platform/upstream/pytorch.git
5 years agoslots with explicit value/setValue make more sense in future patches (#18468)
Zachary DeVito [Fri, 5 Apr 2019 20:33:14 +0000 (13:33 -0700)]
slots with explicit value/setValue make more sense in future patches (#18468)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18468
ghimport-source-id: d4b41c521f2269a695e03c8e7d05d5542731ee48

Stack from [ghstack](https://github.com/ezyang/ghstack):
* #18469 Create Object that represents a Module
* **#18468 slots with explicit value/setValue make more sense in future patches**
* #18467 Make Object hold its ClassType
* #18379 Enforce single parent for script submodules
* #18378 Unify namespace of script::Module
* #18314 Add ability to specialize class types to ArgumentSpec
* #18226 Add Slot type to abstract the raw pointers being used for slots.

Reviewed By: suo

Differential Revision: D14613509

fbshipit-source-id: 9f2208d0efd01465c78cebdc3e8365a9e0adf9ff

5 years agoMake Object hold its ClassType (#18467)
Zachary DeVito [Fri, 5 Apr 2019 20:33:14 +0000 (13:33 -0700)]
Make Object hold its ClassType (#18467)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18467
ghimport-source-id: d51bdd64d2529d08c634c58df1a0870b54ad49fb

Stack from [ghstack](https://github.com/ezyang/ghstack):
* #18469 Create Object that represents a Module
* #18468 slots with explicit value/setValue make more sense in future patches
* **#18467 Make Object hold its ClassType**
* #18379 Enforce single parent for script submodules
* #18378 Unify namespace of script::Module
* #18314 Add ability to specialize class types to ArgumentSpec
* #18226 Add Slot type to abstract the raw pointers being used for slots.

Currently it holds a symbol whose unqualified name is the name of the
class. This will get confusing when there are multiple possible registries,
and it makes getting the class type from the object difficult.
The pointer to the class is only 4 more bytes so this patch just puts
it in the object.

Reviewed By: suo

Differential Revision: D14613510

fbshipit-source-id: b35175ba4be83d2522deaa6dad5070d6ec691fed

5 years agoEnforce single parent for script submodules (#18379) (#18860)
Zachary DeVito [Fri, 5 Apr 2019 20:33:14 +0000 (13:33 -0700)]
Enforce single parent for script submodules (#18379) (#18860)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18860
ghimport-source-id: 96305349bf3db564f43df2263b1e5bddcc9e9dae

Reviewed By: suo

Differential Revision: D14780421

Pulled By: zdevito

fbshipit-source-id: 2bdd89b35866ba035ebea0adab037e441c1006e2

5 years agoCUDA_NVCC_EXECUTABLE is not needed, as nvcc is in PATH (#18958)
Stas Bekman [Fri, 5 Apr 2019 19:46:44 +0000 (12:46 -0700)]
CUDA_NVCC_EXECUTABLE is not needed, as nvcc is in PATH (#18958)

Summary:
As indicated by f0k: https://github.com/pytorch/pytorch/pull/18495#issuecomment-480178763
nvcc via ccache is already first in the PATH in the instructions I provided, so CUDA_NVCC_EXECUTABLE is not needed.

I re-built to test that it's so.

Thank you!
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18958

Differential Revision: D14810732

Pulled By: ezyang

fbshipit-source-id: 3758ae2253c745c5d7cfccedd49fa00cc4629965

5 years agoFix precision issue with expansion that prefers 'probs' over 'logits' (#18614)
Ahmad Salim Al-Sibahi [Fri, 5 Apr 2019 19:45:37 +0000 (12:45 -0700)]
Fix precision issue with expansion that prefers 'probs' over 'logits' (#18614)

Summary:
I have experienced that sometimes both were in `__dict__`, but it chose to copy `probs` which loses precision over `logits`. This is especially important when training (bayesian) neural networks or doing other type of optimization, since the loss is heavily affected.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18614

Differential Revision: D14793486

Pulled By: ezyang

fbshipit-source-id: d4ff5e34fbb4021ea9de9f58af09a7de00d80a63

5 years agoMethod is supposed to be in-place (#18684)
Joakim Rishaug [Fri, 5 Apr 2019 19:44:49 +0000 (12:44 -0700)]
Method is supposed to be in-place (#18684)

Summary:
Tracing models which attempts to return this in-place value doesn't turn out well.

I haven't run any tests to confirm the results to be honest, but regardless of the outcome, the operation happens in-place, so it should work as before.

