platform/upstream/pytorch.git
5 years agoExpose GenerateProposals to PyTorch
Sebastian Messmer [Mon, 11 Feb 2019 22:03:45 +0000 (14:03 -0800)]
Expose GenerateProposals to PyTorch

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/16880

Reviewed By: bwasti

Differential Revision: D13998092

fbshipit-source-id: 23ab886ba137377312557fa718f262f4c8149cc7

5 years agoExpose BBoxTransform to pytorch
Sebastian Messmer [Mon, 11 Feb 2019 22:03:45 +0000 (14:03 -0800)]
Expose BBoxTransform to pytorch

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/16879

Reviewed By: bwasti

Differential Revision: D13998093

fbshipit-source-id: ddfe4bff83e9a1a4cedf1e520e6d2977b21cb3af

5 years agoMinimize templated code in caffe2 operator wrapper (#16965)
Sebastian Messmer [Mon, 11 Feb 2019 22:03:45 +0000 (14:03 -0800)]
Minimize templated code in caffe2 operator wrapper (#16965)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16965

Instead of having one large templated function to wrap the caffe2 op, minimize the amount of templated code.
Non-templated code can be reused between different operators and decreases binary size.

Reviewed By: orionr

Differential Revision: D14018806

fbshipit-source-id: bedd4152eec21dd8c5778446963826316d210543

5 years agoDon't keep unnecessary saved_inputs alive (#16583)
Adam Paszke [Mon, 11 Feb 2019 21:31:06 +0000 (13:31 -0800)]
Don't keep unnecessary saved_inputs alive (#16583)

Summary:
Fixes #16577.

This greatly improves memory efficiency of certain ops like Dropout2d. Previously, they were implemented as `input * mask` where mask never requires_grad, but we didn't use that knowledge in forward, and (in case of a in-place dropout) kept input.clone() for the backward, when it would simply get ignored.

This patch tries to address this situation by emitting some guards for stores like this, but only if they are as simple, as checking if a single value requires_grad.

Interestingly, the same optimizations apply to methods like bmm, baddmm, etc., but _not to mm nor addmm_, because of how their derivatives are defined. Apparently they unnecessarily use `mat1` to compute the derivative of `mat1` just to improve the error message in case `mat1` was sparse. I'd like to apply this optimization to that case, but I don't want to loose the nicer error message, so if anyone has any ideas for solutions, please let me know...

Full list of operators affected by this patch:
* _nnpack_spatial_convolution
* addbmm
* addcdiv
* addcmul
* addmv
* addr
* baddbmm
* bmm
* cross
* div
* dot
* fmod
* ger
* index_add_
* mul
* mv
* scatter_add_
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16583

Differential Revision: D13900881

Pulled By: gchanan

fbshipit-source-id: dd0aeb2ab58c4b6aa95b37b46d3255b3e014291c

5 years agoEnforce same input tensor storage in VariableType functions (#16305)
Will Feng [Mon, 11 Feb 2019 20:48:17 +0000 (12:48 -0800)]
Enforce same input tensor storage in VariableType functions (#16305)

Summary:
In VariableType.cpp, when a function modifies its input tensors, it should only change the input tensors' storage data in-place, and should never change the input tensors' storage pointers. This PR adds checks for this, and also fixes functions that fail this test.

This is part of the Variable/Tensor merge work (https://github.com/pytorch/pytorch/issues/13638).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16305

Differential Revision: D13897855

Pulled By: yf225

fbshipit-source-id: 0c4fc7eb530d30db88037b1f0981f6f8454d3b79

5 years agoRevert unneeded fixes in flat_hash_map (#16907)
Sebastian Messmer [Mon, 11 Feb 2019 20:29:47 +0000 (12:29 -0800)]
Revert unneeded fixes in flat_hash_map (#16907)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16907

The begin()/end() fix actually doesn't make sense, see my comment on https://github.com/skarupke/flat_hash_map/pull/8
This diff removes it.

Reviewed By: ezyang

Differential Revision: D13985779

fbshipit-source-id: f08b02c941069e2a4e728e02a19b65dc72f96b41

5 years agoFix constexpr in KernelRegistrationBuilder (#16906)
Sebastian Messmer [Mon, 11 Feb 2019 20:29:47 +0000 (12:29 -0800)]
Fix constexpr in KernelRegistrationBuilder (#16906)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16906

In C++11, constexpr implies const, so these methods actually wouldn't be rvalue overloads as intended but const rvalue overloads.
Let's only apply the constexpr flag in C++14 to be safe.

Reviewed By: bddppq

Differential Revision: D13998486

fbshipit-source-id: a04d17ef0cc8f45e3d0a1ca9843d194f4f0f6f7f

5 years agoCatch cudaError_t return val (nodiscard in rocm) (#16399)
Xiaodong Wang [Mon, 11 Feb 2019 20:27:12 +0000 (12:27 -0800)]
Catch cudaError_t return val (nodiscard in rocm) (#16399)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16399

Catching cudaError_t return values in a few places, because it's nodiscard in rocm. Unless we add -Wno-unused-result, it'll end up with a compilation error.

Also in c10/cuda/test, check whether a host has GPU or not. We were silently throwing out the error before (so not really testing the cuda api).

Reviewed By: bddppq

Differential Revision: D13828281

fbshipit-source-id: 587d1cc31c20b836ce9594e3c18f067d322b2934

5 years agooptionally zero infinite losses in CTCLoss (#16199)
Thomas Viehmann [Mon, 11 Feb 2019 20:26:47 +0000 (12:26 -0800)]
optionally zero infinite losses in CTCLoss (#16199)

Summary:
Here is a stab at implementing an option to zero out infinite losses (and NaN gradients).
It might be nicer to move the zeroing to the respective kernels.
The default is currently `False` to mimic the old behaviour, but I'd be half inclined to set the default to `True`, because the behaviour wasn't consistent between CuDNN and Native anyways and the NaN gradients aren't terribly useful.

This topic seems to come up regularly, e.g. in  #14335
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16199

Differential Revision: D14020462

Pulled By: ezyang

fbshipit-source-id: 5ba8936c66ec6e61530aaf01175dc49f389ae428

5 years agoMerge binaries "convert_image_to_tensor" and "caffe2_benchmark" (#16875)
Zhizhen Qin [Mon, 11 Feb 2019 20:24:10 +0000 (12:24 -0800)]
Merge binaries "convert_image_to_tensor" and "caffe2_benchmark" (#16875)

Summary:
Merge binaries "convert_image_to_tensor" and "caffe2_benchmark" to remove the overhead of writing to/reading from Tensor file.

*TODO next: TensorProtos is another overhead. No need for de-serialization.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16875

Reviewed By: sf-wind

Differential Revision: D13997726

Pulled By: ZhizhenQin

fbshipit-source-id: 4dec17f0ebb59cf1438b9aba5421db2b41c47a9f

5 years agoFix missing CircleCI GPG key (#16961)
SsnL [Mon, 11 Feb 2019 19:59:17 +0000 (11:59 -0800)]
Fix missing CircleCI GPG key (#16961)

Summary:
I'm seeing a bunch of apt gpg key errors on CI with the following message:
```
An error occurred during the signature verification. The repository is not
updated and the previous index files will be used. GPG error:
https://packagecloud.io trusty InRelease: The following signatures couldn't
be verified because the public key is not available:
NO_PUBKEY 4E6910DFCB68C9CD
```

Most of the times apt will reuse the old cached version, but sometimes this results in a build failure: https://circleci.com/gh/pytorch/pytorch/758366?utm_campaign=vcs-integration-link&utm_medium=referral&utm_source=github-build-link.

