019

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

ezyang commented Jun 18, 2019

Uh oh!

Choose a reason for hiding this comment

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

xmnlab commented Jun 18, 2019

Uh oh!

ezyang commented Jun 18, 2019 via email

Uh oh!

xmnlab commented Jun 19, 2019

Uh oh!

ezyang commented Jun 19, 2019

Uh oh!

xmnlab commented Jun 19, 2019

Uh oh!

ezyang commented Jun 19, 2019

Uh oh!

xmnlab commented Jun 19, 2019

Uh oh!

xmnlab commented Jun 19, 2019

Uh oh!

ezyang commented Jun 20, 2019 via email

Uh oh!

facebook-github-bot commented Jun 20, 2019

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

8 participants

xmnlab commented May 27, 2019 •

edited

Loading