Sample output from traced model attempting to set `max_norm` on `Embedding`:
```
a leaf Variable that requires grad has been used in an in-place operation. (check_inplace at /pytorch/torch/csrc/autograd/VariableTypeUtils.h:49)
frame #0: std::function<std::string ()>::operator()() const + 0x11 (0x7f0ecc5cc021 in /usr/local/lib/python3.7/site-packages/torch/lib/libc10.so)
frame #1: c10::Error::Error(c10::SourceLocation, std::string const&) + 0x2a (0x7f0ecc5cb8ea in /usr/local/lib/python3.7/site-packages/torch/lib/libc10.so)
frame #2: <unknown function> + 0x38ab2f (0x7f0ecb55ab2f in /usr/local/lib/python3.7/site-packages/torch/lib/libtorch.so.1)
frame #3: torch::autograd::VariableType::embedding_renorm_(at::Tensor&, at::Tensor const&, double, double) const + 0x76 (0x7f0ecb5b5966 in /usr/local/lib/python3.7/site-packages/torch/lib/libtorch.so.1)
frame #4: <unknown function> + 0x56c958 (0x7f0ecb73c958 in /usr/local/lib/python3.7/site-packages/torch/lib/libtorch.so.1)
frame #5: <unknown function> + 0x672286 (0x7f0ecb842286 in /usr/local/lib/python3.7/site-packages/torch/lib/libtorch.so.1)
frame #6: torch::jit::InterpreterState::run(std::vector<c10::IValue, std::allocator<c10::IValue> >&) + 0x22 (0x7f0ecb83d842 in /usr/local/lib/python3.7/site-packages/torch/lib/libtorch.so.1)
frame #7: <unknown function> + 0x65c6ac (0x7f0ecb82c6ac in /usr/local/lib/python3.7/site-packages/torch/lib/libtorch.so.1)
frame #8: <unknown function> + 0x3c8ab4 (0x7f0f06bc0ab4 in /usr/local/lib/python3.7/site-packages/torch/lib/libtorch_python.so)
frame #9: <unknown function> + 0x3ad2c3 (0x7f0f06ba52c3 in /usr/local/lib/python3.7/site-packages/torch/lib/libtorch_python.so)
frame #10: <unknown function> + 0x11663e (0x7f0f0690e63e in /usr/local/lib/python3.7/site-packages/torch/lib/libtorch_python.so)
<omitting python frames>
frame #39: python_call + 0x11 (0x5563c3c521c1 in uwsgi)
frame #40: uwsgi_request_wsgi + 0x100 (0x5563c3c54410 in uwsgi)
frame #41: wsgi_req_recv + 0xac (0x5563c3becabc in uwsgi)
frame #42: simple_loop_run + 0xc4 (0x5563c3c35be4 in uwsgi)
frame #43: simple_loop + 0x10 (0x5563c3c35a00 in uwsgi)
frame #44: uwsgi_ignition + 0x241 (0x5563c3c3a3a1 in uwsgi)
frame #45: uwsgi_worker_run + 0x275 (0x5563c3c3ec35 in uwsgi)
frame #46: <unknown function> + 0x8f22c (0x5563c3c3f22c in uwsgi)
frame #47: <unknown function> + 0x3c13e (0x5563c3bec13e in uwsgi)
frame #48: __libc_start_main + 0xf1 (0x7f0f138922e1 in /lib/x86_64-linux-gnu/libc.so.6)
frame #49: _start + 0x2a (0x5563c3bec16a in uwsgi)
:
operation failed in interpreter:
op_version_set = 0
def forward(self,
    input_1: Tensor) -> Tensor:
  _0 = torch.norm(self.item_embedding.weight, 2, 1, True)
  _1 = torch.div(self.item_embedding.weight, _0)
  m_weight = torch.t(_1)
  input_2 = torch.contiguous(input_1)
  weight_1 = torch.embedding_renorm_(self.item_embedding.weight, input_2, 1., 2.)
             ~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
  x = torch.embedding(weight_1, input_2, -1, False, False)
  input_3 = torch.div(x, torch.norm(x, 2, 2, True))
  max_batch_size = ops.prim.NumToTensor(torch.size(input_3, 0))
  hx = torch.zeros([2, int(max_batch_size), 70], dtype=6, layout=0, device=torch.device("cpu"))
  _2 = [self.lstm_layer.weight_ih_l0, self.lstm_layer.weight_hh_l0, self.lstm_layer.weight_ih_l1, self.lstm_layer.weight_hh_l1]
  input_4, _3, _4 = torch.lstm(input_3, [hx, hx], _2, False, 2, 0.10000000000000001, False, False, True)
  input = torch.matmul(input_4, torch.t(self.rnn2item.weight))
  tastevec = torch.div(input, torch.norm(input, 2, 2, True))
  outputs = torch.matmul(tastevec, m_weight)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18684

Differential Revision: D14782041

Pulled By: ezyang

fbshipit-source-id: 7b2fc19b7d5b6600263644498bb728319a19f39d

5 years agofix bug when falling back to acc32 when weight is prepacked (#18881)
Summer Deng [Fri, 5 Apr 2019 19:44:09 +0000 (12:44 -0700)]
fix bug when falling back to acc32 when weight is prepacked (#18881)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18881

Pull Request resolved: https://github.com/pytorch/pytorch/pull/18878

When the weight is prepacked and it doesn't contain a prepacked weight for acc32, we shouldn't fallback to acc32.

TODO: add unit tests with better coverage

Reviewed By: feiyu1990

Differential Revision: D14778810

fbshipit-source-id: d49a8c4b7c815ab29b77feb53ee730ad63780488

5 years agoMore numerically stable lerp (#18871)
Marek Kolodziej [Fri, 5 Apr 2019 19:43:02 +0000 (12:43 -0700)]
More numerically stable lerp (#18871)

Summary:
The C++ and CUDA implementations of the lerp are not numerically stable. This is discussed on Wikipedia [here](https://en.wikipedia.org/wiki/Linear_interpolation#Programming_language_support). I checked the GPU SASS output and there's no overhead from using the more precise implementation, from Kepler all the way to Turing. I haven't looked at CPU ASM though.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18871

Differential Revision: D14793438

Pulled By: ezyang

fbshipit-source-id: 2ddc2e026c5285466cae7d1b4101174253100445

5 years agoIncrease default c10d/ProcessGroupGloo test timeout (#18916)
Pieter Noordhuis [Fri, 5 Apr 2019 19:13:31 +0000 (12:13 -0700)]
Increase default c10d/ProcessGroupGloo test timeout (#18916)

Summary:
See #18659.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18916

Differential Revision: D14808749

Pulled By: pietern

fbshipit-source-id: 9a9c8beddb2dbbb1bf4c5e575743d9e1fa3f07fa

5 years agoremove symbolic variable part 1 (#17986)
Ailing Zhang [Fri, 5 Apr 2019 18:57:17 +0000 (11:57 -0700)]
remove symbolic variable part 1 (#17986)

Summary:
As discussed with gchanan we should deduplicate symbolic_variable and symbolic_script to prepare for the future merge with derivatives.yaml.

This PR moves most easy formulas to symbolic_script.

TODO: run benchmarks to make sure no perf regression

cc: apaszke zdevito wanchaol
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17986

Differential Revision: D14766412

Pulled By: ailzhang

fbshipit-source-id: d95a3f876e256c0f505779a71587c985571d3b8f