This should hopefully fix it.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16961

Differential Revision: D14028151

Pulled By: ezyang

fbshipit-source-id: 7648a0a58ece38d8d04916937a9fa17f34f8833e

5 years agoDisable binary_linux_conda_3.6_cu90_build on PRs. (#16958)
Edward Yang [Mon, 11 Feb 2019 19:43:45 +0000 (11:43 -0800)]
Disable binary_linux_conda_3.6_cu90_build on PRs. (#16958)

Summary:
Issue tracked at https://github.com/pytorch/pytorch/issues/16710

Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16958

Differential Revision: D14028078

Pulled By: ezyang

fbshipit-source-id: 6c68f79775a156ef4a55ac450a5a0ecacc0e6af5

5 years agoInstall Thrust package and stop patching (#16911)
Xiaodong Wang [Mon, 11 Feb 2019 17:44:17 +0000 (09:44 -0800)]
Install Thrust package and stop patching (#16911)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16911

I think the Thrust package has want we want for /opt/rocm/include/thrust. We probably can stop patching it now.

Reviewed By: bddppq

Differential Revision: D14015177

fbshipit-source-id: 8d9128783a790c39083a1b8b4771c2c18bd67d46

5 years agoMake pin_memory and default_collate preserve namedtuples (#16440)
Eskil Jörgensen [Mon, 11 Feb 2019 16:22:15 +0000 (08:22 -0800)]
Make pin_memory and default_collate preserve namedtuples (#16440)

Summary:
Open issue: https://github.com/pytorch/pytorch/issues/3281
Corresponding PR (conflict): https://github.com/pytorch/pytorch/pull/4577

Another open issue: https://github.com/pytorch/pytorch/issues/14613
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16440

Differential Revision: D14020901

Pulled By: ezyang

fbshipit-source-id: 4abe817fc43c281a510715d311bad544511995d3

5 years agoRevert D14020906: [pytorch][PR] Extend support for exporting reshape to onnx.
Edward Yang [Mon, 11 Feb 2019 14:02:08 +0000 (06:02 -0800)]
Revert D14020906: [pytorch][PR] Extend support for exporting reshape to onnx.

Differential Revision:
D14020906

Original commit changeset: 168616873044

fbshipit-source-id: 2730bb6990d41f3a9cef6625ea919c219733433d

5 years agoAdded scientific notation on set_printoptions (#16876)
Ivan Ogasawara [Mon, 11 Feb 2019 12:52:06 +0000 (04:52 -0800)]
Added scientific notation on set_printoptions (#16876)

Summary:
This PR fixes #15683
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16876

Differential Revision: D14021703

Pulled By: soumith

fbshipit-source-id: 1f603a7d24e331831d8d389f4a704c6a5b070b0c

5 years agoExtend support for exporting reshape to onnx.
BowenBao [Mon, 11 Feb 2019 04:12:45 +0000 (20:12 -0800)]
Extend support for exporting reshape to onnx.

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/16632

Differential Revision: D14020906

Pulled By: ezyang

fbshipit-source-id: 168616873044b980145a3554dab942bdec19efb2

5 years agoInt8GivenTensorFill Operator Schema fix typo (#16204)
eyyub.sari@epitech.eu [Mon, 11 Feb 2019 04:05:32 +0000 (20:05 -0800)]
Int8GivenTensorFill Operator Schema fix typo (#16204)

Summary:
Hi,
caffe2/operators/quantized/int8_given_tensor_fill_op.cc expects the value array to be named "values" but the operator schema describe "value" (no s). I guess it is a little typo but it made me losing a bit of time before understanding why I had this error by passing "value" instead of "values":
```
[F int8_given_tensor_fill_op.h:95] Check failed: output->t.numel() == values_.numel() output size: 3 given size: 0
Aborted (core dumped)
```

Thanks,
Eyyüb Sari
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16204

Differential Revision: D14020476

Pulled By: ezyang

fbshipit-source-id: a8a46bfc44ec125e7925ce4b7c79fdf99c890a50

5 years agoAdd support for fusion of half batch norm with float stats (#16735)
Adam Paszke [Mon, 11 Feb 2019 03:32:15 +0000 (19:32 -0800)]
Add support for fusion of half batch norm with float stats (#16735)

Summary:
Fixes #16642.

cc ngimel
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16735

Differential Revision: D14020310

Pulled By: ezyang

fbshipit-source-id: ac78726f471d16d188eb998354d52bc79fe2c282

5 years agoImprove the Sparse matrix multiplication computational speed #16187 (#16905)
musikisomorphie [Mon, 11 Feb 2019 03:31:38 +0000 (19:31 -0800)]
Improve the Sparse matrix multiplication computational speed #16187 (#16905)

Summary:
Instead of converting coo to csr format of the sparse matrix in the original implementation, in my revision I directly use coo format for sparse dense matrix mutliplication.
On my linux machine it is 5 times faster than the original code:

```
(original code)
SIZE: 15000 DENSITY: 0.01 DEVICE: cpu
torch: 0.39403 seconds
np:    0.00496674 seconds
torch/np: 79.3338

----------------------------------------

(my update)
SIZE: 15000 DENSITY: 0.01 DEVICE: cpu
torch: 0.0812583 seconds
np:    0.00501871 seconds
torch/np: 16.1911

```

Further code feedback and running time tests are highly welcomed. I will keep revise my code if needed.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16905

Differential Revision: D14020095

Pulled By: ezyang

fbshipit-source-id: 4ab94075344a55b375f22421e97a690e682baed5

5 years agoAllow dataloader to accept a custom memory pinning function (#16743)
Michael Carilli [Mon, 11 Feb 2019 03:31:23 +0000 (19:31 -0800)]
Allow dataloader to accept a custom memory pinning function (#16743)

Summary:
Renewed attempt at https://github.com/pytorch/pytorch/pull/14171

From the original PR:
> Currently, the pin_memory_batch function in the dataloader will return a batch comprised of any unrecognized type without pinning the data, because it doesn't know how.
>
>This behavior was preventing us from overlapping data prefetching in Mask-RCNN, whose custom collate_fn returns a custom batch type.

The old PR allowed the user to implement batch pinning for custom batch and data types by passing a custom pin function to the dataloader.  slayton58 suggested a cleaner approach:  allow the user to define a `pin_memory` method on their custom types, and have `pin_memory_batch` [check for the presence of that method](https://github.com/pytorch/pytorch/pull/16743/files#diff-9f154cbd884fe654066b1621fad654f3R56) in the incoming batch as a fallback.  I've updated the test and docstrings accordingly.

