platform/upstream/pytorch.git
5 years agoFix math formatting of PairwiseDistance and CosineSimilarity docs and fix math format...
zhiqiang [Tue, 23 Apr 2019 14:20:32 +0000 (07:20 -0700)]
Fix math formatting of PairwiseDistance and CosineSimilarity docs and fix math formatting of CTC loss docs.

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/19534

Differential Revision: D15034011

Pulled By: ezyang

fbshipit-source-id: 60b81c970c919508a57c86fb23edc9f64973117c

5 years agoRevert D15039713: [pytorch][PR] add torch.tensor requires grad
Michael Suo [Tue, 23 Apr 2019 06:10:18 +0000 (23:10 -0700)]
Revert D15039713: [pytorch][PR] add torch.tensor requires grad

Differential Revision:
D15039713

Original commit changeset: 47f1931b6fc4

fbshipit-source-id: fd91ce8ddd6d2f4e0016054dcdc2541dacc0e191

5 years agoBugfix for fusion device check (#19594)
James Reed [Tue, 23 Apr 2019 03:49:21 +0000 (20:49 -0700)]
Bugfix for fusion device check (#19594)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19594

I missed a callsite

Reviewed By: wanchaol

Differential Revision: D15041457

fbshipit-source-id: eef76ad51bee06a56d31b4ab64f19250fe2ad8f0

5 years agoadd torch.tensor requires grad (#19445)
Elias Ellison [Tue, 23 Apr 2019 00:56:51 +0000 (17:56 -0700)]
add torch.tensor requires grad (#19445)

Summary:
Add setting requires_grad = True within torchscript to torch.Tensor

Within constant propagation, we can't insert any constants that require grad.

Also added shape analysis and requires grad analysis to torch.tensor
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19445

Differential Revision: D15039713

Pulled By: eellison

fbshipit-source-id: 47f1931b6fc4a1137c13d80110cc404465bfdf06

5 years agoAdd onnx support for _unique2 operator
Vitaly Fedyunin [Tue, 23 Apr 2019 00:48:43 +0000 (17:48 -0700)]
Add onnx support for _unique2 operator

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/19582

Reviewed By: ezyang, jamesr66a

Differential Revision: D15037375

fbshipit-source-id: 6060476925bf02fa07f852054e06d2107f046e38

5 years agoAutomatic update of fbcode/onnx to 0e8d2bc5e51455c70ef790b9f65aa632ed9bc8a7 (#19568)
Lu Fang [Tue, 23 Apr 2019 00:24:57 +0000 (17:24 -0700)]
update of fbcode/onnx to 0e8d2bc5e51455c70ef790b9f65aa632ed9bc8a7 (#19568)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19568

Previous import was 83dd62659fc07d5b7fa93b5d1c1879f93509c7db

Included changes:
- **[0e8d2bc5](https://github.com/onnx/onnx/commit/0e8d2bc5)**: [Minor need to be in 1.5]Fix an issue in NMS test data which introduce wrong shape. (#1953) <Hector Li>
- **[9346dd5d](https://github.com/onnx/onnx/commit/9346dd5d)**: adding modulus operator (#1874) <Jeff Saremi>
- **[414dbc73](https://github.com/onnx/onnx/commit/414dbc73)**: Fix shape inference for slice (#1950) <Hariharan Seshadri>
- **[6fb0775d](https://github.com/onnx/onnx/commit/6fb0775d)**: Fix shape inference for ConstantOfShape op (#1951) <Ashwini Khade>

Reviewed By: bddppq, zrphercule, benoitsteiner

Differential Revision: D15033070

fbshipit-source-id: f7eb90b142cbdc9bf1600cfd33e5a8df709045fb

5 years agoDon't create FusionGroups for known-CPU producer values (#19342)
James Reed [Mon, 22 Apr 2019 23:54:19 +0000 (16:54 -0700)]
Don't create FusionGroups for known-CPU producer values (#19342)

Summary:
I believe the existing check in FuseGraph was only `false` if PyTorch was built with NO_CUDA=1. Otherwise, we would create fusion groups even if we're on a CPU-only machine running CPU code. This is confusing. Instead I've made it so that the decision to fuse or not is dependent on if the producer Value is a known CPU tensor. If it is, we skip fusion.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19342

Differential Revision: D15038351

Pulled By: jamesr66a

fbshipit-source-id: fce9d83929309a7bf14346833f84b996f3e7f6db

5 years agoExplicitly define supported types (#19516)
Sebastian Messmer [Mon, 22 Apr 2019 23:16:30 +0000 (16:16 -0700)]
Explicitly define supported types (#19516)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19516

Explicitly define types that are supported in kernel inputs and outputs.
Also, this allows us to show much nicer error messages if a user writes kernels with wrong argument types.

Reviewed By: ezyang

Differential Revision: D15020306

fbshipit-source-id: 55ebec81e075e874777acd59aa29a5578fc19ef7

5 years agoIRParser: optionally create name->value map of the parsed IR. (#19551)
Mikhail Zolotukhin [Mon, 22 Apr 2019 23:02:40 +0000 (16:02 -0700)]
IRParser: optionally create name->value map of the parsed IR. (#19551)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19551
ghimport-source-id: e666e3c00786a3b1c747f2dd6e85a48a63bdd69d

Differential Revision: D15028056

Pulled By: ZolotukhinM

fbshipit-source-id: 37e08d6df1d43513748ecfdd8549738eac7ec24e

5 years agoProfiling : Adding Profile Op to provide storage for profiling lambdas
Nikolay Korovaiko [Mon, 22 Apr 2019 22:03:48 +0000 (15:03 -0700)]
Profiling : Adding Profile Op to provide storage for profiling lambdas

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/19365

Differential Revision: D14998968

Pulled By: Krovatkin

fbshipit-source-id: a7f7d1529cbe4e8b30638c6eb8e2ff68f6e114c3

5 years agoStep 5: remove _unique_dim in favor of unique_dim (#18654)
Xiang Gao [Mon, 22 Apr 2019 19:32:14 +0000 (12:32 -0700)]
Step 5: remove _unique_dim in favor of unique_dim (#18654)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18654
ghimport-source-id: 63c84cedc3335719fca4a085fa19bdc57d2bc88a

Differential Revision: D15000635

Pulled By: VitalyFedyunin

fbshipit-source-id: 9e8594622a867a79d8e2b6be96579816aa22ae2d

5 years agoAdd back option to not adjust output batch size (#19442)
Yinghai Lu [Mon, 22 Apr 2019 19:23:05 +0000 (12:23 -0700)]
Add back option to not adjust output batch size (#19442)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19442

For cases like CV, some of ops like transpose and tile will mangle the batch size so that we don't know how to adjust output batch size. In this case, the current solution is just fix the input batch statically and do not adjust output batch size.

Reviewed By: zrphercule

Differential Revision: D15007237

fbshipit-source-id: a21b943a52ee5462d9d7804dfae44360f579f8cf

5 years agoAdd debug logic to c2_ref_test and its helpers (#19359)
Michael Antonov [Mon, 22 Apr 2019 19:04:07 +0000 (12:04 -0700)]
Add debug logic to c2_ref_test and its helpers (#19359)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19359

Even with file IO exception handling, some of the sandcastle c2_ref_tests are still failing in length-check assert, as can be seen here:
https://our.intern.facebook.com/intern/test/844424932589974?ref_report_id=0

This is an attempt to add printing logic to debug what's going on.

