avg_pool2d avg_pool3d for LongTensor #22433

ifedan · 2019-07-02T02:22:20Z

Generate avg_pool2d/avg_pool3d for LongTensor for CPU.
Added divisor_override parameter.

facebook-github-bot

@ifedan has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

ifedan · 2019-07-03T18:51:12Z

gchanan · 2019-07-15T18:20:51Z

aten/src/ATen/Dispatch.h


+template<>
+struct ScalarTypeToCType<at::ScalarType::Long> {
+  using type = int64_t ;


nit: spacing.

gchanan · 2019-07-15T19:06:22Z

aten/src/ATen/native/cuda/AveragePool3d.cu

  int padT, int padH, int padW,
  bool count_include_pad,
-  int offsetZ)
+  int offsetZ, int divisor_override)


I think I asked this before, but is there a reason that optional doesn't work herE?

it's not allowed to call a host function c10::optional::value from a global function

ah, you might be able to fix that by making value a C10_HOST_DEVICE function, but this seems fine.

gchanan · 2019-07-15T19:08:41Z

torch/nn/functional.py

        to compute the output shape. Default: ``False``
    count_include_pad: when True, will include the zero-padding in the
        averaging calculation. Default: ``True``
+    divisor_override: if specified, it will be used as divisor, otherwise


I don't think this is accurate, count_include_pad is also potentially used.

gchanan · 2019-07-15T19:08:46Z

torch/nn/functional.py

        to compute the output shape
    count_include_pad: when True, will include the zero-padding in the
        averaging calculation
+    divisor_override: if specified, it will be used as divisor, otherwise


gchanan · 2019-07-15T19:10:07Z

torch/onnx/symbolic_opset9.py

+    def symbolic_fn(g, input, kernel_size, stride, padding, ceil_mode, count_include_pad, divisor_override=None):
        if ceil_mode and input.type().kind() != "CompleteTensorType":
            return _unimplemented(name, "input size not accesible")
+        if divisor_override and divisor_override.node().kind() != 'prim::Constant':


shouldn't this check if it's None type or something? This looks like a constant would pass but not actually be passed to the operator.

It's similar to ceil_mode. In case if the value will be provided, type().kind() == "ONNX:Constant"

I don't see what ceil_mode has to do with it. @houseroad can you look at this?

I checked two cases: divisor_override not provided, divisor_override is provided. And if divisor_override is not provided it will converted to prim::Constant, otherwise it will be ONNX:Constant.

what is the meaning of divisor_override? We would like to check the value of divisor_override to determine whether we can express it in ONNX or not. If not, call _unimplemented is the right choice.

ifedan · 2019-07-16T14:13:21Z

@pytorchbot retest this please

facebook-github-bot

@ifedan has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

ifedan · 2019-07-16T19:49:43Z

@pytorchbot retest this please

gchanan

would prefer if @houseroad would take a look at the onnx stuff.

facebook-github-bot

@ifedan has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

ifedan · 2019-07-17T16:28:10Z

@pytorchbot retest this please

ifedan · 2019-07-17T18:10:33Z

@ailzhang can we do the changes on XLA side? We blocked on rocm failures

ifedan · 2019-07-17T18:20:49Z

@ailzhang I also can create PR for review

ailzhang · 2019-07-17T18:23:07Z

@ifedan This should be easy to follow I think. Feel free to land when it's ready. I will prepare a fix on xla side. :)

ifedan · 2019-07-17T18:29:46Z

@ifedan This should be easy to follow I think. Feel free to land when it's ready. I will prepare a fix on xla side. :)

@ailzhang PR is ready to land, but I'm not able to do it, because XLA test is failing. So should wait for the fix first

@ailzhang pytorch/xla#831

facebook-github-bot

@ifedan has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

facebook-github-bot · 2019-07-18T04:05:55Z

@ifedan merged this pull request in c2df54d.