The old PR was merged but then reverted due to weird cuda OOM errors on windows that may or may not have been related.  I have no idea why my changes would cause such errors (then or now) but it's something to keep an eye out for.

fmassa and yf225 who were my POCs on the old PR.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16743

Differential Revision: D13991745

Pulled By: ezyang

fbshipit-source-id: 74e71f62a03be453b4caa9f5524e9bc53467fa17

5 years agoAdd abs for ByteTensor and CharTensor. (#16893)
Hameer Abbasi [Mon, 11 Feb 2019 03:28:50 +0000 (19:28 -0800)]
Add abs for ByteTensor and CharTensor. (#16893)

Summary:
Fixes #15089
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16893

Differential Revision: D14020115

Pulled By: ezyang

fbshipit-source-id: 6f3be6ed28d2d37667159be45959d400bc473451

5 years agoSupport named tuple return from operators on JIT (#16253)
Xiang Gao [Mon, 11 Feb 2019 02:10:59 +0000 (18:10 -0800)]
Support named tuple return from operators on JIT (#16253)

Summary:
Fixes: https://github.com/pytorch/pytorch/issues/16233

The following changes are made:
- Modify `TupleType` to store optional field names
- Modify schema matching to return fill in those field names when creating  `TupleType` as return type.
- Modify codegen of JIT to copy field names to schema string
- Modify `SchemaParser` to set field names of returned schema.
- Modify `SimpleValue::attr` to emit tuple indexing for named tuple.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16253

Reviewed By: ezyang

Differential Revision: D13954298

Pulled By: zdevito

fbshipit-source-id: 247d483d78a0c9c12d1ba36e1f1ec6c3f1a3007b

5 years agoEnhance the documentation for torch.nn.DataParallel (#15993)
Derek Kim [Sun, 10 Feb 2019 23:52:37 +0000 (15:52 -0800)]
Enhance the documentation for torch.nn.DataParallel (#15993)

Summary:
I found a few sentences in DataParallel docstring confusing, so I suggest this enhancement.

- Arbitrary arguments are allowed to be passed .... *INCLUDING* tensors (Not *EXCLUDING*)
- The original author said that "other types" are shallow-copied but I think actually only some builtin types are (effectively) shallow-copied.  And "other types" are shared. Here is an example.

```python
import torch
from torch.nn import Module, DataParallel
from collections import deque

class MyModel(Module):
    def forward(self, x):
        x.append(None)

model = MyModel(); model.cuda()
model = DataParallel(model)

d = deque()
model.forward(d)
print(d)
```

This is a side note.

As far as I know, copying objects is not a specially frequent operation in python unlike some other languages. Notably, no copying is involved in assignment or function parameter passing. They are only name bindings and it is the whole point of "everything is object" python philosophy, I guess. If one keep this in mind, it may help you dealing with things like multithreading.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15993

Differential Revision: D14020404

Pulled By: ezyang

fbshipit-source-id: a38689c94d0b8f77be70447f34962d3a7cd25e2e

5 years agoDOC: correct docstring for torch and torch.Tensor package (#16842)
ZhuBaohe [Sun, 10 Feb 2019 22:34:31 +0000 (14:34 -0800)]
DOC: correct docstring for torch and torch.Tensor package (#16842)

Summary:
This PR is a simple fix for the mistake in the  "tensor"  and "torch.Tensor"doc.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16842

Differential Revision: D14020300

Pulled By: ezyang

fbshipit-source-id: 3ab04f1223d6e60f8da578d04d759e385d23acbb

5 years agofind libnvToolsExt instead of using only hardcoded path (#16714)
Thomas Viehmann [Sun, 10 Feb 2019 21:57:57 +0000 (13:57 -0800)]
find libnvToolsExt instead of using only hardcoded path (#16714)

Summary:
This changes the libnvToolsExt dependency to go through CMake find_library.

I have a machine where cuda libs, and libnvToolsExt in particular, are in the "usual library locations". It would be neat if we could find libnvToolsExt and use the path currently hardcoded as default.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16714

Differential Revision: D14020315

Pulled By: ezyang

fbshipit-source-id: 00be27be10b1863ca92fd585f273d50bded850f8

5 years agoClean up autograd method tests
Xiang Gao [Sun, 10 Feb 2019 21:42:49 +0000 (13:42 -0800)]
Clean up autograd method tests

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/16790

Differential Revision: D14020305

Pulled By: ezyang

fbshipit-source-id: 3aa3362830cde35967a3895837a25b3cf3287569

5 years agofixed LogSigmoid math string that wasn't rendering in documentation (#16900)
Travis Johnston [Sun, 10 Feb 2019 19:44:54 +0000 (11:44 -0800)]
fixed LogSigmoid math string that wasn't rendering in documentation (#16900)

Summary:
The documentation for LogSigmoid says:

> Applies the element-wise function:
> \<blank\>

Now the documentation properly displays the math string.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16900

Differential Revision: D14020097

Pulled By: ezyang

fbshipit-source-id: 41e229d0fcc6b9bb53367be548bf85286dc13546

5 years agoctc_loss error message bug fix. (#16917)
drkw [Sun, 10 Feb 2019 18:41:46 +0000 (10:41 -0800)]
ctc_loss error message bug fix. (#16917)

Summary:
CTCLLoss argument error message is wrong.
Please fix this. (sorry if I made some mistakes.)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16917

Differential Revision: D14019983

Pulled By: ezyang

fbshipit-source-id: 3337a2e86da6f3f7594c73fddb73340494a19ce2

5 years agoUse non-Variable type for callsites that check type equality (#16325)
Will Feng [Sun, 10 Feb 2019 17:38:50 +0000 (09:38 -0800)]
Use non-Variable type for callsites that check type equality (#16325)

Summary:
When Variable and Tensor are merged, the dynamic type of the tensors passed to certain functions will become variables, and expecting `type()` on those variables to still return non-Variable types will cause type mismatch error.

One way to fix this problem is to use the thread-local guard `at::AutoNonVariableTypeMode` to force `type()` to return non-Variable type, but ideally we want to limit the use of `at::AutoNonVariableTypeMode` to be only in VariableType.cpp. Another way to fix the problem is to use `at::globalContext().getNonVariableType()` instead to get the non-Variable type of the tensor, which is what this PR is trying to achieve.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16325

Differential Revision: D14012022

Pulled By: yf225

fbshipit-source-id: 77ef1d2a02f78bff0063bdd72596e34046f1e00d

5 years agoFix the error in the note about `torch.device` documentation. (#16839)
Jiren Jin [Sun, 10 Feb 2019 04:12:32 +0000 (20:12 -0800)]
Fix the error in the note about `torch.device` documentation. (#16839)

Summary:
This PR is a simple fix for the mistake in the first note for `torch.device` in the "tensor attributes" doc.
![image](https://user-images.githubusercontent.com/8536399/52399611-1becaa00-2b00-11e9-85bf-cac04b29842d.png)

```
>>> # You can substitute the torch.device with a string
>>> torch.randn((2,3), 'cuda:1')
```
Above code will cause error like below:
```
---------------------------------------------------------------------------
TypeError                                 Traceback (most recent call last)
<ipython-input-53-abdfafb67ab1> in <module>()
----> 1 torch.randn((2,3), 'cuda:1')

TypeError: randn() received an invalid combination of arguments - got (tuple, str), but expected one of:
 * (tuple of ints size, torch.Generator generator, Tensor out, torch.dtype dtype, torch.layout layout, torch.device device, bool requires_grad)
 * (tuple of ints size, Tensor out, torch.dtype dtype, torch.layout layout, torch.device device, bool requires_grad)
```

Simply adding the argument name `device` solves the problem: `torch.randn((2,3), device='cuda:1')`.