Reviewed By: dzhulgakov

Differential Revision: D14966274

fbshipit-source-id: adce6d4780d664c5ef59f9341b6133b0d09324cb

5 years agofix variable shadowing issus (#19567)
Dehua Cheng [Mon, 22 Apr 2019 18:52:12 +0000 (11:52 -0700)]
fix variable shadowing issus (#19567)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19567

fix variable shadowing

Reviewed By: bddppq, wx1988

Differential Revision: D15032114

fbshipit-source-id: 895ea21f22b87db8c7c8684f54fa186d22f24d10

5 years agoAdd manual_seed in script (#19510)
Elias Ellison [Mon, 22 Apr 2019 17:52:28 +0000 (10:52 -0700)]
Add manual_seed in script (#19510)

Summary:
Add manual_seed to torch script.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19510

Reviewed By: suo, driazati

Differential Revision: D15018823

Pulled By: eellison

fbshipit-source-id: d7734a8ad05ba254c0d88abf3fb58c4ce6a4e53b

5 years agoAutomatic update of fbcode/onnx to 83dd62659fc07d5b7fa93b5d1c1879f93509c7db (#19454)
Lu Fang [Mon, 22 Apr 2019 17:37:15 +0000 (10:37 -0700)]
update of fbcode/onnx to 83dd62659fc07d5b7fa93b5d1c1879f93509c7db (#19454)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19454

Previous import was ad7313470a9119d7e1afda7edf1d654497ee80ab

Included changes:
- **[83dd6265](https://github.com/onnx/onnx/commit/83dd6265)**: Add NonMaxSuppression operator (#1703) <Hector Li>
- **[31ca5d6f](https://github.com/onnx/onnx/commit/31ca5d6f)**: add node tests for quantized ops (#1944) <Ashwini Khade>
- **[e6076c1d](https://github.com/onnx/onnx/commit/e6076c1d)**: Fix test stat coverage script (#1948) <Raymond Yang>
- **[ad036405](https://github.com/onnx/onnx/commit/ad036405)**: Add IsInf to detect infinity values (#1884) <Wei-Sheng Chin>

Reviewed By: benoitsteiner

Differential Revision: D15010015

fbshipit-source-id: 4b29de21de60f8e6a2db75309809a4e619c92532

5 years agoGet rid of unnecessary matches_jit_signature: True specifications. (#19549)
Gregory Chanan [Mon, 22 Apr 2019 17:19:17 +0000 (10:19 -0700)]
Get rid of unnecessary matches_jit_signature: True specifications. (#19549)

Summary:
Unstacked version of https://github.com/pytorch/pytorch/pull/19431.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19549

Reviewed By: ezyang

Differential Revision: D15027965

Pulled By: gchanan

fbshipit-source-id: a4456326a999d77d6baeb0edbb1bb5db5208a8f8

5 years agoRename potri to cholesky_inverse (#19498)
vishwakftw [Mon, 22 Apr 2019 15:14:49 +0000 (08:14 -0700)]
Rename potri to cholesky_inverse (#19498)

Summary:
Changelog:
- Rename `potri` to `cholesky_inverse` to remain consistent with names of `cholesky` methods (`cholesky`, `cholesky_solve`)
- Fix all callsites
- Rename all tests
- Create a tentative alias for `cholesky_inverse` under the name `potri` and add a deprecation warning to not promote usage
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19498

Differential Revision: D15029901

Pulled By: ezyang

fbshipit-source-id: 2074286dc93d8744cdc9a45d54644fe57df3a57a

5 years agoAdd assertion to make sure init op is always fp16 compatible in fp16 training
Jiyan Yang [Mon, 22 Apr 2019 06:40:24 +0000 (23:40 -0700)]
Add assertion to make sure init op is always fp16 compatible in fp16 training

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/18498

Reviewed By: kennyhorror

Differential Revision: D14626755

fbshipit-source-id: d8a0b3c02920ab3835911a21bf05e8956853fcd7

5 years agoGenerate only one Type class per backend (#19295)
Roy Li [Mon, 22 Apr 2019 04:12:21 +0000 (21:12 -0700)]
Generate only one Type class per backend (#19295)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19295
ghimport-source-id: 9345110f91f044a449804ddd5116cc9179444a00

Differential Revision: D14948581

Pulled By: li-roy

fbshipit-source-id: a317b03d58d621e8df162918038f7543bfb13ba2

5 years agoMake complex its own backend (#19275)
Roy Li [Mon, 22 Apr 2019 04:12:21 +0000 (21:12 -0700)]
Make complex its own backend (#19275)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19275
ghimport-source-id: 73fd40b02152aed6f24225a88d7ffde7f700899e

Differential Revision: D14948582

Pulled By: li-roy

fbshipit-source-id: a1be6e57057defc74a007c5351c5edb2b9dcaf30

5 years agoAdd ScalarType argument to Type::options() (#19270)
Roy Li [Mon, 22 Apr 2019 04:12:21 +0000 (21:12 -0700)]
Add ScalarType argument to Type::options() (#19270)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19270
ghimport-source-id: a5ade6131f3260066c5750ea1fa9ed5c998bb791

Differential Revision: D14938707

Pulled By: li-roy

fbshipit-source-id: 018fb3f01706531a06515d6d861e5683a455a705

5 years agoGenerate cases for all ScalarTypes in Type functions that call to TH (#19230)
Roy Li [Mon, 22 Apr 2019 04:12:21 +0000 (21:12 -0700)]
Generate cases for all ScalarTypes in Type functions that call to TH (#19230)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19230
ghimport-source-id: 81f360f2ebd137b8e7d8e885b85246cc219761aa

Differential Revision: D14927991

Pulled By: li-roy

fbshipit-source-id: 1b6a57918ecdc9c87858d3e50578edef0b6e7ad5

5 years agoFix clang-format. (#19550)
Mikhail Zolotukhin [Mon, 22 Apr 2019 03:28:15 +0000 (20:28 -0700)]
Fix clang-format. (#19550)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19550
ghimport-source-id: 980d96762426d3e97c26839edbaf107a3fc18b2f

Differential Revision: D15028055

Pulled By: ZolotukhinM

fbshipit-source-id: a50a0aaa74d0f1b9249ad79ab80e4b7747c3bffc

5 years agoFix some typos in jit README
Shen Li [Mon, 22 Apr 2019 02:39:54 +0000 (19:39 -0700)]
Fix some typos in jit README

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/19548

Differential Revision: D15028275

Pulled By: mrshenli

fbshipit-source-id: 84ff635be3b4681962451b4c301271683174d7a8

5 years agoMatch JIT signature with triu_indices / tril_indices. (#19484)
Gregory Chanan [Sun, 21 Apr 2019 22:52:37 +0000 (15:52 -0700)]
Match JIT signature with triu_indices / tril_indices. (#19484)

Summary:
This just plugs into the existing mechanism to do a direct translation to TensorOptions in the backend, so no codegen changes.