5 years agoRevert D14742020: Wrap workaround for cpp custom types a bit prettier and add an...
Edward Yang [Fri, 5 Apr 2019 18:55:38 +0000 (11:55 -0700)]
Revert D14742020: Wrap workaround for cpp custom types a bit prettier and add an example

Differential Revision:
D14742020

Original commit changeset: 0f2fd83ae56a

fbshipit-source-id: 5640255aef0319b7d8996e07132e87213130d31c

5 years agoDecompose more Windows scripts (#18917)
Karl Ostmo [Fri, 5 Apr 2019 18:26:31 +0000 (11:26 -0700)]
Decompose more Windows scripts (#18917)

Summary:
This PR:

* pulls four distinct installation steps out of `build_pytorch.bat` and into their own scripts.
* eliminates the copy step for helper scripts called by `win-build.sh` and `win-test.sh`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18917

Differential Revision: D14807236

Pulled By: kostmo

fbshipit-source-id: 03e91a5834dfd6d68903ad9725eacc099bbf6d53

5 years agoWrap workaround for cpp custom types a bit prettier and add an example (#18791)
Dmytro Dzhulgakov [Fri, 5 Apr 2019 18:14:11 +0000 (11:14 -0700)]
Wrap workaround for cpp custom types a bit prettier and add an example (#18791)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18791

As a temporary demonstration on how to extend this hack further until custom C types are ready.

Reviewed By: jamesr66a

Differential Revision: D14742020

fbshipit-source-id: 0f2fd83ae56ab2abe16977a1829ed421e6abe74b

5 years agoRemove cuda::compat functions in aten (#18905)
bddppq [Fri, 5 Apr 2019 18:09:15 +0000 (11:09 -0700)]
Remove cuda::compat functions in aten (#18905)

Summary:
Looks like the issue of using `std::` functions is fixed in new rocm version
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18905

Differential Revision: D14792943

Pulled By: bddppq

fbshipit-source-id: af11acbb85872943f23b6e55415db1f0699e7b8f

5 years agofix side-effects and aliasing for custom ops (#18711)
Michael Suo [Fri, 5 Apr 2019 17:40:19 +0000 (10:40 -0700)]
fix side-effects and aliasing for custom ops (#18711)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18711
ghimport-source-id: c9caedc0660b2b7ba3730cd0e1a2e0e9c3cf422b

Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#18711 [jit] fix side-effects and aliasing for custom ops**

Previously we didn't track aliasing, mutation, or side effects for
custom ops. This PR adds in guards with the most conservative
assumptions possible: the op will
1) have side effects,
2) write to everything
3) produce a wildcard.

In order to tell whether a given operator is a custom op, this PR introduces
the concept of a "reserved" namespace (basically all our builtin namespaces).
Custom ops live in non-reserved namespaces, so a check on the namespace
is sufficient to tell whether a schema/node is "custom" or not.

This is just to get things correct for now. Follow-ups to this:
- Users should be able to specify aliasing/mutability without having to learn
the whole alias annotation schema.
- Relax assumptions a bit. In particular outputs can only alias input tensors,
they don't have to be wildcards.

Fixes #18490

Differential Revision: D14730978

fbshipit-source-id: 540b47a24ccf24145051609bdcc99c97e46e0fe0

5 years agoExpand the list of ops that mutate an inputs shape (#18812)
Elias Ellison [Fri, 5 Apr 2019 17:37:58 +0000 (10:37 -0700)]
Expand the list of ops that mutate an inputs shape (#18812)

Summary:
Expand the list of ops that resize an input in-place to include broadcasting ops and other ops that affect shape. Whoever is reviewing the PR could you please look through pytorch in place ops and see if I missed any.

Expanding the PR from: https://github.com/pytorch/pytorch/pull/17518

This is already being tested in test_resize_input_ops.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18812

Differential Revision: D14793410

Pulled By: eellison

fbshipit-source-id: 125f4f5375ac1036fb96fabc9da2aaccc9adc778

5 years agoadd launch bounds, enable more tests (#18909)
J M Dieterich [Fri, 5 Apr 2019 17:11:43 +0000 (10:11 -0700)]
add launch bounds, enable more tests (#18909)

Summary:
Add launch bounds annotations for ROCm arising from maxThreadsPerBlock and apply threads use.

Enable tests that now work.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18909

Differential Revision: D14801490

Pulled By: ezyang

fbshipit-source-id: b81c97fc783a2627bc7e31b32036a364cfe40cc7

5 years agoAdd backward pass to infer single missing input shape for Concat opportunitiscally...
Yinghai Lu [Fri, 5 Apr 2019 17:09:14 +0000 (10:09 -0700)]
Add backward pass to infer single missing input shape for Concat opportunitiscally (#18911)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18911

Att.

Reviewed By: bddppq

Differential Revision: D14791295

fbshipit-source-id: 4b7a775924f0eadb0cb73aa6c434a6a5be8b92be

5 years agochange to use clang if NDK >= 18 (#18914)
Jiakai Liu [Fri, 5 Apr 2019 16:54:27 +0000 (09:54 -0700)]
change to use clang if NDK >= 18 (#18914)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18914
ghimport-source-id: 4d9d9322ee5559d96e13533ec37ff3be86a0227c

Reviewed By: ezyang

Differential Revision: D14794162

Pulled By: ljk53

fbshipit-source-id: caac55e12b1e62bf6ebcc6e2062d5ed122ad4e64

5 years agoRevert D14673459: [pytorch][PR] [jit] Replace Slot on script::Method with NamedIValue
Zachary DeVito [Fri, 5 Apr 2019 16:46:10 +0000 (09:46 -0700)]
Revert D14673459: [pytorch][PR] [jit] Replace Slot on script::Method with NamedIValue

Differential Revision:
D14673459

Original commit changeset: 21200180c47f

fbshipit-source-id: 9c01de4cf5bb7c87ac0c55705b901db990cd917b

5 years agoDisable flaky test_proper_exit test. (#18950)
Edward Yang [Fri, 5 Apr 2019 16:37:11 +0000 (09:37 -0700)]
Disable flaky test_proper_exit test. (#18950)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18950
ghimport-source-id: 27bd575fd3c73a51ace1360aa020fa63a792a5d2

Differential Revision: D14802009

Pulled By: ezyang

fbshipit-source-id: 051e1d038892c2c6e8337357fa80771b8dc42680

5 years agoCheckout pytorch_sphinx_theme with https. (#18859)