However, another concern is that this note seems redundant as **there is already another note covering this usage**:
![image](https://user-images.githubusercontent.com/8536399/52399583-0ecfbb00-2b00-11e9-914f-e95da4edecd1.png)

So maybe it's better to just remove this note?
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16839

Reviewed By: ezyang

Differential Revision: D13989209

Pulled By: gchanan

fbshipit-source-id: ac255d52528da053ebfed18125ee6b857865ccaf

5 years agoRegister coalescer bug was fixed in ROCm 2.1 (#16923)
Johannes M Dieterich [Sat, 9 Feb 2019 19:20:18 +0000 (11:20 -0800)]
Register coalescer bug was fixed in ROCm 2.1 (#16923)

Summary:
Remove specialization/workaround for ROCm.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16923

Differential Revision: D14018521

Pulled By: bddppq

fbshipit-source-id: d88162740bca6dc8ad37397dfbf8c84408074a00

5 years agoAlignas is now correctly handled on ROCm (#16920)
Johannes M Dieterich [Sat, 9 Feb 2019 19:19:30 +0000 (11:19 -0800)]
Alignas is now correctly handled on ROCm (#16920)

Summary:
Post 2.1 release, packing is fixed and alignas works as expected.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16920

Differential Revision: D14018539

Pulled By: bddppq

fbshipit-source-id: 0ed4d9e9f36afb9b970812c3870082fd7f905455

5 years agoEnable buildin bitonic sort (#16919)
Johannes M Dieterich [Sat, 9 Feb 2019 19:16:05 +0000 (11:16 -0800)]
Enable buildin bitonic sort (#16919)

Summary:
It now works post ROCm 2.1 release.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16919

Differential Revision: D14018538

Pulled By: bddppq

fbshipit-source-id: c4e1bafb53204a6d718b2d5054647d5715f23243

5 years agoChange the default image size from 227 to 224 in resnet50 trainer (#16924)
Junjie Bai [Sat, 9 Feb 2019 19:15:29 +0000 (11:15 -0800)]
Change the default image size from 227 to 224 in resnet50 trainer (#16924)

Summary:
cc xw285cornell
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16924

Differential Revision: D14018509

Pulled By: bddppq

fbshipit-source-id: fdbc9e94816ce6e4b1ca6f7261007bda7b80e1e5

5 years agoenable unit tests working on ROCm 2.1 (#16871)
Johannes M Dieterich [Sat, 9 Feb 2019 08:09:33 +0000 (00:09 -0800)]
enable unit tests working on ROCm 2.1 (#16871)

Summary:
This is the first round of enabling unit tests that work on ROCm 2.1 in my tests.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16871

Differential Revision: D13997662

Pulled By: bddppq

fbshipit-source-id: d909a3f7dd5fc8f85f126bf0613751c8e4ef949f

5 years agoAdd suggest add to __constants__ message on save fail
Elias Ellison [Sat, 9 Feb 2019 03:06:41 +0000 (19:06 -0800)]
Add suggest add to __constants__ message on save fail

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/16850

Differential Revision: D14014735

Pulled By: eellison

fbshipit-source-id: 7b6d5d5b64b9b107743cea1548cb4ee1b653977e

5 years agoMake the exception raised from "numpy.dtype(numpy.void, (INT,))" less cryptic (#16809)
Chandler Zuo [Fri, 8 Feb 2019 23:20:15 +0000 (15:20 -0800)]
Make the exception raised from "numpy.dtype(numpy.void, (INT,))" less cryptic (#16809)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16809

https://fb.facebook.com/groups/582508038765902/permalink/736710343345670/?comment_id=824042307945806&reply_comment_id=824318864584817

numpy.dtype(numpy.void, (<INT>, )) raises a cryptic message "invalid itemsize in generic type tuple" that is hard to debug.

This diff adds the message to ask the user to investigate the error causing blob.

Reviewed By: kennyhorror

Differential Revision: D13973359

fbshipit-source-id: 43a0c492ffafbabdfd7f7541c08a258e5ac0280f

5 years agoRevert D13970381: [caffe2] Add visibility to registry class to fix ubsan error
Bram Wasti [Fri, 8 Feb 2019 22:57:57 +0000 (14:57 -0800)]
Revert D13970381: [caffe2] Add visibility to registry class to fix ubsan error

Differential Revision:
D13970381

Original commit changeset: 763db24b8a98

fbshipit-source-id: dda8672ed0bc6fecc4dde5ce73feb99e15205978

5 years agoExtend Net.RunAllOnGPU() to support RecurrentNetwork op (#15713)
Nikita Shulga [Fri, 8 Feb 2019 22:20:31 +0000 (14:20 -0800)]
Extend Net.RunAllOnGPU() to support RecurrentNetwork op (#15713)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15713

[caffe2] Extend Net.RunAllOnGPU() to support RecurrentNetwork op

Reviewed By: dzhulgakov

Differential Revision: D13576507

fbshipit-source-id: f517127492c9d516ece663d42fef84338c70344e

5 years agodelete critical section in TH*Tensor_addmm (#16889)
James Reed [Fri, 8 Feb 2019 21:45:43 +0000 (13:45 -0800)]
delete critical section in TH*Tensor_addmm (#16889)

Summary:
This was serializing all calls to `addmm` (and any op that used it, in my case `bmm`) in the entire process, and led to downright atrocious performance in the TorchScript threaded runtime. Removing this gives a 2x throughput boost for high-load machine translation inference.

The original justification for this is dubious: there are other `gemm` callsites in the codebase that are not protected by critical sections. And in caffe2 land we never had any issues with nonreentrant BLAS libraries
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16889

Differential Revision: D14008928

Pulled By: jamesr66a

fbshipit-source-id: 498e2133bd6564dba539a2d9751f4e61afbce608

5 years agoRevert D13806753: [pytorch][PR] TensorIterator cuda launch configs update
Bram Wasti [Fri, 8 Feb 2019 19:42:22 +0000 (11:42 -0800)]
Revert D13806753: [pytorch][PR] TensorIterator cuda launch configs update

Differential Revision:
D13806753

Original commit changeset: 37e45c7767b5

fbshipit-source-id: 74ac9f54f86853287b372ccf21fb37ed0e04a5d3

5 years agoAllow sequential modules in module list (#16882)
Elias Ellison [Fri, 8 Feb 2019 19:34:40 +0000 (11:34 -0800)]
Allow sequential modules in module list (#16882)

Summary:
Fix for https://github.com/pytorch/pytorch/issues/16845
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16882

Differential Revision: D14007746

Pulled By: eellison

fbshipit-source-id: d7918275cc1de6a67320619c3203463f66783343

5 years agoImpl ExpandDims op and fallback to CPU if needed (#15264)
Gu, Jinghui [Fri, 8 Feb 2019 19:17:59 +0000 (11:17 -0800)]
Impl ExpandDims op and fallback to CPU if needed (#15264)

Summary:
Impl ExpandDims op and fallback to CPU if needed
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15264

Differential Revision: D13808797

Pulled By: yinghai

fbshipit-source-id: 7795ec303a46e85f84e5490273db0ec76e8b9374

5 years agoAdd visibility to registry class to fix ubsan error (#16792)
Bram Wasti [Fri, 8 Feb 2019 18:00:49 +0000 (10:00 -0800)]
Add visibility to registry class to fix ubsan error (#16792)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16792

fix

Reviewed By: ezyang

Differential Revision: D13970381

fbshipit-source-id: 763db24b8a98a2757a63b77c70c8c68ba47f31e6

5 years agoRemove Legacy entry point. (#16721)
Edward Yang [Fri, 8 Feb 2019 17:29:59 +0000 (09:29 -0800)]
Remove Legacy entry point. (#16721)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16721

The very key line is we have to set the stream to the default
stream before calling the allocator.  This is very interesting.
It shouldn't be necessary, but seemingly is!