After this lands, all native_functions will match the JIT signature.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19484

Differential Revision: D15013051

Pulled By: gchanan

fbshipit-source-id: 6818f868d2f765ca3e56e7e6f75fe4f68492466c

5 years agoMake one_hot non-differentiable. (#19524)
Gregory Chanan [Sun, 21 Apr 2019 21:11:14 +0000 (14:11 -0700)]
Make one_hot non-differentiable. (#19524)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19524
ghimport-source-id: ceda3ad43471242ebbd272a21de11731c7d8bef6

Differential Revision: D15021417

Pulled By: gchanan

fbshipit-source-id: 65d1f17a32f81f47dba5e58e343d0b7b828e1d51

5 years agoRemove 'BoolTensor', 'IndexTensor' from frontend specifications. (#19523)
Gregory Chanan [Sun, 21 Apr 2019 21:11:14 +0000 (14:11 -0700)]
Remove 'BoolTensor', 'IndexTensor' from frontend specifications. (#19523)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19523
ghimport-source-id: 618a15c2d1d9af9f87b46e32f10ff77111c2e3b7

Differential Revision: D15021420

Pulled By: gchanan

fbshipit-source-id: 048af8da3128de10bdee5827b6fbc169c3ad25a8

5 years agoHave _embedding_bag_dense_backward match JIT signature. (#19522)
Gregory Chanan [Sun, 21 Apr 2019 21:11:14 +0000 (14:11 -0700)]
Have _embedding_bag_dense_backward match JIT signature. (#19522)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19522
ghimport-source-id: ad645d87396de645a1aff5fd9d9939cb79cf6558

Differential Revision: D15021419

Pulled By: gchanan

fbshipit-source-id: bd7017edadb4ec9d43cefddf0aee8c52c5cca6a4

5 years agoHave embedding_dense_backward match JIT signature. (#19521)
Gregory Chanan [Sun, 21 Apr 2019 21:11:14 +0000 (14:11 -0700)]
Have embedding_dense_backward match JIT signature. (#19521)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19521
ghimport-source-id: 817d3defb5f4ee98bae1f0488f99cb0e9a5226a2

Differential Revision: D15021376

Pulled By: gchanan

fbshipit-source-id: 2e29f1d3913f94fab3347dc48676303510d7da46

5 years agoUpdate mkldnn-bridge to fix crash issue in DNNLOWP dequantize op (#19159)
Gu, Jinghui [Sun, 21 Apr 2019 20:59:26 +0000 (13:59 -0700)]
Update mkldnn-bridge to fix crash issue in DNNLOWP dequantize op (#19159)

Summary:
Remove an useless format checker in mkldnn-bridge to fix the crash issue in DNNLOWP dequantize op.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19159

Differential Revision: D15027670

Pulled By: yinghai

fbshipit-source-id: ac97d6ff94de013105108b9596b1bd7621c5aa75

5 years agoHook up non_differentiability in derivatives.yaml when no autograd function is genera...
Gregory Chanan [Sun, 21 Apr 2019 20:43:02 +0000 (13:43 -0700)]
Hook up non_differentiability in derivatives.yaml when no autograd function is generated. (#19520)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19520
ghimport-source-id: a1272aa0b23692fb189974c4daba7b2e4e0dad50

Differential Revision: D15021380

Pulled By: gchanan

fbshipit-source-id: ec83efd4bb6d17714c060f13a0527a33a10452db

5 years agoMove non_differentiable_arg_names from autograd functions to differentiability_info...
Gregory Chanan [Sun, 21 Apr 2019 18:03:09 +0000 (11:03 -0700)]
Move non_differentiable_arg_names from autograd functions to differentiability_info. (#19519)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19519
ghimport-source-id: 74e603688b2e4ed33f6c46c7da9d009336140e74

Differential Revision: D15021378

Pulled By: gchanan

fbshipit-source-id: e366a914c67a90ba0552b67d0bf5b347edbaf189

5 years agoMove cuFFT plan cache note outside Best Practices (#19538)
Tongzhou Wang [Sun, 21 Apr 2019 04:36:54 +0000 (21:36 -0700)]
Move cuFFT plan cache note outside Best Practices (#19538)

Summary:
I mistakenly put it there.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19538

Differential Revision: D15026500

Pulled By: soumith

fbshipit-source-id: 0c13499571fdfd789c3bd1c4b58abd870725d422

5 years agoRevert D14689639: [pytorch] Allow passing lists as trace inputs.
Michael Suo [Sat, 20 Apr 2019 15:45:22 +0000 (08:45 -0700)]
Revert D14689639: [pytorch] Allow passing lists as trace inputs.

Differential Revision:
D14689639

Original commit changeset: 6dcec8a64319

fbshipit-source-id: 03a5e7c80e7f2420e33b056b5844a78d7fd41141

5 years agoImprove optimizations for DNNLOWP support on MKL-DNN (#18843)
Gu, Jinghui [Sat, 20 Apr 2019 09:09:15 +0000 (02:09 -0700)]
Improve optimizations for DNNLOWP support on MKL-DNN (#18843)

Summary:
In this PR, the fusion alogrithms are improved to support DNNLOWP.
1. Enabled conv fusions for DNNLOWP
2. Fused order switch op into following quantize op
3. Improve conv+sum fusion to parse larger scope/window
4. re-org fusion code to fix random crash issue due to changing graph
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18843

Differential Revision: D15021030

Pulled By: yinghai

fbshipit-source-id: 88d2199d9fc69f392de9bfbe1f291e0ebf78ab08

5 years agoMake Observer class as template Quant class for QuantConfig (#19418)
Nishant Pandit [Sat, 20 Apr 2019 04:39:00 +0000 (21:39 -0700)]
Make Observer class as template Quant class for QuantConfig (#19418)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19418

This change makes Observer class template which always
takes an observer function as argument. Second test-case becomes redundant, hence removing
it.

Reviewed By: jerryzh168

Differential Revision: D15000594

fbshipit-source-id: 9555fe98a5f2054b8fd01e64e9ac2db72c043bfa

5 years agoSupport compilation on gcc-7.4.0 (#19470)
Sam Leeman-Munk [Sat, 20 Apr 2019 04:35:35 +0000 (21:35 -0700)]
Support compilation on gcc-7.4.0 (#19470)

Summary:
There are two corrections in this pull request.
The first is specific to gcc-7.4.0.
compiled with -std=c++14 gcc-7.4.0 has __cplusplus = 201402L
This does not meet the check set in Deprecated.h, which asks for >201402L.
The compiler goes down to the __GNUC__ check, which passes and sets C10_DEPRECATED_MESSAGE to a value that c++14 does not appear to support or even recognize, leading to a compile time error.
My recommended solution, which worked for my case, was to change the = into a >=

The second correction comes in response to this error:
caffe2/operators/crash_op.cc: In member function ‘virtual bool caffe2::CrashOp::RunOnDevice()’:
caffe2/operators/crash_op.cc:14:11: error: ‘SIGABRT’ was not declared in this scope

I am merely committing to the repository the solution suggested here (which worked for me)
https://discuss.pytorch.org/t/building-pytorch-from-source-in-conda-fails-in-pytorch-caffe2-operators-crash-op-cc/42859
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19470

Differential Revision: D15019529

Pulled By: ailzhang

fbshipit-source-id: 9ce9d713c860ee5fd4266e5c2a7f336a97d7a90d

5 years agoImprove embedding_bag add kernel (#19329)
James Reed [Sat, 20 Apr 2019 02:13:10 +0000 (19:13 -0700)]
Improve embedding_bag add kernel (#19329)

Summary:
This was actually getting pretty poor throughput with respect to memory bandwidth. I used this test to measure the memory bandwidth specifically for the AXPY call: https://gist.github.com/jamesr66a/b27ff9ecbe036eed5ec310c0a3cc53c5

And I got ~8 GB/s before this change, but ~14 GB/s after this change.