Edward Yang [Fri, 5 Apr 2019 16:33:08 +0000 (09:33 -0700)]
Checkout pytorch_sphinx_theme with https. (#18859)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18859
ghimport-source-id: fbbcb8a2dd9c9f0a317de489b6bbb83e9071a7d8

Differential Revision: D14801989

Pulled By: ezyang

fbshipit-source-id: a9bc02e1383adafcac01994e6346b28551d95c71

5 years agoAdd tests for reducer class (#18845)
Pieter Noordhuis [Fri, 5 Apr 2019 16:04:43 +0000 (09:04 -0700)]
Add tests for reducer class (#18845)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18845

This adds a few CPU only test cases for the reducer class.

Reviewed By: mrshenli

Differential Revision: D14768432

fbshipit-source-id: c008a52206826304e634a95bc14167ed94c97662

5 years agoFix a few instances of notifying on a CV while holding the lock (#18857)
Owen Anderson [Fri, 5 Apr 2019 15:34:41 +0000 (08:34 -0700)]
Fix a few instances of notifying on a CV while holding the lock (#18857)

Summary:
Fix a few instances of notifying on a CV while holding the lock to release the lock before notifying.  This avoids an extra thread suspension when the notified thread tries to grab the lock.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18857

Differential Revision: D14779132

Pulled By: resistor

fbshipit-source-id: b18a05c4c15be1426ebfdffac1c8f002b771cfd7

5 years agoUnify caffe2 and libtorch build scripts on Windows (#18683)
peter [Fri, 5 Apr 2019 14:44:43 +0000 (07:44 -0700)]
Unify caffe2 and libtorch build scripts on Windows (#18683)

Summary:
`scripts/build_windows.bat` is the original way to build caffe2 on Windows, but since it is merged into libtorch, the build scripts should be unified because they actually do the same thing except there are some different flags.

The follow-up is to add the tests. Looks like the CI job for caffe2 windows is defined [here](https://github.com/pytorch/ossci-job-dsl/blob/master/src/jobs/caffe2.groovy#L906). Could we make them a separate file, just like what we've done in `.jenkins/pytorch/win-build.sh`? There's a bunch of things we can do there, like using ninja and sccache to accelerate build.

cc orionr yf225
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18683

Differential Revision: D14730188

Pulled By: ezyang

fbshipit-source-id: ea287d7f213d66c49faac307250c31f9abeb0ebe

5 years agoSimplify storage wrapping in TH. (#18855)
Gregory Chanan [Fri, 5 Apr 2019 14:18:39 +0000 (07:18 -0700)]
Simplify storage wrapping in TH. (#18855)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18855
ghimport-source-id: 01faa229fa4db901ab8539d3778b716d909ba4cf

Reviewed By: dzhulgakov

Differential Revision: D14790669

Pulled By: gchanan

fbshipit-source-id: 167b9bc9c9872743fa8f6040a26ddf7ff5789c27

5 years agoCache device on TensorImpl; clean up TensorImpl constructors. (#18833)
Gregory Chanan [Fri, 5 Apr 2019 14:18:38 +0000 (07:18 -0700)]
Cache device on TensorImpl; clean up TensorImpl constructors. (#18833)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18833
ghimport-source-id: 6f2be25fcc5e6be3ffe20582e604bd2c1fbab66b

Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#18833 [STACK] Cache device on TensorImpl; clean up TensorImpl constructors.**
* #18832 [STACK] Disallow changing the device of a tensor via set_.
* #18831 [STACK] Stop swapping in Storages of the wrong device for Tensors.

1) We cache device on TensorImpl.  This means we can access the device without a virtual function and allows us to more easily extend TensorImpls (because they don't need to figure out how to store the Device for themselves).

2) Clean up TensorImpl APIs.  We had a constructor that took a TensorTypeId and an allocator and would allocate a Storage based on the recognized types of TensorTypeIds.  Instead, we just have two different constructors: one for types with a storage, one without.

Reviewed By: dzhulgakov

Differential Revision: D14766230

fbshipit-source-id: 745b8db84dcd6cb58f1a8675ad3ff8d033bc50df

5 years agoRevert "Adding pin_memory kwarg to zeros, ones, empty,... (#18854)
Vitaly Fedyunin [Fri, 5 Apr 2019 13:19:58 +0000 (06:19 -0700)]
Revert "Adding pin_memory kwarg to zeros, ones, empty,... (#18854)

Summary:
This reverts commit c484cf43a02863efd2f4a76aad43246fb0191ab5.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18854

Differential Revision: D14778393

Pulled By: VitalyFedyunin

fbshipit-source-id: 4b5a1f5b1c091bbc4a8e75614734cc011d26b452

5 years agoSilence compiler warnings (#18912)
Sebastian Messmer [Fri, 5 Apr 2019 08:46:58 +0000 (01:46 -0700)]
Silence compiler warnings (#18912)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18912

We intentionally test a deprecated API, no need to show the warnings here.

Reviewed By: dzhulgakov

Differential Revision: D14792617

fbshipit-source-id: 9ea2a4106d566064283726eed2c274b98f49a2e5

5 years agoScriptModuleOp in caffe2 (#18716)
Dmytro Dzhulgakov [Fri, 5 Apr 2019 08:04:58 +0000 (01:04 -0700)]
ScriptModuleOp in caffe2 (#18716)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18716

Might be useful as an intermediate stage for some systems that currently use Caffe2 nets as an execution mechanism.

Not sure it's a good idea all together, please comment.

Limitations:
- only Tensor types as inputs/outputs
- the entire module is serialized as a zip archive inside a proto in Caffe2 db, it'd be subject to 4Gb limit and is likely very slow. For small models it'd work though.
- no autograd, though it can be attached in principle
- no way to retrieve parameters inside the script module from C2 runtime perspective (though they potentially can be alias-fetched and stored as individual blobs)
- after deserialization, python wrappers returned don't have correct type (as we don't do module_lookup trick)

Build-wise, I had to add dependency from pybind_state to libtorch.so. I don't think we build Caffe2 python frontend independently anymore, so it should be fine.

Reviewed By: amirshim, houseroad

Differential Revision: D14339599

fbshipit-source-id: 88a37a8abd1f1c4703e5ef937031f222535d4080

5 years agoflake8 fix on extracted python script