Reviewed By: dzhulgakov

Differential Revision: D13943193

fbshipit-source-id: c21014917d9fe504fab0ad8abbc025787f559287

5 years agoDeduplicate instances caching allocator, so that we only have one instance. (#16720)
Edward Yang [Fri, 8 Feb 2019 17:29:59 +0000 (09:29 -0800)]
Deduplicate instances caching allocator, so that we only have one instance. (#16720)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16720

I'm taking the deduplication slowly because there is something here
that is causing problems, and I want to figure out what it is.

Reviewed By: dzhulgakov

Differential Revision: D13943194

fbshipit-source-id: cbc08fee5862fdcb393b9dd5b1d2ac7250f77c4b

5 years agoDelete duplicate copy of THCCachingAllocator (round two). (#16615)
Edward Yang [Fri, 8 Feb 2019 17:29:58 +0000 (09:29 -0800)]
Delete duplicate copy of THCCachingAllocator (round two). (#16615)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16615

This is another go at landing https://github.com/pytorch/pytorch/pull/16226
Now that the caching allocator is moved to c10_cuda, we can
delete the duplicate copy from Caffe2.

The difference between this and the previous PR is that this
version faithfully maintains the binding code; in particular,
we end up with a SECOND copy of the caching allocator in
this patch.  I verified that this code does NOT cause a crash
in the workflow we canaried last time.

In further diffs, I plan to eliminate the second copy, and then
adjust the binding code.

Reviewed By: dzhulgakov

Differential Revision: D13901067

fbshipit-source-id: 66331fd4eadffd0a5defb3cea532d5cd07287872

5 years agoBump caffe2 docker images to 248 (#16863)
Junjie Bai [Fri, 8 Feb 2019 08:32:49 +0000 (00:32 -0800)]
Bump caffe2 docker images to 248 (#16863)

Summary:
Jenkins jobs update will be separate.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16863

Differential Revision: D13994672

Pulled By: bddppq

fbshipit-source-id: 5b27879dc6ac11a42016fe7835e9124345005ebb

5 years agoAlso register op schema when no kernels are registered
Sebastian Messmer [Fri, 8 Feb 2019 04:47:46 +0000 (20:47 -0800)]
Also register op schema when no kernels are registered

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/16878

Reviewed By: bwasti

Differential Revision: D13997959

fbshipit-source-id: 7527a560b03f672f76e95d4f22ae28ce24698cc1

5 years agoDon't automatically handle context parameter (#16867)
Sebastian Messmer [Fri, 8 Feb 2019 04:47:45 +0000 (20:47 -0800)]
Don't automatically handle context parameter (#16867)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16867

Some caffe2 operators (example: BBoxTransform) have not just one template parameter which is the context, but might have multiple template parameters.
Because of this, we can't handle the context parameter inside the macro.

Reviewed By: bwasti

Differential Revision: D13995696

fbshipit-source-id: f55c3be913c8b125445a8d486846fc2fab587a63

5 years agoSupport onnxifi with partially shaped inferred net (#16877)
Yinghai Lu [Fri, 8 Feb 2019 04:41:18 +0000 (20:41 -0800)]
Support onnxifi with partially shaped inferred net (#16877)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16877

That's it.

Reviewed By: ipiszy

Differential Revision: D13997771

fbshipit-source-id: f512c7f30b4a4747aca335a0769712c2a2cc2206

5 years agoRobust determination of cudnn library and relevant conda packages. (#16859)
Pearu Peterson [Fri, 8 Feb 2019 04:27:40 +0000 (20:27 -0800)]
Robust determination of cudnn library and relevant conda packages. (#16859)

Summary:
This PR implements:
1. a fix to issue #12174 - determine the location of cudnn library using `ldconfig`
2. a fix to determine the installed conda packages (in recent versions of conda, the command `conda` is a Bash function that cannot be called within a python script, so using CONDA_EXE environment variable instead)
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16859

Differential Revision: D14000399

Pulled By: soumith

fbshipit-source-id: 905658ecacb0ca0587a162fade436de9582d32ab

5 years agoSpecialize LengthsRangeFill and SparseLengthsWeightedSum in bound shape inference...
Yinghai Lu [Fri, 8 Feb 2019 04:08:39 +0000 (20:08 -0800)]
Specialize LengthsRangeFill and SparseLengthsWeightedSum in bound shape inference (#16869)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16869

TSIA.

Reviewed By: ipiszy, rdzhabarov

Differential Revision: D13994946

fbshipit-source-id: 7e507abc5a3c2834c92910e521387085c56e8b2e

5 years agoActivation histogram net observer with multiple histogram files as output (#16855)
Summer Deng [Fri, 8 Feb 2019 03:42:17 +0000 (19:42 -0800)]
Activation histogram net observer with multiple histogram files as output (#16855)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16855

Save the histogram of each net to a separate file

Reviewed By: jspark1105

Differential Revision: D13991610

fbshipit-source-id: a5be4e37a5e63567dcd7fdf99f451ee31bb350a5

5 years agoAllow dicts in C++ frontend (#16846)
David Riazati [Fri, 8 Feb 2019 02:21:30 +0000 (18:21 -0800)]
Allow dicts in C++ frontend (#16846)

Summary:
Fixes #16856
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16846

Differential Revision: D13991103

Pulled By: driazati

fbshipit-source-id: 4830dd6f707fa90429b5d3070eeda0bee53d2f2b

5 years agoSeparate elementwise level2 math functions (#16753)
Xiaomeng Yang [Fri, 8 Feb 2019 02:19:46 +0000 (18:19 -0800)]
Separate elementwise level2 math functions (#16753)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16753

Separate elementwise level2 math functions

i-am-not-moving-c2-to-c10

Reviewed By: houseroad

Differential Revision: D13954928

fbshipit-source-id: 1ca7a5d3da96e32510f502e5e4e79168854bee67

5 years agoFix (#2) ppc64le build break on git status --porcelain check (#16852)
Freddie Mendoza [Fri, 8 Feb 2019 02:15:44 +0000 (18:15 -0800)]
Fix (#2) ppc64le build break on git status --porcelain check (#16852)

Summary:
Add test/.hypothesis/ to .gitignore to pass git status --porcelain check in CI build
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16852

Differential Revision: D14000206

Pulled By: soumith

fbshipit-source-id: 5da99a4bb242c12aa35776f7254f6399a7fa6d8c

5 years agodoc updates for TorchScript (#16866)
Michael Suo [Fri, 8 Feb 2019 01:56:10 +0000 (17:56 -0800)]
doc updates for TorchScript (#16866)

Summary:
Some batched updates:
1. bool is a type now
2. Early returns are allowed now
3. The beginning of an FAQ section with some guidance on the best way to do GPU training + CPU inference
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16866

Differential Revision: D13996729

Pulled By: suo

fbshipit-source-id: 3b884fd3a4c9632c9697d8f1a5a0e768fc918916

5 years agoFix autodiff of nll_loss
Alex Şuhan [Fri, 8 Feb 2019 01:31:52 +0000 (17:31 -0800)]
Fix autodiff of nll_loss

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/16851

Differential Revision: D13995046

Pulled By: wanchaol

fbshipit-source-id: 557c99f1d1825fa9b6031dd9fa8ba9b54205e8c4

5 years agoaten::_convolution now participates in shape analysis (#16837)
James Reed [Fri, 8 Feb 2019 01:22:00 +0000 (17:22 -0800)]
aten::_convolution now participates in shape analysis (#16837)