This seems to speed up the operator overall by around 1.3x (benchmark: https://gist.github.com/jamesr66a/c533817c334d0be432720ef5e54a4166):

== Before ==

time_per_iter 0.0001298875093460083
GB/s 3.082544287868467

== After ==

time_per_iter 0.00010104801654815674
GB/s 3.9623142905451076

The large difference between the local BW increase and the full-op BW increase likely indicates significant time is being spent elsewhere in the op, so I will investigate that.

EDIT: I updated this PR to include a call into caffe2/perfkernels. This is the progression:

before

time_per_iter 8.983819484710693e-05
GB/s 4.456723564864611

After no axpy
time_per_iter 7.19951868057251e-05
GB/s 5.56126065872172

AFter perfkernels
time_per_iter 5.6699180603027346e-05
GB/s 7.061548257694262

After perfkernels no grad
time_per_iter 4.388842582702637e-05
GB/s 9.122769670026413
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19329

Reviewed By: dzhulgakov

Differential Revision: D14969630

Pulled By: jamesr66a

fbshipit-source-id: 42d1015772c87bedd119e33c0aa2c8105160a738

5 years agoMake finding unused model parameters optional (#19515)
Pieter Noordhuis [Sat, 20 Apr 2019 00:20:37 +0000 (17:20 -0700)]
Make finding unused model parameters optional (#19515)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19515

This is still done by default, but can now be disabled by specifying
`find_unused_parameters=False`. There are use cases where finding
unused parameters results in erroneous behavior, because a subset of
model parameters is used *outside* the `forward` function. One can
argue that doing this is not a good idea, but we should not break
existing use cases without an escape hatch. This configuration
parameter is that escape hatch.

Reviewed By: bddppq

Differential Revision: D15016381

fbshipit-source-id: f2f86b60771b3801ab52776e62b5fd6748ddeed0

5 years agoDisallow std::vector arguments (#19511)
Sebastian Messmer [Fri, 19 Apr 2019 23:59:50 +0000 (16:59 -0700)]
Disallow std::vector arguments (#19511)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19511

In the c10 operator registration API, disallow std::vector arguments and show a nice error message
pointing users towards using ArrayRef instead.

Reviewed By: ezyang

Differential Revision: D15017423

fbshipit-source-id: 157ecc1298bbc598d2e310a16041edf195aaeff5

5 years agoDrop instead of pop (#19503)
Sebastian Messmer [Fri, 19 Apr 2019 23:59:50 +0000 (16:59 -0700)]
Drop instead of pop (#19503)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19503

After reading the arguments from the stack, the c10 kernel wrapper accidentally popped them again, causing a vector to be allocated.
Instead, it should just drop them because they have already been read.

Reviewed By: ezyang

Differential Revision: D15016023

fbshipit-source-id: b694a2929f97fa77cebe247ec2e49820a3c818d5

5 years agoAdd minimalistic implementation of subgraph matcher. (#19322)
Mikhail Zolotukhin [Fri, 19 Apr 2019 23:29:02 +0000 (16:29 -0700)]
Add minimalistic implementation of subgraph matcher. (#19322)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19322
ghimport-source-id: 93c713f829d1b2a9aa5d104cb1f30148dd37c967

Differential Revision: D14962182

Pulled By: ZolotukhinM

fbshipit-source-id: 3989fba06502011bed9c24f12648d0baa2a4480c

5 years agoFix op benchmarks error in OSS environment (#19518)
Mingzhe Li [Fri, 19 Apr 2019 23:22:13 +0000 (16:22 -0700)]
Fix op benchmarks error in OSS environment (#19518)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19518

Previous design needs to run the op benchmarks from PyTorch root directory which could lead to `module not found` error in OSS environment. This diff fixes that issue by making the benchmark to be launched in the `benchmarks` folder.

Reviewed By: ilia-cher

Differential Revision: D15020787

fbshipit-source-id: eb09814a33432a66cc857702bc86538cd17bea3b

5 years agofix AI-PEP path error (#19514)
Mingzhe Li [Fri, 19 Apr 2019 23:22:12 +0000 (16:22 -0700)]
fix AI-PEP path error (#19514)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19514

as title

Reviewed By: hl475

Differential Revision: D15018499

fbshipit-source-id: 9ce38e3a577432e0575a6743f5dcd2e907d3ab9d

5 years agoFirst step at container aliasing (#18710)
eellison [Fri, 19 Apr 2019 23:04:01 +0000 (16:04 -0700)]
First step at container aliasing (#18710)

Summary:
First step at allowing container types within alias analysis.

Since the current implementation hides the concept of Wildcards within alias analysis and does not expose it to memory dag, we cannot represent whether a container type holds a wildcard. As a result, only handle TupleConstruct, where we can directly inspect if any input values are wildcards, and don't handle nested containers.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18710

Differential Revision: D15017068

Pulled By: eellison

fbshipit-source-id: 3ee76a5482cef1cc4a10f034593ca21019161c18

5 years agoFix relu bug for empty tensor (#19451)
Xiaomeng Yang [Fri, 19 Apr 2019 22:14:50 +0000 (15:14 -0700)]
Fix relu bug for empty tensor (#19451)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19451

Fix relu bug for empty tensor

Reviewed By: xianjiec

Differential Revision: D15009811

fbshipit-source-id: b75e567c3bec08d7d12b950d8f1380c50c138704

5 years agoAllow passing lists as trace inputs.
Eric Faust [Fri, 19 Apr 2019 20:28:42 +0000 (13:28 -0700)]
Allow passing lists as trace inputs.

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/18636

Differential Revision: D14689639

fbshipit-source-id: 6dcec8a64319ae3c4da9a93f574a13ce8ec223a5

5 years agoAllow for segmented printing in PythonPrint (#19238)
Michael Suo [Fri, 19 Apr 2019 19:48:39 +0000 (12:48 -0700)]
Allow for segmented printing in PythonPrint (#19238)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19238
ghimport-source-id: 469d33cd187fa68840b201d625800a0f4fead547

Differential Revision: D14928291

Reviewed By: zdevito

Pulled By: suo

fbshipit-source-id: 257fce3dd1601ba192092d3fc318374e3752907e

5 years agoadd resolveType to Resolver (#19237)
Michael Suo [Fri, 19 Apr 2019 19:48:39 +0000 (12:48 -0700)]
add resolveType to Resolver (#19237)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19237
ghimport-source-id: 70777ec37155be37efef1b743d564752e4dff9de