Karl Ostmo [Fri, 5 Apr 2019 07:49:06 +0000 (00:49 -0700)]
flake8 fix on extracted python script

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/18931

Differential Revision: D14796114

Pulled By: kostmo

fbshipit-source-id: 25971be5a36fffc61e29db981af7298a0fe0ed8c

5 years agoReplace Slot on script::Method with NamedIValue (#18252)
David Riazati [Fri, 5 Apr 2019 06:27:05 +0000 (23:27 -0700)]
Replace Slot on script::Method with NamedIValue (#18252)

Summary:
This refactor lets us track the types of initial values added onto a `Method`. The main motivation for this is the change in `module.cpp`.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18252

Differential Revision: D14673459

Pulled By: driazati

fbshipit-source-id: 21200180c47f25bb70898771adfb569856e6c34a

5 years agoU/kostmo/windows offload scripts 3
Karl Ostmo [Fri, 5 Apr 2019 04:05:13 +0000 (21:05 -0700)]
U/kostmo/windows offload scripts 3

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/18754

Differential Revision: D14794893

Pulled By: kostmo

fbshipit-source-id: 05187d9b53615ffbcc7253accdc692c4ecaf25d9

5 years agofix lint in optim doc
Tongzhou Wang [Fri, 5 Apr 2019 02:03:08 +0000 (19:03 -0700)]
fix lint in optim doc

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/18883

Differential Revision: D14793365

Pulled By: ezyang

fbshipit-source-id: c1b46c98e3319badec3e0e772d0ddea24cbf9c89

5 years agoFixed the comment to reference gist example instead of private repo (#18852)
Iurii Zdebskyi [Fri, 5 Apr 2019 01:23:38 +0000 (18:23 -0700)]
Fixed the comment to reference gist example instead of private repo (#18852)

Summary:
Replace link to a file in a private repo with a gist
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18852

Reviewed By: ezyang

Differential Revision: D14778481

Pulled By: izdeby

fbshipit-source-id: 8389aa4bf115ddcfd85079cc2c861404efa678e7

5 years agoreturn missing keys from load_state_dict (#18668)
Sepehr Sameni [Fri, 5 Apr 2019 01:07:54 +0000 (18:07 -0700)]
return missing keys from load_state_dict (#18668)

Summary:
return missing_keys and unexpected_keys from load_state_dict so the user can handle them when strict mode is off; also removed an unused variable
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18668

Differential Revision: D14782073

Pulled By: ezyang

fbshipit-source-id: ab3b855eb77bb7422594d971988067e86eef20f2

5 years agoFix caffe2 miopen conv transpose gradient op for case of no dX gradient
Junjie Bai [Fri, 5 Apr 2019 00:21:41 +0000 (17:21 -0700)]
Fix caffe2 miopen conv transpose gradient op for case of no dX gradient

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/18809

Reviewed By: ezyang

Differential Revision: D14759762

Pulled By: bddppq

fbshipit-source-id: ff795b7e58c82f67a1d7284b5ab06b0e0e5fd3ae

5 years agodon't attempt to multiply by a sparse matrix (#18737)
Brennan Vincent [Fri, 5 Apr 2019 00:18:11 +0000 (17:18 -0700)]
don't attempt to multiply by a sparse matrix (#18737)

Summary:
Tested by running the script in #16562 , and there was no error.

Then:
```
>>> print(mat.grad)
tensor([[1., 2., 3.],
        [1., 2., 3.],
        [1., 2., 3.]])
```

which is correct.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18737

Differential Revision: D14773078

Pulled By: umanwizard

fbshipit-source-id: 8aa36eb6f6aa104263a467d9ac91d61b3bfd05f5

5 years agoadd Fast-RNN to AI-PEP
Wanchao Liang [Fri, 5 Apr 2019 00:00:46 +0000 (17:00 -0700)]
add Fast-RNN to AI-PEP

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/18885

Reviewed By: hl475

Differential Revision: D14728854

fbshipit-source-id: 7e7a2946929551963f7c938e3d82a260a9efdfbd

5 years agoAllow override of backend in dist.new_group() (#18595)
Pieter Noordhuis [Thu, 4 Apr 2019 21:14:50 +0000 (14:14 -0700)]
Allow override of backend in dist.new_group() (#18595)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18595

There is no need to force the backend to be the same as the global
process group, as long as the backend is "nccl" or "gloo".

Reviewed By: mrshenli

Differential Revision: D14657204

fbshipit-source-id: 868817b9f219e3be8db0761a487f0027ed46663b

5 years agoONNX Export All Cases of Softmax
Lara [Thu, 4 Apr 2019 20:15:18 +0000 (13:15 -0700)]
ONNX Export All Cases of Softmax

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/18482

Reviewed By: zrphercule

Differential Revision: D14630697

Pulled By: houseroad

fbshipit-source-id: c06f1e3bead10a265c5f4ac3723d49f4caf46801

5 years agoAdded bool and half support for resize_as_ and view methods (#18821)
Iurii Zdebskyi [Thu, 4 Apr 2019 20:01:10 +0000 (13:01 -0700)]
Added bool and half support for resize_as_ and view methods (#18821)

Summary:
Enabled **resize_as_** and **view** methods for bool and half tensors.
tested via unit tests
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18821

Reviewed By: ezyang

Differential Revision: D14762852

Pulled By: izdeby

fbshipit-source-id: 4312079fb4e893fea6f71ff4f163094b2674f1e8

5 years agoAutomatic update of fbcode/onnx to 079c2639f9bb79b1774d1e3bfa05b0c093816ca7 (#18841)
Lu Fang [Thu, 4 Apr 2019 19:57:31 +0000 (12:57 -0700)]
update of fbcode/onnx to 079c2639f9bb79b1774d1e3bfa05b0c093816ca7 (#18841)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18841

Previous import was f0d7df2c643c4e37f1fd7735ef02c972c4d19fb5

Included changes:
- **[079c2639](https://github.com/onnx/onnx/commit/079c2639)**: update the squeeze and unsqueeze doc (#1905) <Lu Fang>
- **[a8b45d62](https://github.com/onnx/onnx/commit/a8b45d62)**: fix the ir_version onnx-operators.proto (#1903) <Lu Fang>

Reviewed By: zrphercule

Differential Revision: D14767158

fbshipit-source-id: 2d772fece45e25d72bf1d10fad156189397f3f86

5 years agoActually model scalar type promotion in shape analysis (#18811)
James Reed [Thu, 4 Apr 2019 19:53:44 +0000 (12:53 -0700)]
Actually model scalar type promotion in shape analysis (#18811)

Summary:
This was causing some numerical issues in the fuser
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18811

Differential Revision: D14767390

Pulled By: jamesr66a