Summary:
During tracing, we record `aten::_convolution` rather than `aten::convolution`. The schema for the former was not present in the shape analysis pass, and resulted in some missing shape information.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16837

Differential Revision: D13993831

Pulled By: jamesr66a

fbshipit-source-id: ebb63bf628d81613258caf773a3af5930303ce5a

5 years agoEnable arg_ops_test/unique_ops_test on AMD/rocm (#16853)
peter.yeh@amd.com [Fri, 8 Feb 2019 00:10:50 +0000 (16:10 -0800)]
Enable arg_ops_test/unique_ops_test on AMD/rocm (#16853)

Summary:
Verified both tests are passing on rocm 2.1 env.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16853

Differential Revision: D13996279

Pulled By: bddppq

fbshipit-source-id: c0df610d7d9ca8d80ed2d1339cdadef59105a71c

5 years agoUpdate CI to recently released ROCm 2.1 release (#16808)
Johannes M Dieterich [Thu, 7 Feb 2019 22:17:14 +0000 (14:17 -0800)]
Update CI to recently released ROCm 2.1 release (#16808)

Summary:
* we do not need EAP packages any longer as the antistatic feature is now in the release
* consistently install the rccl package
* Skip one unit test that has regressed with 2.1
* Follow-up PRs will use 2.1 features once deployed on CI
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16808

Differential Revision: D13992645

Pulled By: bddppq

fbshipit-source-id: 37ca9a1f104bb140bd2b56d403e32f04c4fbf4f0

5 years agoUse bound shape inference in SparseNN tests (#16834)
Yinghai Lu [Thu, 7 Feb 2019 22:11:44 +0000 (14:11 -0800)]
Use bound shape inference in SparseNN tests (#16834)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16834

Inserting AdjustBatch ops will possibly change the names of the input/output, so we need to create a mapping and use the renamed names for external_inputs/outputs and input_shape_info for the onnxifi_net.

Reviewed By: ipiszy

Differential Revision: D13982731

fbshipit-source-id: c18b8a03d01490162929b2ca30c182d166001626

5 years agoAdd recognition for XLA device types.
Davide Libenzi [Thu, 7 Feb 2019 22:11:33 +0000 (14:11 -0800)]
Add recognition for XLA device types.

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/16844

Differential Revision: D13988805

Pulled By: gchanan

fbshipit-source-id: 4e89d6d2cde8bdac41739efa65cc91569a360953

5 years agoFix and re-enable test case (#16643)
Sebastian Messmer [Thu, 7 Feb 2019 21:52:49 +0000 (13:52 -0800)]
Fix and re-enable test case (#16643)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16643

The test was disabled in D13908117 because it conflicted with another diff that was about to land.
Now fixed the merge conflict and re-landing it.

Reviewed By: ezyang

Differential Revision: D13911775

fbshipit-source-id: b790f1c3a3f207916eea41ac93bc104d011f629b

5 years agoC10_REGISTER_CAFFE2_OPERATOR: Macro for registering c2 kernels (#16548)
Sebastian Messmer [Thu, 7 Feb 2019 21:52:49 +0000 (13:52 -0800)]
C10_REGISTER_CAFFE2_OPERATOR: Macro for registering c2 kernels (#16548)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16548

With this macro, a caffe2 operator can now directly be registered with c10.
No need to write custom wrapper kernels anymore.

Differential Revision: D13877076

fbshipit-source-id: e56846238c5bb4b1989b79855fd44d5ecf089c9c

5 years agoFix Anaconda logins on binary builds
Jesse Hellemn [Thu, 7 Feb 2019 21:39:18 +0000 (13:39 -0800)]
Fix Anaconda logins on binary builds

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/16848

Differential Revision: D13993614

Pulled By: pjh5

fbshipit-source-id: 16854b06d01460b78d9dbe7bd0341b7332984795

5 years agonew embedding label type in image input op (#16835)
Zhicheng Yan [Thu, 7 Feb 2019 21:08:06 +0000 (13:08 -0800)]
new embedding label type in image input op (#16835)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16835

We were using label type `multi_label_dense` to denote both 1) dense representation of integer label 2) embedding label of data type floating number.

This cause some issues as two cases have different assumption, such as for integer label, we will check whether label value is in [0, number_class - 1]. But such check should be skipped for `embedding label`.

Reviewed By: BIT-silence

Differential Revision: D13985048

fbshipit-source-id: 1202cdfeea806eb47647e3f4a1ed9c104f72ad2c

5 years agoUpdate ATen internals to use int64_t for dimension indexing (#16739)
Michael Antonov [Thu, 7 Feb 2019 20:42:38 +0000 (12:42 -0800)]
Update ATen internals to use int64_t for dimension indexing (#16739)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16739

Some code ATen locations seemed to use int, etc. inclorrectly where either
int64_t or size_t was required. Update them to use int64_t for dimension indexing where necessary.

Reviewed By: ezyang

Differential Revision: D13950124

fbshipit-source-id: aaf1cef783bf3c657aa03490f2616c35c816679f

5 years agoMake JIT attributes t_ and ts_ store Variable instead of Tensor (#16596)
Will Feng [Thu, 7 Feb 2019 19:58:50 +0000 (11:58 -0800)]
Make JIT attributes t_ and ts_ store Variable instead of Tensor (#16596)

Summary:
Discussed with zdevito and we want to use Variable (with `set_requires_grad(false)`) instead of Tensor in all parts of JIT, to eliminate the distinction and the conceptual overhead when trying to figure out which one to use.

This also helps with the Variable/Tensor merge work tracked at https://github.com/pytorch/pytorch/issues/13638, which will make common functions (such as `numel()` / `sizes()` / `dim()`) on Variable much faster when finished.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16596

Differential Revision: D13979971

Pulled By: yf225

fbshipit-source-id: c69119deec5bce0c22809081115f1012fdbb7d5a

5 years agoBetter error when using a constant tensor (#16724)
David Riazati [Thu, 7 Feb 2019 19:50:27 +0000 (11:50 -0800)]
Better error when using a constant tensor (#16724)

Summary:
Fixes #16284
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16724

Differential Revision: D13990531

Pulled By: driazati

fbshipit-source-id: adbf47a07eddb3813fbe1322944abfe5fcff89fa

5 years agoBackport the stable doc build on v1.0.1 to master (#16503)
Richard Zou [Thu, 7 Feb 2019 19:09:10 +0000 (11:09 -0800)]
Backport the stable doc build on v1.0.1 to master (#16503)

Summary:
List of changes:
- Always push the final state of the doc build docker for debugging purposes.
- Adds code for the stable doc build. This code is never actually run on master, only the v1.0.1 branch. There is a big note for this behavior.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16503

Differential Revision: D13972469

Pulled By: zou3519

fbshipit-source-id: 68f459650ef0de200a34edd43fc1372143923972

5 years agoRemove undefined tensor in jit script (#16379)
Wanchao Liang [Thu, 7 Feb 2019 18:32:02 +0000 (10:32 -0800)]
Remove undefined tensor in jit script (#16379)

Summary:
This PR is a follow up of #15460, it did the following things:

* remove the undefined tensor semantic in jit script/tracing mode
* change ATen/JIT schema for at::index and other index related ops with `Tensor?[]` to align with what at::index is really doing and to adopt `optional[tensor]` in JIT
* change python_print to correctly print the exported script
* register both TensorList and ListOfOptionalTensor in JIT ATen ops to support both
* Backward compatibility for `torch.jit.annotate(Tensor, None)`

List of follow ups:

* remove the undefined tensor semantic in jit autograd, autodiff and grad_of
* remove prim::Undefined fully

For easy reviews, please turn on `hide white space changes` in diff settings.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16379

Differential Revision: D13855677

Pulled By: wanchaol

fbshipit-source-id: 0e21c14d7de250c62731227c81bfbfb7b7da20ab

5 years agoSupport multiple inheritance in torch.distributions (#16772)
Fritz Obermeyer [Thu, 7 Feb 2019 09:33:41 +0000 (01:33 -0800)]
Support multiple inheritance in torch.distributions (#16772)

Summary:
This adds calls to `super().__init__()` in three classes in torch.distributions.