Summary: Generate avg_pool2d/avg_pool3d for LongTensor for CPU. Added divisor_override parameter. Pull Request resolved: pytorch/pytorch#22433 Differential Revision: D16108809 Pulled By: ifedan fbshipit-source-id: 8de7ff585a0479702cceafb5ccf9dfea62a9cc50

avg_pool2d avg_pool3d for LongTensor

9312755

ifedan added 6 commits July 1, 2019 23:11

avg_pool2d avg_pool3d for LongTensor

7362978

Fix issues with CUDA

c54d3bc

Fix issues with CUDA

8f9e44b

Fix issues with CUDA

c877b40

Fix issues with CUDA

ca9fb48

Fix issues with CUDA

6aefe58

ifedan mentioned this pull request Jul 2, 2019

sum_pool2d/sum_pool3d for LongTensor #20731

Closed

ifedan requested a review from gchanan July 2, 2019 18:46

Fix for Functional binding with optional argument

c8b7fe5

ifedan requested review from ebetica, goldsborough and yf225 as code owners July 3, 2019 04:01

pytorchbot added the module: cpp Related to C++ API label Jul 3, 2019

ifedan added 3 commits July 3, 2019 00:04

Fix for Functional binding with optional argument

0694a16

Fix for Functional binding with optional argument

020a5fa

Fix for Functional binding with optional argument

6af6e82

facebook-github-bot reviewed Jul 3, 2019

View reviewed changes

Reverted changes for Functional

7762532

ifedan removed request for ebetica, goldsborough and yf225 July 3, 2019 18:44

Some code refactoring

3820a3e

ifedan added 2 commits July 15, 2019 12:24

Fix ONNX test

5976636

Fix ONNX test

5cf1672

ifedan requested a review from gchanan July 15, 2019 17:46

Fix based on PR

c41210a

gchanan requested changes Jul 15, 2019

View reviewed changes

Fix based on PR

086164d

gchanan mentioned this pull request Jul 15, 2019

Enable Add, sub, mul, and div on CPU for bfloat16 type. #22851

Closed

ifedan added 2 commits July 15, 2019 17:33

Fix based on PR

4ed181f

Merge branch 'master' into avg_pool2d

464c634

ifedan requested a review from gchanan July 15, 2019 22:35

Merge branch 'master' into avg_pool2d

7229f13

facebook-github-bot reviewed Jul 16, 2019

View reviewed changes

gchanan approved these changes Jul 16, 2019

View reviewed changes

facebook-github-bot reviewed Jul 17, 2019

View reviewed changes

Merge branch 'master' into avg_pool2d

eca2882

ailzhang mentioned this pull request Jul 17, 2019

PT BREAK: avg_pool2d avg_pool3d for LongTensor pytorch/xla#830

Closed

facebook-github-bot reviewed Jul 17, 2019

View reviewed changes

facebook-github-bot closed this in c2df54d Jul 18, 2019

facebook-github-bot added the merged label Jul 18, 2019

mruberry added the Merged label Oct 28, 2020

avg_pool2d avg_pool3d for LongTensor #22433

avg_pool2d avg_pool3d for LongTensor #22433

Uh oh!

Conversation

ifedan commented Jul 2, 2019

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

ifedan commented Jul 3, 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!

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!

ifedan commented Jul 16, 2019

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

ifedan commented Jul 16, 2019

Uh oh!

gchanan left a comment

Choose a reason for hiding this comment

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

ifedan commented Jul 17, 2019

Uh oh!

ifedan commented Jul 17, 2019

Uh oh!

ifedan commented Jul 17, 2019

Uh oh!

ailzhang commented Jul 17, 2019

Uh oh!

ifedan commented Jul 17, 2019 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

facebook-github-bot left a comment

Choose a reason for hiding this comment

Uh oh!

facebook-github-bot commented Jul 18, 2019

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

9 participants

ifedan commented Jul 17, 2019 •

edited

Loading