fix issue #1549, make and operator correct by stegben · Pull Request #1556 · pytorch/pytorch

stegben · 2017-05-15T06:27:41Z

A quick attempt to fix #1549

apaszke · 2017-05-15T07:27:31Z

Thanks, but it's not a long-term solution. A better fix for this issue is to expose bitwise operations from TH and THC (in our C extension).

stegben · 2017-05-15T07:34:10Z

OK, let me try implementing it in C modules (in torch/csrc/generic/Tensor.cpp right?)

apaszke · 2017-05-15T07:38:10Z

The only C file you should need to modify is torch/csrc/generic/methods/Tensor.cwrap

apaszke · 2017-05-15T07:42:12Z

Eh I can't update the comment on my phone. You need to add new cwrap declarations (parts in [[ ]]) for the bitwise functions, call them __and__, etc. (using python_name) and remove them from the Python file.

stegben · 2017-05-15T07:47:37Z

Got it.
By the way, Is THPTensor_ an alias for both THTensor and THCTensor_?

apaszke · 2017-05-15T08:15:32Z

THP is for objects that appear in Python. THTensor expands to THTensor when compiling CPU objects and to THCTensor when building CUDA objects (macro magic).

stegben · 2017-05-15T12:17:47Z

Fixed. Not sure if I use the correct method. I only found bitand in THTensorMath.h that matches what we need.

apaszke

That looks great!

If you exposed other bit functions (or, xor, shifts), it would resolve #768 too. If you don't want to do it I can merge this and add that later myself.

torch/csrc/generic/methods/Tensor.cwrap

test/test_torch.py

stegben · 2017-05-15T12:36:21Z

I could add other bitwise operations. WIP

stegben · 2017-05-15T12:59:10Z

Done

apaszke

One last thing that needs a fix and it's good to go

torch/tensor.py

soumith · 2017-05-15T17:11:45Z

@pytorchbot test this please

soumith · 2017-05-15T18:49:19Z

thank you so much for the patch @stegben !

stegben · 2017-05-16T01:55:39Z

It's my honor to commit to such a great library :)

And many thanks to @apaszke for these detailed instructions.

hmansell · 2017-05-16T13:57:47Z

This is awesome @stegben, thanks!

@soumith / @apaszke - isn't invert still incorrect? I don't think the TH/THC has a bitwise not, but we can xor with 255.

apaszke · 2017-05-16T15:01:10Z

You're right I forgot that there's invert.

BTW how do bitwise operations work with floats/doubles?

hmansell · 2017-05-16T15:03:21Z

I suspect they don't work in any logical way and should only be enabled for ByteTensor (or any other integers we might have).

…1556) * Propagate root domain mappings from rfactor to root domains in ComputeAtRootDomainMap The main purpose of ComputeAtRootDomainMap is to find unmappable domains for comptueAt. This analyais is done by traversing a fusion in a backward direction. Currently, the traversal only visits arithmetic expressions, so information propagation is done from consumer tensors to producer tensors. This propagation is also required from rfactor domains to root domains. Previously it doesn't really matter as rfactor is limited reduction domains, but that's not the case with view. This change also means that ComputeAtRootDomain does not guarantee one-to-one mappings. For example, ``` tv0: [I0, I1] tv1 = view(tv0); // tv1: [I0*I1/N, N] ``` I.e., the view op is done first merging the two domains of `tv0` and then splitting it by N. Note that both of the two rfactor axes of `tv1` are now mapped with the two axes of `tv0`. Because of this change, `ComputeAtRootDomainMap:mapBestEffort` and other mapping functions between a producer and a consumer that is supposed to return a one-to-one map can fail. `ComputeAtRootDomainMap::getMappableDims` is fine as it just grabs any domain that is mappable. `ComputeAtRootDomainMap::mapConsumerToProducer` and `ComputeAtRootDomainMap::mapProducerToConsumer` were used in `TransformReplay::replayPasC` and `TransformReplay::replayCasP`, but they don't really need to use `ComputeAtRootDomainMap` but just `PairwiseRootDomainMap` is sufficient, so replaed the usages with the pairwise variant.

pytorch#1556) This PR pins sympy==1.12.1 in the .ci/docker/requirements-ci.txt file Also it skips pytorch-nightly installation in docker images Installation of pytorch-nightly is needed to prefetch mobilenet_v2 avd v3 models for some tests. Came from ROCm@85bd6bc Models are downloaded on first use to the folder /root/.cache/torch/hub But pytorch-nightly installation also overrides .ci/docker/requirements-ci.txt settings and upgrades some of python packages (sympy from 1.12.0 to 1.13.0) which causes several 'dynamic_shapes' tests to fail Skip prefetching models affects these tests without any errors (but **internet access required**): - python test/mobile/model_test/gen_test_model.py mobilenet_v2 - python test/quantization/eager/test_numeric_suite_eager.py -k test_mobilenet_v3 Issue ROCm/frameworks-internal#8772 Also, in case of some issues these models can be prefetched after pytorch building and before testing (cherry picked from commit b92b34d)