This is needed when `Distribution` and `Transform` objects are used with multiple inheritance, as e.g. combined with `torch.nn.Module`s. For example
```py
class MyModule(torch.distributions.Transform, torch.nn.Module):
    ...
```
cc  martinjankowiak esling who have wanted to use this pattern, e.g. in #16756
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16772

Differential Revision: D13978633

Pulled By: soumith

fbshipit-source-id: 8bc6cca1747cd74d32135ee2fe588bba2ea796f1

5 years agoRemove redundant wrappers in torch.distributions (#16807)
vishwakftw [Thu, 7 Feb 2019 09:10:54 +0000 (01:10 -0800)]
Remove redundant wrappers in torch.distributions (#16807)

Summary:
Changelog:
- Remove torch.distributions.multivariate_normal._batch_diag : same functionality is provided by torch.diagonal
- Remove torch.distributions.lowrank_multivariate_normal._batch_vector_diag : same functionality is provided by torch.diag_embed
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16807

Differential Revision: D13985550

Pulled By: soumith

fbshipit-source-id: 25c7d00c52ff7f85e431134e9ce0d5dda453667b

5 years agoInsert AdjustBatchSizeOp into the predict_net. (#16811)
Ying Zhang [Thu, 7 Feb 2019 08:33:29 +0000 (00:33 -0800)]
Insert AdjustBatchSizeOp into the predict_net. (#16811)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16811

As the title. The AdjustBatch ops will be inserted before and after the Onnxifi op to:
1) adjust batch/seq sizes to the ideal batch/seq size before these tensors are processed by the Onnxifi op;
2) adjust batch size to the original batch size for batches generated by the Onnxifi op.

Reviewed By: yinghai

Differential Revision: D13967711

fbshipit-source-id: 471b25ae6a60bf5b7ebee1de6449e0389b6cafff

5 years agoUnify gpu_support variable in python tests (#16748)
rohithkrn [Thu, 7 Feb 2019 08:21:21 +0000 (00:21 -0800)]
Unify gpu_support variable in python tests (#16748)

Summary:
Assign `has_gpu_support = has_cuda_support or has_hip_support` and make according changes in python tests.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16748

Differential Revision: D13983132

Pulled By: bddppq

fbshipit-source-id: ca496fd8c6ae3549b736bebd3ace7fa20a6dad7f

5 years agoUpdate Docker file section in README.md (#16812)
Mohana Rao [Thu, 7 Feb 2019 07:33:40 +0000 (23:33 -0800)]
Update Docker file section in README.md (#16812)

Summary:
Emphasize on the fact that docker build should be triggered from pytorch repo directory.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16812

Differential Revision: D13985531

Pulled By: soumith

fbshipit-source-id: c6511d1e81476eb795b37fb0ad23e8951dbca617

5 years agoTensorIterator cuda launch configs update (#16224)
Jie [Thu, 7 Feb 2019 07:05:49 +0000 (23:05 -0800)]
TensorIterator cuda launch configs update (#16224)

Summary:
Update launch configs for TensorIterator gpu_reduce_kernel. Enable flexible
block dimension to improve efficiency for reduction cases with small fast
dimension.

Previously TensorIterator launches blocks with fixed 32x16 threads.
For cases like:

  import torch
  torch.randn(2**20, 4, device='cuda').sum(0)

The fixed launch config does handle coalesced memory access efficiently.

Updated launch configure enables flexible block dimension. Combining with
improved reduction scheme (using flexible vertical / horizontal reduction
instead of limited warp / block reduction in the old code), it ensures optimal
memory access pattern even with reduction on dimension with small stride.

Possible future improvements:
1. Precise dynamic shared memory allocation.
2. Using warp shuffle for vertical (block_y) reduction.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16224

Differential Revision: D13806753

Pulled By: soumith

fbshipit-source-id: 37e45c7767b5748cf9ecf894fad306e040e2f79f

5 years agoDefine layer_norm schema in caffe2 (#16535)
Sebastian Messmer [Thu, 7 Feb 2019 05:14:21 +0000 (21:14 -0800)]
Define layer_norm schema in caffe2 (#16535)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16535

There is now no need anymore to define the layer norm schema in a central location.
It can just be defined in caffe2 next to the kernel implementation.

Reviewed By: ezyang

Differential Revision: D13869503

fbshipit-source-id: c478153f8fd712ff6d507c794500286eb3583149

5 years agoAutomatically register c10 ops with JIT (#16534)
Sebastian Messmer [Thu, 7 Feb 2019 05:14:20 +0000 (21:14 -0800)]
Automatically register c10 ops with JIT (#16534)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16534

All c10 ops from the c10 dispatcher are now automatically registered with JIT

Reviewed By: dzhulgakov

Differential Revision: D13869275

fbshipit-source-id: 5ab5dec5b983fe661f977f9d29d8036768cdcab6

5 years agoAdd AdjustBatch Op (#16676)
Yinghai Lu [Thu, 7 Feb 2019 03:12:32 +0000 (19:12 -0800)]
Add AdjustBatch Op (#16676)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16676

This op is used for changing batch size (first dimension) of the tensor.

Reviewed By: bertmaher, ipiszy

Differential Revision: D13929200

fbshipit-source-id: 4f2c3faec072d468be8301bf00c80d33adb3b5b3

5 years agoBring back running pytorch tests in rocm CI
bddppq [Thu, 7 Feb 2019 01:52:12 +0000 (17:52 -0800)]
Bring back running pytorch tests in rocm CI

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/16829

Differential Revision: D13982323

Pulled By: bddppq

fbshipit-source-id: 6ffadb96b9e2ebd64a29e38674a51401dfb211db

5 years agoRename DynamicType -> TensorType (#16787)
Zachary DeVito [Thu, 7 Feb 2019 01:22:47 +0000 (17:22 -0800)]
Rename DynamicType -> TensorType (#16787)