Differential Revision: D14928289

Reviewed By: zdevito

Pulled By: suo

fbshipit-source-id: 46827da9ace16730669fc654bf781d83172d18b1

5 years agoTurn resolver into a class (#19236)
Michael Suo [Fri, 19 Apr 2019 19:48:39 +0000 (12:48 -0700)]
Turn resolver into a class (#19236)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19236
ghimport-source-id: d36705ea5ecff085d0d84ea57bb96d18d7c260dd

Differential Revision: D14928292

Reviewed By: zdevito

Pulled By: suo

fbshipit-source-id: cd038100ac423fa1c19d0547b9e5487a633a2258

5 years agoFix bad annotation in docs (#19501)
davidriazati [Fri, 19 Apr 2019 19:38:23 +0000 (12:38 -0700)]
Fix bad annotation in docs (#19501)

Summary:
Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#19501 [jit] Fix bad annotation in docs**

Pull Request resolved: https://github.com/pytorch/pytorch/pull/19501

Pulled By: driazati

Differential Revision: D15016062

fbshipit-source-id: 3dcd0481eb48b84e98ffe8c5df2cbc9c2abf99f9

5 years agoFix out-of-topological-order issue in Nomnigraph (#19458)
Yinghai Lu [Fri, 19 Apr 2019 19:15:59 +0000 (12:15 -0700)]
Fix out-of-topological-order issue in Nomnigraph (#19458)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19458

The algorithm in https://fburl.com/ggh9iyvc fails to really ensure topological ordering of nodes. The fix is ugly but effective. I think we need a real topological sort to fix this issue more nicely. Mikhail Zolotukhin, Bram Wasti.

Differential Revision: D15011893

fbshipit-source-id: 130c3aa442f5d578adfb14fbe5f16aa722434942

5 years agoRemove uses of TypeID (#19452)
Roy Li [Fri, 19 Apr 2019 18:55:46 +0000 (11:55 -0700)]
Remove uses of TypeID (#19452)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19452
ghimport-source-id: 816ae7fe1a18d76f064d5796dec44dca6a138a21

Differential Revision: D15009920

Pulled By: li-roy

fbshipit-source-id: 722f05a927528148555561da62839f84dba645c6

5 years agoExpose QScheme in frontend (#19381)
Jerry Zhang [Fri, 19 Apr 2019 18:53:46 +0000 (11:53 -0700)]
Expose QScheme in frontend (#19381)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19381

Expose QScheme enum in frontend so that people can use it in
quantization configs in modules.

Differential Revision: D14922992

fbshipit-source-id: ab07b8a7ec42c1c1f5fe84a4a0c805adbcad408d

5 years agoRevert D15003385: Have embedding_dense_backward match JIT signature.
Gregory Chanan [Fri, 19 Apr 2019 18:23:53 +0000 (11:23 -0700)]
Revert D15003385: Have embedding_dense_backward match JIT signature.

Differential Revision:
D15003385

Original commit changeset: 53cbe18aa454

fbshipit-source-id: be904ee2212aa9e402715c436a84d95f6cde326f

5 years agoRevert D15003379: Have _embedding_bag_dense_backward match JIT signature.
Gregory Chanan [Fri, 19 Apr 2019 18:23:53 +0000 (11:23 -0700)]
Revert D15003379: Have _embedding_bag_dense_backward match JIT signature.

Differential Revision:
D15003379

Original commit changeset: f8e82800171f

fbshipit-source-id: 55f83557998d166aeb41d00d7a590acdc76fcf22

5 years agoRevert D15003387: Remove 'BoolTensor', 'IndexTensor' from frontend specifications.
Gregory Chanan [Fri, 19 Apr 2019 18:23:52 +0000 (11:23 -0700)]
Revert D15003387: Remove 'BoolTensor', 'IndexTensor' from frontend specifications.

Differential Revision:
D15003387

Original commit changeset: e518e8ce3228

fbshipit-source-id: af5b107239446ea8d6f229a427d5b157fcafd224

5 years agoRevert D15003382: Make one_hot non-differentiable.
Gregory Chanan [Fri, 19 Apr 2019 18:23:52 +0000 (11:23 -0700)]
Revert D15003382: Make one_hot non-differentiable.

Differential Revision:
D15003382

Original commit changeset: e9244c7a5f0a

fbshipit-source-id: 84789cf4c46c77cce655e70c2a8ff425f32f48bd

5 years agoMake empty_affine_quantized private (#19446)
Jerry Zhang [Fri, 19 Apr 2019 18:06:02 +0000 (11:06 -0700)]
Make empty_affine_quantized private (#19446)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19446

change empty_affine_quantized to _empty_affine_quantized

Reviewed By: dzhulgakov

Differential Revision: D15008757

fbshipit-source-id: c7699ac0c208a8f17d88e95193970c75ba7219d3

5 years agoMake one_hot non-differentiable. (#19430)
Gregory Chanan [Fri, 19 Apr 2019 17:57:46 +0000 (10:57 -0700)]
Make one_hot non-differentiable. (#19430)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19430
ghimport-source-id: 6787473873fdc21400138a4322e17fee8db62607

Differential Revision: D15003382

Pulled By: gchanan

fbshipit-source-id: e9244c7a5f0ad7cd2f79635944a8b37f910231c9

5 years agoRemove 'BoolTensor', 'IndexTensor' from frontend specifications. (#19429)
Gregory Chanan [Fri, 19 Apr 2019 17:57:03 +0000 (10:57 -0700)]
Remove 'BoolTensor', 'IndexTensor' from frontend specifications. (#19429)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19429
ghimport-source-id: 6116682b84210a34babb8b87a92e7050433e5d59

Differential Revision: D15003387

Pulled By: gchanan

fbshipit-source-id: e518e8ce322810e06175bb4e6672d4ea1eb18efd

5 years agoHave embedding_dense_backward match JIT signature. (#19427)
Gregory Chanan [Fri, 19 Apr 2019 17:56:00 +0000 (10:56 -0700)]
Have embedding_dense_backward match JIT signature. (#19427)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19427
ghimport-source-id: 93438cd495129a1e41118c62e6339909783035fd

Differential Revision: D15003385

Pulled By: gchanan

fbshipit-source-id: 53cbe18aa4541a2501f496abfee526e40093c0ff

5 years agoHave _embedding_bag_dense_backward match JIT signature. (#19428)
Gregory Chanan [Fri, 19 Apr 2019 17:53:33 +0000 (10:53 -0700)]
Have _embedding_bag_dense_backward match JIT signature. (#19428)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19428
ghimport-source-id: 037efa3df95efc1fbff631826351d1698a3c49ec

Differential Revision: D15003379

Pulled By: gchanan

fbshipit-source-id: f8e82800171f632e28535e416283d858156068ec

5 years agoStop generating autograd functions for derivatives.yaml entries that only specify...
Gregory Chanan [Fri, 19 Apr 2019 17:53:13 +0000 (10:53 -0700)]
Stop generating autograd functions for derivatives.yaml entries that only specify output differentiability. (#19424)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19424
ghimport-source-id: e9d1b86742607f5cbe39fb278fa7f378739cd6ef

Differential Revision: D15003380

Pulled By: gchanan

fbshipit-source-id: 8efb94fbc0b843863021bf25deab57c492086237

5 years agoFix ord() when dealing with utf8 chars (#19423)
David Riazati [Fri, 19 Apr 2019 17:20:43 +0000 (10:20 -0700)]
Fix ord() when dealing with utf8 chars (#19423)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19423
ghimport-source-id: e7449489fbc86ec1116f94027b3c1561942413ee