This PR skips pytorch-nightly installation in docker images Installation of pytorch-nightly is needed to prefetch mobilenet_v2 avd v3 models for some tests. Came from ROCm@85bd6bc Models are downloaded on first use to the folder /root/.cache/torch/hub But pytorch-nightly installation also overrides .ci/docker/requirements-ci.txt settings and upgrades some of python packages (sympy from 1.12.0 to 1.13.0) which causes several 'dynamic_shapes' tests to fail Skip prefetching models affects these tests without any errors (but **internet access required**): - python test/mobile/model_test/gen_test_model.py mobilenet_v2 - python test/quantization/eager/test_numeric_suite_eager.py -k test_mobilenet_v3 Issue ROCm/frameworks-internal#8772 Also, in case of some issues these models can be prefetched after pytorch building and before testing (cherry picked from commit b92b34d) (cherry picked from commit d5608f3)

This PR skips pytorch-nightly installation in docker images Installation of pytorch-nightly is needed to prefetch mobilenet_v2 avd v3 models for some tests. Came from ROCm@85bd6bc Models are downloaded on first use to the folder /root/.cache/torch/hub But pytorch-nightly installation also overrides .ci/docker/requirements-ci.txt settings and upgrades some of python packages (sympy from 1.12.0 to 1.13.0) which causes several 'dynamic_shapes' tests to fail Skip prefetching models affects these tests without any errors (but **internet access required**): - python test/mobile/model_test/gen_test_model.py mobilenet_v2 - python test/quantization/eager/test_numeric_suite_eager.py -k test_mobilenet_v3 Issue ROCm/frameworks-internal#8772 Also, in case of some issues these models can be prefetched after pytorch building and before testing (cherry picked from commit b92b34d) (cherry picked from commit d5608f3) (cherry picked from commit aaa3134)

This PR skips pytorch-nightly installation in docker images Installation of pytorch-nightly is needed to prefetch mobilenet_v2 avd v3 models for some tests. Came from ROCm@85bd6bc Models are downloaded on first use to the folder /root/.cache/torch/hub But pytorch-nightly installation also overrides .ci/docker/requirements-ci.txt settings and upgrades some of python packages (sympy from 1.12.0 to 1.13.0) which causes several 'dynamic_shapes' tests to fail Skip prefetching models affects these tests without any errors (but **internet access required**): - python test/mobile/model_test/gen_test_model.py mobilenet_v2 - python test/quantization/eager/test_numeric_suite_eager.py -k test_mobilenet_v3 Issue ROCm/frameworks-internal#8772 Also, in case of some issues these models can be prefetched after pytorch building and before testing (cherry picked from commit b92b34d) (cherry picked from commit d5608f3) (cherry picked from commit aaa3134) (cherry picked from commit d33dd98)

stegben force-pushed the fix-and-operator branch from 258286d to 2c3c765 Compare May 15, 2017 07:18

stegben force-pushed the fix-and-operator branch from 3ddb25e to ffd19f2 Compare May 15, 2017 12:16

apaszke reviewed May 15, 2017

View reviewed changes

fix issue pytorch#1549, expose bitwise and

0925da6

stegben force-pushed the fix-and-operator branch from ffd19f2 to 0925da6 Compare May 15, 2017 12:36

stegben added 2 commits May 15, 2017 20:42

expose C bitwise or of Tensor

7e0a691

expose C bitwise xor of Tensor

1e0fa8f

apaszke reviewed May 15, 2017

View reviewed changes

torch/tensor.py Outdated

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

This comment was marked as off-topic.

Sign in to view

stegben added 2 commits May 15, 2017 21:42

use built-in method for inplace and, or, xor

1941539

expose C bitwise lshift(ilshift) and rshift(irshift) of Tensor

1549ccd

stegben force-pushed the fix-and-operator branch from 183ee10 to 1549ccd Compare May 15, 2017 13:50

soumith approved these changes May 15, 2017

View reviewed changes

soumith merged commit ba885a1 into pytorch:master May 15, 2017

soumith mentioned this pull request May 15, 2017

Feature Request: Bitwise Operations #768

Closed

stegben mentioned this pull request May 16, 2017

Remove bitwise functions from floating point Tensors #1425

Closed

ezyang added the open source label Jun 24, 2019

Conversation

stegben commented May 15, 2017

Uh oh!

apaszke commented May 15, 2017

Uh oh!

stegben commented May 15, 2017

Uh oh!

apaszke commented May 15, 2017

Uh oh!

apaszke commented May 15, 2017

Uh oh!

stegben commented May 15, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

apaszke commented May 15, 2017

Uh oh!

stegben commented May 15, 2017

Uh oh!

apaszke left a comment

Choose a reason for hiding this comment

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

stegben commented May 15, 2017

Uh oh!

stegben commented May 15, 2017

Uh oh!

apaszke left a comment

Choose a reason for hiding this comment

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

This comment was marked as off-topic.

Uh oh!

soumith commented May 15, 2017

Uh oh!

soumith commented May 15, 2017

Uh oh!

stegben commented May 16, 2017 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

hmansell commented May 16, 2017

Uh oh!

apaszke commented May 16, 2017

Uh oh!

hmansell commented May 16, 2017

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

5 participants

stegben commented May 15, 2017 •

edited

Loading

stegben commented May 16, 2017 •

edited

Loading