fbshipit-source-id: f1123d1aab5501abad850d2edc996f8aa8dafe04

5 years agoAdd a .ctags.d/ toplevel directory (#18827)
Max Wang [Thu, 4 Apr 2019 19:42:12 +0000 (12:42 -0700)]
Add a .ctags.d/ toplevel directory (#18827)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18827
ghimport-source-id: 38f857bc29b2c2c6a71069d00c4c69ed0bef1574

Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#18827 Add a .ctags.d/ toplevel directory**

Exclude build artifacts by default.

Reviewed By: ezyang

Differential Revision: D14765721

fbshipit-source-id: a785dbb2ef1df96af8e23cc65c8db2a6b67b4fce

5 years agoFix typo
Wanwannodao [Thu, 4 Apr 2019 19:40:46 +0000 (12:40 -0700)]
Fix typo

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/18802

Differential Revision: D14781874

Pulled By: ezyang

fbshipit-source-id: 0f94c40bd84c84558ea3329117580f6c749c019f

5 years agoAdd support for group ConvTranspose (#18794)
Xiaomeng Yang [Thu, 4 Apr 2019 18:46:37 +0000 (11:46 -0700)]
Add support for group ConvTranspose (#18794)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18794

Add support for group ConvTranspose

Reviewed By: houseroad

Differential Revision: D14741327

fbshipit-source-id: 5d947ca044bf8495dd7f8f56122441ebbcc6c7e4

5 years agoDisallow changing the device of a tensor via set_. (#18832)
Gregory Chanan [Thu, 4 Apr 2019 18:12:13 +0000 (11:12 -0700)]
Disallow changing the device of a tensor via set_. (#18832)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18832
ghimport-source-id: fde4ad90541ba52dfa02bdd83466f17e6541e535

Stack from [ghstack](https://github.com/ezyang/ghstack):
* #18833 [STACK] Cache device on TensorImpl; clean up TensorImpl constructors.
* **#18832 [STACK] Disallow changing the device of a tensor via set_.**
* #18831 [STACK] Stop swapping in Storages of the wrong device for Tensors.

This is necessary to cache the device on a TensorImpl.

Differential Revision: D14766231

fbshipit-source-id: bba61634b2d6252ac0697b96033c9eea680956e8

5 years agoU/kostmo/win test offload scripts
Karl Ostmo [Thu, 4 Apr 2019 17:38:09 +0000 (10:38 -0700)]
U/kostmo/win test offload scripts

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/18694

Differential Revision: D14766339

Pulled By: kostmo

fbshipit-source-id: a2300e72129979f866430ca5c09dd7fff6df0a89

5 years agoRevert D14603722: Enforce single parent for script submodules
Zachary DeVito [Thu, 4 Apr 2019 17:22:27 +0000 (10:22 -0700)]
Revert D14603722: Enforce single parent for script submodules

Differential Revision:
D14603722

Original commit changeset: 63ab5d0cccf7

fbshipit-source-id: 2c4174def102eda4589e08c4dbd67ce8af975199

5 years agoFix deviceCount on FakeGuardImpl. (#18745)
Edward Yang [Thu, 4 Apr 2019 16:20:20 +0000 (09:20 -0700)]
Fix deviceCount on FakeGuardImpl. (#18745)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18745
ghimport-source-id: 3ed111efe83b3061652869e33d9b5910b7daa732

Differential Revision: D14759198

Pulled By: ezyang

fbshipit-source-id: 70a8db767f310fe0e0079c7b0693e9330d7cd472

5 years agoStop swapping in Storages of the wrong device for Tensors. (#18831)
Gregory Chanan [Thu, 4 Apr 2019 13:19:54 +0000 (06:19 -0700)]
Stop swapping in Storages of the wrong device for Tensors. (#18831)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18831
ghimport-source-id: 2741e0d70ebe2c2217572c3af54ddd9d2047e342

Stack from [ghstack](https://github.com/ezyang/ghstack):
* #18833 [STACK] Cache device on TensorImpl; clean up TensorImpl constructors.
* #18832 [STACK] Disallow changing the device of a tensor via set_.
* **#18831 [STACK] Stop swapping in Storages of the wrong device for Tensors.**

This is necessary to support device caching, see https://github.com/pytorch/pytorch/pull/18751 and https://github.com/pytorch/pytorch/pull/18578.

In library code, we potentially swap in Storages with the wrong device when device_guard is False.  This happens as follows with "view-like" operations.
1) We allocate a tensor on the 'wrong' device (because device_guard is false).
2) We swap out the 'wrong' storage with the 'right' storage using e.g. THCTensor_setStorage.

Instead, we can just construct the Tensor with the correct Storage from the beginning.  This is what we do with 'view'.

Note there are two other "view-like" cases where this happens:
1) unfold
2) set_()

Because these aren't performance critical, I just added the device_guard instead of applying the above correction.

For completeness, this also includes a test that all `device_guard: false` functions behave properly under these conditions.

Reviewed By: dzhulgakov

Differential Revision: D14766232

fbshipit-source-id: 0865c3ddae3f415df5da7a9869b1ea9f210e81bc

5 years agoPass ScalarType separately from Type in python constructors
Roy Li [Thu, 4 Apr 2019 09:21:09 +0000 (02:21 -0700)]
Pass ScalarType separately from Type in python constructors

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17786

Reviewed By: ezyang

Differential Revision: D14379075

fbshipit-source-id: 3abf066563b789a30cafe5b0c868a41326f5b833

5 years agoStore ScalarType and Backend instead of Type in TensorIterator
Roy Li [Thu, 4 Apr 2019 09:21:09 +0000 (02:21 -0700)]
Store ScalarType and Backend instead of Type in TensorIterator

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/17601

Reviewed By: ezyang

Differential Revision: D14274754

fbshipit-source-id: b08880ae586b6ae57d4c0bbeb203796d087926c4

5 years agoIntroduce DeprecatedTypeProperties class (#17991)
Roy Li [Thu, 4 Apr 2019 09:21:09 +0000 (02:21 -0700)]
Introduce DeprecatedTypeProperties class (#17991)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17991

changes:
-Breaks bc: Tensor::type() now returns DeprecatedTypeProperties& rather than Type&.
-Added DeprecatedTypeProperties, it serves as a temporary replacement for Type as the return value of Tensor::type(). This contributes to making Type just for dispatch purposes so that we can make it dtype agnostic.
-Tensor::dispatch_type() now returns Type& like Tensor::type() used to do.
-Changed callsites of Tensor::type() appropriately.

Reviewed By: ezyang

Differential Revision: D14443117