Reviewed By: eellison

Differential Revision: D15002847

Pulled By: driazati

fbshipit-source-id: 4560cebcfca695447423d48d65ed364e7dbdbedb

5 years agoFix copied optimizer (#19308)
barrh [Fri, 19 Apr 2019 17:12:46 +0000 (10:12 -0700)]
Fix copied optimizer (#19308)

Summary:
Add the defaults field to the copied object.
Prior to this patch, optimizer.__getattr__ has excluded the defaults
attribute of optimizer source object, required by some LR schedulers. (e.g. CyclicLR with momentum)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19308

Differential Revision: D15012801

Pulled By: soumith

fbshipit-source-id: 95801b269f6f9d78d531d4fed95c973b280cc96f

5 years agoAdd an identity module (#19249)
MilesCranmer [Fri, 19 Apr 2019 17:08:50 +0000 (10:08 -0700)]
Add an identity module (#19249)

Summary:
This is a simple yet useful addition to the torch.nn modules: an identity module. This is a first draft - please let me know what you think and I will edit my PR.

 There is no identity module - nn.Sequential() can be used, however it is argument sensitive so can't be used interchangably with any other module. This adds nn.Identity(...) which can be swapped with any module because it has dummy arguments. It's also more understandable than seeing an empty Sequential inside a model.

See discussion on #9160. The current solution is to use nn.Sequential(). However this won't work as follows:

```python
batch_norm = nn.BatchNorm2d
if dont_use_batch_norm:
    batch_norm = Identity
```

Then in your network, you have:

```python
nn.Sequential(
    ...
    batch_norm(N, momentum=0.05),
    ...
)
```

If you try to simply set `Identity = nn.Sequential`, this will fail since `nn.Sequential` expects modules as arguments. Of course there are many ways to get around this, including:

- Conditionally adding modules to an existing Sequential module
- Not using Sequential but writing the usual `forward` function with an if statement
- ...

**However, I think that an identity module is the most pythonic strategy,** assuming you want to use nn.Sequential.

Using the very simple class (this isn't the same as the one in my commit):

```python
class Identity(nn.Module):
    def __init__(self, *args, **kwargs):
        super().__init__()
    def forward(self, x):
        return x
```

we can get around using nn.Sequential, and `batch_norm(N, momentum=0.05)` will work. There are of course other situations this would be useful.

Thank you.
Best,
Miles
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19249

Differential Revision: D15012969

Pulled By: ezyang

fbshipit-source-id: 9f47e252137a1679e306fd4c169dca832eb82c0c

5 years agoRemove no-fork workaround for running tests with ROCm (#19436)
Junjie Bai [Fri, 19 Apr 2019 16:47:52 +0000 (09:47 -0700)]
Remove no-fork workaround for running tests with ROCm (#19436)

Summary:
This should have been fixed in newest ROCm version.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19436

Reviewed By: ezyang

Differential Revision: D15004685

Pulled By: bddppq

fbshipit-source-id: 19fd4cca94c914dc54aabfbb4e62b328aa348a35

5 years agoDelete defunct test/ffi directory. (#19168)
Edward Yang [Fri, 19 Apr 2019 15:11:01 +0000 (08:11 -0700)]
Delete defunct test/ffi directory. (#19168)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19168
ghimport-source-id: 5190a8d00c529735e99e8745c5e7cf1901fdb800

Differential Revision: D14938318

Pulled By: ezyang

fbshipit-source-id: eaeb6814178c434f737b99ae1fce63fd9ecdb432

5 years agoFix missing doc out= for torch.cumprod (#19340)
Bharat123rox [Fri, 19 Apr 2019 14:56:58 +0000 (07:56 -0700)]
Fix missing doc out= for torch.cumprod (#19340)

Summary:
Fix #19255 by adding the `out=None` argument for `torch.cumprod` missing [here](https://pytorch.org/docs/master/torch.html#torch.cumprod) also added the docstring for `out` in torch.cumsum which was missing [here](https://pytorch.org/docs/master/torch.html#torch.cumsum)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19340

Differential Revision: D14973931

Pulled By: ezyang

fbshipit-source-id: 232f5c9a606b749d67d068afad151539866fedda

5 years agoMention packed accessors in tensor basics doc (#19464)
Clément Pinard [Fri, 19 Apr 2019 14:17:09 +0000 (07:17 -0700)]
Mention packed accessors in tensor basics doc (#19464)

Summary:
This is a continuation of efforts into packed accessor awareness.
A very simple example is added, along with the mention that the template can hold more arguments.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19464

Differential Revision: D15012564

Pulled By: soumith

fbshipit-source-id: a19ed536e016fae519b062d847cc58aef01b1b92

5 years agoRename 'not_differentiable' to 'non_differentiable'. (#19272)
Gregory Chanan [Fri, 19 Apr 2019 13:58:41 +0000 (06:58 -0700)]
Rename 'not_differentiable' to 'non_differentiable'. (#19272)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19272
ghimport-source-id: 755e91efa68c5a1c4377a6853f21b3eee3f8cab5

Differential Revision: D15003381

Pulled By: gchanan

fbshipit-source-id: 54db27c5c5e65acf65821543db3217de9dd9bdb5

5 years agoClean the onnx constant fold code a bit (#19398)
Lu Fang [Fri, 19 Apr 2019 06:56:32 +0000 (23:56 -0700)]
Clean the onnx constant fold code a bit (#19398)

Summary:
This is a follow up PR of https://github.com/pytorch/pytorch/pull/18698 to lint the code using clang-format.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19398

Differential Revision: D14994517

Pulled By: houseroad

fbshipit-source-id: 2ae9f93e66ce66892a1edc9543ea03932cd82bee

5 years agoAllow passing dicts as trace inputs. (#18092)
Eric Faust [Fri, 19 Apr 2019 06:48:59 +0000 (23:48 -0700)]
Allow passing dicts as trace inputs. (#18092)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18092

Previously, tracing required all inputs to be either tensors,
or tuples of tensor. Now, we allow users to pass dicts as well.

Differential Revision: D14491795

fbshipit-source-id: 7a2df218e5d00f898d01fa5b9669f9d674280be3

5 years agoskip test_trace_c10_ops if _caffe2 is not built (#19099)
Lu Fang [Fri, 19 Apr 2019 06:31:32 +0000 (23:31 -0700)]
skip test_trace_c10_ops if _caffe2 is not built (#19099)

Summary:
fix https://github.com/pytorch/pytorch/issues/18142
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19099

Differential Revision: D15010452

Pulled By: houseroad

fbshipit-source-id: 5bf158d7fce7bfde109d364a3a9c85b83761fffb

5 years agoremove needless ## in REGISTER_ALLOCATOR definition. (#19261)
Gemfield [Fri, 19 Apr 2019 05:34:52 +0000 (22:34 -0700)]
remove needless ## in REGISTER_ALLOCATOR definition. (#19261)

Summary:
remove needless ## in REGISTER_ALLOCATOR definition.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19261

Differential Revision: D15002025

Pulled By: soumith

fbshipit-source-id: 40614b1d79d1fe05ccf43f0ae5aab950e4c875c2

5 years agoStrip doc_string from exported ONNX models (#18882)
Lara Haidar-Ahmad [Fri, 19 Apr 2019 05:25:04 +0000 (22:25 -0700)]
Strip doc_string from exported ONNX models (#18882)

Summary:
Strip the doc_string by default from the exported ONNX models (this string has the stack trace and information about the local repos and folders, which can be confidential).