Summary:
```
import json
from subprocess import check_call
from pprint import pprint
renames = {
    'c10::TensorType': 'DimentionedTensorType',
    'c10::DynamicType': 'TensorType',
    'c10::TensorTypePtr': 'DimentionedTensorTypePtr',
    'c10::DynamicTypePtr': 'TensorTypePtr',
    'c10::TypeKind::DynamicType': 'TensorType',
    'c10::TypeKind::TensorType': 'DimentionedTensorType',
}

entries = json.loads(open('compile_commands.json', 'r').read())

build = None
sources = []

for e in entries:
    name = e['file']
    if not ('jit' in name or 'ATen/core' in name):
        continue
    build = e['directory']
    sources.append(name)

args = ['clang-rename', '-i', '-force', '-pl']
for name in sorted(renames.keys()):
    args += ['-qualified-name={}'.format(name), '-new-name={}'.format(renames[name])]

for source in sources:
    cmd = args + [source]
    pprint(args)
    check_call(cmd, cwd=build)
    check_call(['git', 'stash', 'push', '-m', 'rename'])
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16787

Differential Revision: D13974132

Pulled By: zdevito

fbshipit-source-id: 8368fd53e17cff83707bbe77f2d7aad74f8ce60e

5 years agoUse bound shape inference in onnxifi transform (#16598)
Yinghai Lu [Thu, 7 Feb 2019 00:13:24 +0000 (16:13 -0800)]
Use bound shape inference in onnxifi transform (#16598)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16598

ATT.

Reviewed By: bertmaher, rdzhabarov

Differential Revision: D13893698

fbshipit-source-id: 8d2ad9814fe76924a46b450eb7ebd3601fbdbbc7

5 years agoimprove error message (#16719)
Soumith Chintala [Wed, 6 Feb 2019 23:46:39 +0000 (15:46 -0800)]
improve error message (#16719)

Summary:
Fixes https://github.com/pytorch/pytorch/issues/16712
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16719

Differential Revision: D13978688

Pulled By: ezyang

fbshipit-source-id: 61f8fa4c54c6969a58550e32e18be2eb9254ced7

5 years agoint8 SpatialBN (#16796)
Jongsoo Park [Wed, 6 Feb 2019 23:14:17 +0000 (15:14 -0800)]
int8 SpatialBN (#16796)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16796

SpatialBN int8 version

Reviewed By: dskhudia

Differential Revision: D13971224

fbshipit-source-id: e55fd608c161069daaa4e62c618bc14b01f32cb7

5 years agocall istringstream clear after str (#16820)
Jongsoo Park [Wed, 6 Feb 2019 23:10:07 +0000 (15:10 -0800)]
call istringstream clear after str (#16820)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16820

Sometimes parsing histogram was not working correctly due to changes in D13633256
We need to call istringstream clear after str

Reviewed By: csummersea

Differential Revision: D13977509

fbshipit-source-id: ce3e8cb390641d8f0b5c9a7d6d6daadffeddbe11

5 years agoReplace resize_dim() with set_sizes_and_strides() (#16732)
Narine Kokhlikyan [Wed, 6 Feb 2019 22:29:35 +0000 (14:29 -0800)]
Replace resize_dim() with set_sizes_and_strides() (#16732)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16732

Use set_sizes_and_strides instead of resize_dim with.

Reviewed By: ezyang

Differential Revision: D13947867

fbshipit-source-id: 067b096b1fde14b039690992a5fe3ace386b2789

5 years agono EIGEN engine for DeformConv (#16785)
Jongsoo Park [Wed, 6 Feb 2019 19:52:23 +0000 (11:52 -0800)]
no EIGEN engine for DeformConv (#16785)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16785

There's no EIGEN engine implemented for DeformConv but unit test was checking it.

Reviewed By: BIT-silence

Differential Revision: D13967306

fbshipit-source-id: e29c19f59f5700fc0501c59f45d60443b87ffedc

5 years agoformat deform_conv_test.py (#16786)
Jongsoo Park [Wed, 6 Feb 2019 19:52:23 +0000 (11:52 -0800)]
format deform_conv_test.py (#16786)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16786

Format to prepare D13967306

Reviewed By: BIT-silence

Differential Revision: D13967317

fbshipit-source-id: 2de895f8474b04c55ba067fbf788c553dc010c60

5 years agoFix/Improve bound shape inference with real net tests (#16597)
Yinghai Lu [Wed, 6 Feb 2019 18:23:01 +0000 (10:23 -0800)]
Fix/Improve bound shape inference with real net tests (#16597)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16597

This diff fixes some bugs in shape inference for `SparseLengthsSumFused8BitRowwise`. And added input shape inference for `Concat` when `add_axis=1`.

Reviewed By: bertmaher

Differential Revision: D13892452

fbshipit-source-id: 6cd95697a6fabe6d78a5ce3cb749a3a1e51c68e7

5 years agoTypofix (#16800)
Edward Yang [Wed, 6 Feb 2019 18:19:37 +0000 (10:19 -0800)]
Typofix (#16800)

Summary:
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16800

Differential Revision: D13972592

Pulled By: ezyang

fbshipit-source-id: 45c352ac6090c8060bf75f44dec7205556986d88

5 years agocaffe2 | MSVS compatibility fixes (#16765)
Oleg Bogdanov [Wed, 6 Feb 2019 17:40:00 +0000 (09:40 -0800)]
caffe2 | MSVS compatibility fixes (#16765)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16765

Code changes required to build caffe2 for windows with toolchain used by FB.

Reviewed By: orionr

Differential Revision: D13953258

fbshipit-source-id: 651823ec9d81ac70e32d4cce5bc2472434104733

5 years agoFallback sum/add to CPU if needed (#15267)
Gu, Jinghui [Wed, 6 Feb 2019 17:25:42 +0000 (09:25 -0800)]
Fallback sum/add to CPU if needed (#15267)

Summary:
Fallback sum/add to CPU if needed
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15267

Differential Revision: D13935064

Pulled By: yinghai

fbshipit-source-id: eb228683d00a0462a1970f849d35365bc98340d6

5 years agoAutomatic update of fbcode/onnx to 822d8df0a2a32233c6022f50a158817a0f19bdc7 (#16791)
Lu Fang [Wed, 6 Feb 2019 17:17:37 +0000 (09:17 -0800)]
update of fbcode/onnx to 822d8df0a2a32233c6022f50a158817a0f19bdc7 (#16791)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16791

Previous import was bfa8b335ab6d1ed7b688dc2ec96421a3fe9e644c

Included changes:
- **[822d8df](https://github.com/onnx/onnx/commit/822d8df)**: allow removed experimental ops in the checker for now (#1792) <Lu Fang>

Reviewed By: MisterTea

Differential Revision: D13970103

fbshipit-source-id: 5feaaa6c65ba10901eeba0b63724d7e451e9b642

5 years agoAdding torch/lib64 in .gitignore for ppc64le CI build to pass (#16782)
Freddie Mendoza [Wed, 6 Feb 2019 17:02:18 +0000 (09:02 -0800)]
Adding torch/lib64 in .gitignore for ppc64le CI build to pass (#16782)

Summary:
Adding torch/lib64 in .gitignore so that a git status --porcelain
check during CI build and test passes for ppc64le. During build
torch/lib64 is created and contains third-party libraries. This
should be ignored by the porcelain check
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16782

Differential Revision: D13972794

Pulled By: ezyang

fbshipit-source-id: 5459c524eca42d396ac46e756a327980b4b1fa53

5 years agoRevert D13854304: [redo][c10] LayerNorm Registration Example
Edward Yang [Wed, 6 Feb 2019 16:17:55 +0000 (08:17 -0800)]
Revert D13854304: [redo][c10] LayerNorm Registration Example

Differential Revision:
D13854304

Original commit changeset: ec463ce22721

fbshipit-source-id: 4262b9a2ef486e1c7c0283ea021331ac97cc5f56

5 years agoRevert D13855525: [c10] Expose RoIAlign to torch
Edward Yang [Wed, 6 Feb 2019 16:17:54 +0000 (08:17 -0800)]
Revert D13855525: [c10] Expose RoIAlign to torch

Differential Revision:
D13855525

Original commit changeset: cfee7bb1544d

fbshipit-source-id: 0b4124b78c4082b52e592a1275069c879a9aed39