fbshipit-source-id: 239ccb7a09626279a71d1a37f8f82e7f57bf7d9e

5 years agoFix to handle null strides in DLPack tensor (#18510)
Bram Wasti [Thu, 4 Apr 2019 07:24:16 +0000 (00:24 -0700)]
Fix to handle null strides in DLPack tensor (#18510)

Summary:
DLPack can have non-strided tensors, which is represented by a nullptr in the place of dl_tensor.strides.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18510

Differential Revision: D14647328

Pulled By: bwasti

fbshipit-source-id: 5364282810a5772cfc2319fc8133fe86fdd84dd1

5 years agoAdd shape inference function for Split (#18838)
Yinghai Lu [Thu, 4 Apr 2019 07:19:21 +0000 (00:19 -0700)]
Add shape inference function for Split (#18838)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18838

It turns out that we don't have shape inference function of `Split` op at all. This diff adds that.

Reviewed By: bertmaher

Differential Revision: D14766871

fbshipit-source-id: 535cb4f24bdada603c76579e00e7a39aee93e19f

5 years agoFix the duplication problem in _unique_state_dict (#18139)
Lu Fang [Thu, 4 Apr 2019 06:14:07 +0000 (23:14 -0700)]
Fix the duplication problem in _unique_state_dict (#18139)

Summary:
Since parameter.data will create a new torch.Tensor each time, we get duplicate tensors when call _unique_state_dict now. Try to deduplicate it before creating new tensor.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18139

Reviewed By: dzhulgakov

Differential Revision: D14511262

Pulled By: houseroad

fbshipit-source-id: cb69795d0b6509721220650bbb19edeb3459a503

5 years agofold col offset into bias; optimize A symmetric quant (#17026)
Jongsoo Park [Thu, 4 Apr 2019 05:50:05 +0000 (22:50 -0700)]
fold col offset into bias; optimize A symmetric quant (#17026)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17026

D14013931 was for FC. This diff is similar optimizations for Conv.
A subtle difference is that in FC, once we fold col_offset into bias during pre-processing step, we can treat everything as if A_zero_offset == 0 (symmetric quantization of A).
In Conv, we can't do this because padding still needs to use the original A_zero_offset.
From requantization point of view, once col_offset folded into bias, we can treat as if we're doing symmetric A quantization.
But, for steps involving padding like im2col, im2col fused with packing, and direct conv for depth-wise/group convolution we still need to pass the original A_zero_offset.

Reviewed By: jianyuh

Differential Revision: D14020276

fbshipit-source-id: c29caefd1127bbc6aff0e9d535939bb0c1ecb66c

5 years agofix flake8 lint (#18835)
Michael Suo [Thu, 4 Apr 2019 05:18:09 +0000 (22:18 -0700)]
fix flake8 lint (#18835)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18835
ghimport-source-id: 7b1f433ae51232822704d62699233688072cbc23

Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#18835 fix flake8 lint**
* #18826 [jit] run cpp tests for non-cuda builds in test_jit.py

...again

Reviewed By: ZolotukhinM

Differential Revision: D14766790

fbshipit-source-id: 29361a407589092831dfbc3c5d63d2834934cd02

5 years agorun cpp tests for non-cuda builds in test_jit.py (#18826)
Michael Suo [Thu, 4 Apr 2019 05:18:09 +0000 (22:18 -0700)]
run cpp tests for non-cuda builds in test_jit.py (#18826)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18826
ghimport-source-id: 7ffa3bc7ef7402a6d6eb6ba5849e197019d77bf8

Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#18826 [jit] run cpp tests for non-cuda builds in test_jit.py**

We did all the work of nicely separating our cpp tests that don't require
CUDA, but they aren't run from test_jit.py if CUDA is missing.

Reviewed By: ZolotukhinM

Differential Revision: D14766287

fbshipit-source-id: 9326b3a5c90f6c20fc8cfaf1a1885a363b91f30a

5 years agoFix the linter (#18842)
Lu Fang [Thu, 4 Apr 2019 04:29:36 +0000 (21:29 -0700)]
Fix the linter (#18842)

Summary:
Remove extra empty line
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18842

Differential Revision: D14767334

Pulled By: houseroad

fbshipit-source-id: 63224bc407949949e1eb5123d3f151e4ac8f6988

5 years agoEnforce single parent for script submodules (#18379)
Zachary DeVito [Thu, 4 Apr 2019 03:21:27 +0000 (20:21 -0700)]
Enforce single parent for script submodules (#18379)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18379
ghimport-source-id: 9895ecc1ff7897e98853dc00675341f36726e7c7

Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#18379 Enforce single parent for script submodules**
* #18378 Unify namespace of script::Module
* #18314 Add ability to specialize class types to ArgumentSpec
* #18226 Add Slot type to abstract the raw pointers being used for slots.

The assumption that a ScriptModule has a single parent is present in
our serialization format, and likely a few other places. It is not
enforced on creation of script module hierarchies though, meaning that
problems associated with (e.g. replicating a module twice in the output
format) will not be caught until much later in the development cycle.

This patch enforces the property when a submodule is registered.
It also removes NamedModule since it is no longer necessary in this regime.
This will also allow the easy discover of a modules fully-qualified name
without needing to traverse the Module hierarchy.

Differential Revision: D14603722

fbshipit-source-id: 63ab5d0cccf7d66c7833e0adf9023024ca9607cb

5 years agoAllow ints, floats, and tensors in conditional (#18755)
Elias Ellison [Thu, 4 Apr 2019 00:09:37 +0000 (17:09 -0700)]
Allow ints, floats, and tensors in conditional (#18755)

Summary:
Per our offline discussion, allow Tensors, ints, and floats to be casted to be bool when used in a conditional

Fix for https://github.com/pytorch/pytorch/issues/18381
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18755

Reviewed By: driazati

Differential Revision: D14752476

Pulled By: eellison

fbshipit-source-id: 149960c92afcf7e4cc4997bccc57f4e911118ff1

5 years agoFix layernorm ad formula on weight and bias (#18233)
Wanchao Liang [Wed, 3 Apr 2019 23:50:46 +0000 (16:50 -0700)]