The users can still generate the doc_string by specifying add_doc_string=True in torch.onnx.export().
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18882

Differential Revision: D14889684

Pulled By: houseroad

fbshipit-source-id: 26d2c23c8dc3f484544aa854b507ada429adb9b8

5 years agoimprove dim sort performance (#19379)
Natalia Gimelshein [Fri, 19 Apr 2019 05:20:44 +0000 (22:20 -0700)]
improve dim sort performance (#19379)

Summary:
We are already using custom comparators for sorting (for a good reason), but are still making 2 sorting passes - global sort and stable sorting to bring values into their slices. Using a custom comparator to sort within a slice allows us to avoid second sorting pass and brings up to 50% perf improvement.
t-vi I know you are moving sort to ATen, and changing THC is discouraged, but #18350 seems dormant. I'm fine with #18350 landing first, and then I can put in these changes.
cc umanwizard for review.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19379

Differential Revision: D15011019

Pulled By: soumith

fbshipit-source-id: 48e5f5aef51789b166bb72c75b393707a9aed57c

5 years agoFix missing import sys in pin_memory.py (#19419)
SsnL [Fri, 19 Apr 2019 05:16:05 +0000 (22:16 -0700)]
Fix missing import sys in pin_memory.py (#19419)

Summary:
kostmo pointed this out in #15331. Thanks :)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19419

Differential Revision: D15002846

Pulled By: soumith

fbshipit-source-id: c600fab3f7a7a5147994b9363910af4565c7ee65

5 years agoupdate documentation of PairwiseDistance#19241 (#19412)
Ran [Fri, 19 Apr 2019 05:10:34 +0000 (22:10 -0700)]
update documentation of PairwiseDistance#19241 (#19412)

Summary:
Fix the documentation of PairwiseDistance [#19241](https://github.com/pytorch/pytorch/issues/19241)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19412

Differential Revision: D14998271

Pulled By: soumith

fbshipit-source-id: bcb2aa46d3b3102c4480f2d24072a5e14b049888

5 years agofixes link in TripletMarginLoss (#19417)
Soumith Chintala [Fri, 19 Apr 2019 05:10:30 +0000 (22:10 -0700)]
fixes link in TripletMarginLoss (#19417)

Summary:
Fixes https://github.com/pytorch/pytorch/issues/19245
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19417

Differential Revision: D15001610

Pulled By: soumith

fbshipit-source-id: 1b85ebe196eb5a3af5eb83d914dafa83b9b35b31

5 years agomake separate operators as independent binaries (#19450)
Mingzhe Li [Fri, 19 Apr 2019 02:56:50 +0000 (19:56 -0700)]
make separate operators as independent binaries (#19450)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19450

We want to make each operator benchmark as a separate binary. The previous way to run the benchmark is by collecting all operators into a single binary, it is unnecessary when we want to filter a specific operator. This diff aims to resolve that issue.

Reviewed By: ilia-cher

Differential Revision: D14808159

fbshipit-source-id: 43cd25b219c6e358d0cd2a61463b34596bf3bfac

5 years agoUpdating submodules
svcscm [Fri, 19 Apr 2019 01:26:20 +0000 (18:26 -0700)]
Updating submodules

Reviewed By: cdelahousse

fbshipit-source-id: a727513842c0a240b377bda4e313fbedbc54c2e8

5 years agoStep 4: add support for unique with dim=None (#18651)
Xiang Gao [Fri, 19 Apr 2019 01:24:50 +0000 (18:24 -0700)]
Step 4: add support for unique with dim=None (#18651)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18651
ghimport-source-id: e11988130a3f9a73529de0b0d08b4ec25fbc639c

Differential Revision: D15000463

Pulled By: VitalyFedyunin

fbshipit-source-id: 9e258e473dea6a3fc2307da2119b887ba3f7934a

5 years agoallow bools to be used as attributes (#19440)
Michael Suo [Fri, 19 Apr 2019 01:06:21 +0000 (18:06 -0700)]
allow bools to be used as attributes (#19440)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19440
ghimport-source-id: 9c962054d760526bf7da324b114455fcb1038521

Differential Revision: D15005723

Pulled By: suo

fbshipit-source-id: 75fc87ae33894fc34d3b913881defb7e6b8d7af0

5 years agoFix test build (#19444)
David Riazati [Fri, 19 Apr 2019 01:02:14 +0000 (18:02 -0700)]
Fix test build (#19444)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19444
ghimport-source-id: c85db00e8037e7f6f0424eb8bd17f957d20b7247

Reviewed By: eellison

Differential Revision: D15008679

Pulled By: driazati

fbshipit-source-id: 0987035116d9d0069794d96395c8ad458ba7c121

5 years agopow scalar exponent / base autodiff, fusion (#19324)
Thomas Viehmann [Fri, 19 Apr 2019 00:52:33 +0000 (17:52 -0700)]
pow scalar exponent / base autodiff, fusion (#19324)

Summary:
Fixes: #19253

Fixing pow(Tensor, float) is straightforward.
The breakage for pow(float, Tensor) is a bit more subtle to trigger, and fixing needs `torch.log` (`math.log` didn't work) from the newly merged #19115  (Thanks ngimel for pointing out this has landed.)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19324

Differential Revision: D15003531

Pulled By: ailzhang

fbshipit-source-id: 8b22138fa27a43806b82886fb3a7b557bbb5a865

5 years agoImprove unique CPU performance for returning counts (#19352)
Gao, Xiang [Fri, 19 Apr 2019 00:46:43 +0000 (17:46 -0700)]
Improve unique CPU performance for returning counts (#19352)

Summary:
Benchmark on a tensor of shape `torch.Size([15320, 2])`. Benchmark code:

```python
print(torch.__version__)
print()
a = tensor.flatten()
print('cpu, sorted=False:')
%timeit torch._unique2_temporary_will_remove_soon(a, sorted=False)
%timeit torch._unique2_temporary_will_remove_soon(a, sorted=False, return_inverse=True)
%timeit torch._unique2_temporary_will_remove_soon(a, sorted=False, return_counts=True)
%timeit torch._unique2_temporary_will_remove_soon(a, sorted=False, return_inverse=True, return_counts=True)
print()
print('cpu, sorted=True:')
%timeit torch._unique2_temporary_will_remove_soon(a)
%timeit torch._unique2_temporary_will_remove_soon(a, return_inverse=True)
%timeit torch._unique2_temporary_will_remove_soon(a, return_counts=True)
%timeit torch._unique2_temporary_will_remove_soon(a, return_inverse=True, return_counts=True)
print()
```

Before
```
1.1.0a0+36854fe

cpu, sorted=False:
340 µs ± 4.05 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
724 µs ± 6.28 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
54.3 ms ± 469 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
54.6 ms ± 659 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)

cpu, sorted=True:
341 µs ± 7.21 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
727 µs ± 7.05 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
54.7 ms ± 795 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
54.3 ms ± 647 µs per loop (mean ± std. dev. of 7 runs, 10 loops each)
```

After
```
1.1.0a0+261d9e8

cpu, sorted=False:
350 µs ± 865 ns per loop (mean ± std. dev. of 7 runs, 1000 loops each)
771 µs ± 598 ns per loop (mean ± std. dev. of 7 runs, 1000 loops each)
1.09 ms ± 6.86 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
1.09 ms ± 4.74 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)

cpu, sorted=True:
324 µs ± 4.99 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
705 µs ± 3.18 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
1.09 ms ± 5.22 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
1.09 ms ± 5.63 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19352

Differential Revision: D14984717

Pulled By: VitalyFedyunin

fbshipit-source-id: 3c56f85705ab13a92ec7406f4f30be77226a3210

5 years agoRevert D14909203: Remove usages of TypeID
Pieter Noordhuis [Fri, 19 Apr 2019 00:44:37 +0000 (17:44 -0700)]
Revert D14909203: Remove usages of TypeID

Differential Revision:
D14909203

Original commit changeset: d716179c484a

fbshipit-source-id: 992ff1fcd6d35d3f2ae768c7e164b7a0ba871914

5 years agoAdd tests for argument types (#19290)
Sebastian Messmer [Fri, 19 Apr 2019 00:16:58 +0000 (17:16 -0700)]
Add tests for argument types (#19290)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19290

Add test cases for the supported argument types
And TODOs for some unsupported ones that we might want to support.

Reviewed By: dzhulgakov

Differential Revision: D14931920

fbshipit-source-id: c47bbb295a54ac9dc62569bf5c273368c834392c

5 years agoAllow optionals arguments from C++ (#19311)
David Riazati [Fri, 19 Apr 2019 00:06:09 +0000 (17:06 -0700)]
Allow optionals arguments from C++ (#19311)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19311
ghimport-source-id: 699f62eb2bbad53ff2045fb2e217eb1402f2cdc5

Reviewed By: eellison

Differential Revision: D14983059

Pulled By: driazati

fbshipit-source-id: 442f96d6bd2a8ce67807ccad2594b39aae489ca5

5 years agoEnhance front-end to add op (#19433)
Mingzhe Li [Fri, 19 Apr 2019 00:03:56 +0000 (17:03 -0700)]
Enhance front-end to add op (#19433)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19433

For operator benchmark project, we need to cover a lot of operators, so the interface for adding operators needs to be very clean and simple. This diff is implementing a new interface to add op.

Here is the logic to add new operator to the benchmark:
```
long_config = {}
short_config = {}

map_func

add_test(
  [long_config, short_config],
  map_func,
  [caffe2 op]
  [pt op]
)
```

Reviewed By: zheng-xq

Differential Revision: D14791191

fbshipit-source-id: ac6738507cf1b9d6013dc8e546a2022a9b177f05

5 years agoFix cpp_custom_type_hack variable handling (#19400)
Dmytro Dzhulgakov [Thu, 18 Apr 2019 23:34:20 +0000 (16:34 -0700)]
Fix cpp_custom_type_hack variable handling (#19400)

Summary:
My bad - it might be called in variable and non-variable context. So it's better to just inherit variable-ness from the caller.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19400

Reviewed By: ezyang

Differential Revision: D14994781

Pulled By: dzhulgakov

fbshipit-source-id: cb9d055b44a2e1d7bbf2e937d558e6bc75037f5b

5 years agofix hub doc formatting issues (#19434)
Ailing Zhang [Thu, 18 Apr 2019 22:58:45 +0000 (15:58 -0700)]
fix hub doc formatting issues (#19434)

Summary:
minor fixes for doc
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19434

Differential Revision: D15003903

Pulled By: ailzhang

fbshipit-source-id: 400768d9a5ee24f9183faeec9762b688c48c531b

5 years agoRecursively find tensors in DDP module output (#19360)
Pieter Noordhuis [Thu, 18 Apr 2019 21:51:37 +0000 (14:51 -0700)]
Recursively find tensors in DDP module output (#19360)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19360

We'll return the output object verbatim since it is a freeform object.
We need to find any tensors in this object, though, because we need to
figure out which parameters were used during this forward pass, to
ensure we short circuit reduction for any unused parameters.

Before this commit only lists were handled and the functionality went
untested. This commit adds support for dicts and recursive structures,
and also adds a test case.

Closes #19354.

Reviewed By: mrshenli

Differential Revision: D14978016

fbshipit-source-id: 4bb6999520871fb6a9e4561608afa64d55f4f3a8

5 years agoMoving at::Tensor into caffe2::Tensor without bumping refcount (#19388)
Sebastian Messmer [Thu, 18 Apr 2019 21:07:30 +0000 (14:07 -0700)]
Moving at::Tensor into caffe2::Tensor without bumping refcount (#19388)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19388

The old implementation forced a refcount bump when converting at::Tensor to caffe2::Tensor.
Now, it is possible to move it without a refcount bump.

Reviewed By: dzhulgakov

Differential Revision: D14986815

fbshipit-source-id: 92b4b0a6f323ed38376ffad75f960cad250ecd9b

5 years agoFix pickling torch.float32 (#18045)
Ailing Zhang [Thu, 18 Apr 2019 19:07:17 +0000 (12:07 -0700)]
Fix pickling torch.float32 (#18045)

Summary:
Attempt fix for #14057 . This PR fixes the example script in the issue.
The old behavior is a bit confusing here. What happened to pickling is python2 failed to recognize `torch.float32` is in module `torch`, thus it's looking for `torch.float32` in module `__main__`. Python3 is smart enough to handle it.
According to the doc [here](https://docs.python.org/2/library/pickle.html#object.__reduce__), it seems `__reduce__` should return `float32` instead of the old name `torch.float32`. In this way python2 is able to find `float32` in `torch` module.
> If a string is returned, it names a global variable whose contents are pickled as normal. The string returned by __reduce__() should be the object’s local name relative to its module
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18045

Differential Revision: D14990638

Pulled By: ailzhang

fbshipit-source-id: 816b97d63a934a5dda1a910312ad69f120b0b4de

5 years agoRespect order of Parameters in rnn.py (#18198)
David Riazati [Thu, 18 Apr 2019 18:07:45 +0000 (11:07 -0700)]
Respect order of Parameters in rnn.py (#18198)

Summary:
Previously to get a list of parameters this code was just putting them in the reverse order in which they were defined, which is not always right. This PR allows parameter lists to define the order themselves. To do this parameter lists need to have a corresponding function that provides the names of the parameters.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18198

Differential Revision: D14966270

Pulled By: driazati

fbshipit-source-id: 59331aa59408660069785906304b2088c19534b2

5 years agoRefactor EmitLoopCommon to make it more amenable to future extensions (#19341)
Nikolay Korovaiko [Thu, 18 Apr 2019 16:56:02 +0000 (09:56 -0700)]
Refactor EmitLoopCommon to make it more amenable to future extensions (#19341)

Summary:
This PR paves the way for support more iterator types in for-in loops.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19341

Differential Revision: D14992749

Pulled By: Krovatkin

fbshipit-source-id: e2d4c9465c8ec3fc74fbf23006dcb6783d91795f