Fix layernorm ad formula on weight and bias (#18233)

Summary:
Fix the layernorm formula when weight and bias passed in.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18233

Differential Revision: D14760375

Pulled By: wanchaol

fbshipit-source-id: d6bd3b137bc04c391aa5c24d021d1f811ba2a877

5 years agoUnify namespace of script::Module (#18378)
Zachary DeVito [Wed, 3 Apr 2019 22:58:08 +0000 (15:58 -0700)]
Unify namespace of script::Module (#18378)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18378
ghimport-source-id: 55c29bb436a2153d29ff2f4488d99d8863c187b1

Stack from [ghstack](https://github.com/ezyang/ghstack):
* #18379 Enforce single parent for script submodules
* **#18378 Unify namespace of script::Module**
* #18314 Add ability to specialize class types to ArgumentSpec
* #18226 Add Slot type to abstract the raw pointers being used for slots.

This removes individual OrderedDicts in favor of a single unified
namespace for all things in a script::Module. This removes a whole
class of bugs where both a method and an parameter could get the
same name, for instance.

Since we no longer have to expose OrderedDict::Item objects, a lot of
downstream code can be simplified.

We no longer now double-store names (both in the key of the dictionary,
and in the object itself).

Differential Revision: D14603723

fbshipit-source-id: b5f7551b3074679623edd6ea70269830353b4d4c

5 years agoStep 1: Secretly add return_counts to unique, and refactor unique_dim for performance...
Vitaly Fedyunin [Wed, 3 Apr 2019 22:26:34 +0000 (15:26 -0700)]
Step 1: Secretly add return_counts to unique, and refactor unique_dim for performance (#18648)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18648
ghimport-source-id: 1cf4a8fe91492621e02217f38cae5d7e0699fb05

Stack from [ghstack](https://github.com/ezyang/ghstack):
* #18661 Step 7: remove _unique
* #18655 Step 6: Rename _unique2 to unique and add int? dim
* #18654 Step 5: remove _unque_dim in favor of unique_dim
* #18651 Step 4: add support for unique with dim=None
* #18650 Step 3: Add support for return_counts to torch.unique for dim not None
* #18649 Step 2: Rename _unique_dim2_temporary_will_remove_soon to unique_dim
* **#18648 Step 1: Secretly add return_counts to unique, and refactor unique_dim for performance**

`unique` is fragile, previously I tried to change it in #18391 and #17097, they all pass OSS tests but finally get reverted due to internal failure. My previous work of refactoring unique #18459 is based on #18391, and after #18391 get reverted, I could not work on #18459. To continue working on #18459, #18391, and #17097 without worrying about internal failures, I am suggesting the following steps for the improvements of `unique` and `unique_dim`. soumith Please take this and there is no need to put #18391 back.

The motivation is basically to move forward as much as possible without causing any internal failures. So I will try to divide it into steps and sort from low probability of internal failure to high probability. (I don't know what the internal failure is, so I have to guess). Let's merge these PR stack one by one until we enounter internal failure.

Step 1: Create two new ATen operators, `_unique2_temporary_will_remove_soon` and `_unique_dim2_temporary_will_remove_soon` and keep `_unique` and `_unique_dim` unchanged. The backend of these two functions and `_unique` and `_unique_dim` are all the same, the only difference is the temporary ones support `return_counts` but not the `_unique` and `_unique_dim`. Step one is mostly #18391 + #18459. The cuda8 errors has been fixed. At this point, there is no user visible API change, so no docs are updated. `torch.unique` does not support `return_counts` yet, and `return_counts` is tested through the newly added temporary operators. This step just added two new ATen operators, so there shouldn't be any internal failure.

Step 2: Rename `_unique_dim2_temporary_will_remove_soon` to `unique_dim`. This should cause no internal failure either, because no change to existing operators. The only thing to worry about is to delete `unique_dim` from python side because we don't want users to use it. At this point, C++ users now have `return_counts` support for `unique_dim`.

Step 3: Update the docs of `torch.unique` and use `unique_dim` inside `torch.unique` to support `return_counts` In the docs, we should say `torch.unique` with None dim support does not support `return_counts` yet. This might cause internal failure.

Step 4: Rename `_unique2_temporary_will_remove_soon` to `_unique2` and use `_unique2` inside `torch.unique` to support `return_counts`. Update the docs saying that `torch.unique` with None dim now support `return_counts`. This might cause internal failure.

Step 5: Remove `_unique_dim`. This might cause internal failure.

Step 6: Rename `_unique2` to `unique`, add optional `dim` argument to make it looks like the signature of Python's `torch.unique`. Inside `torch.unique`, use `unique` and get rid of `unique_dim`. Unbind `unique_dim` totally from Python at codegen. This is likely to cause internal fail.

Step 7: Remove `_unique`. This is very likely to cause internal failure.

This PR
======

This PR is for step 1. This create two new ATen operators, `_unique2_temporary_will_remove_soon` and `_unique_dim2_temporary_will_remove_soon` and implement `return_counts` inside them and do refactor for performance improvements.

Please review ngimel VitalyFedyunin. They are mostly copied from #18391 and #18459, so the review should be easy.

Below is a benchmark on a tensor of shape `torch.Size([15320, 2])`:

Before
---------

```python
print(torch.__version__)
%timeit a.unique(dim=0, sorted=True, return_inverse=False); torch.cuda.synchronize()
%timeit a.unique(dim=0, sorted=True, return_inverse=True); torch.cuda.synchronize()
```

```
1.0.1
192 µs ± 1.61 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
548 ms ± 3.39 ms per loop (mean ± std. dev. of 7 runs, 1 loop each)
```

```python
print(torch.__version__)
%timeit a.unique(sorted=True, return_inverse=False); torch.cuda.synchronize()
%timeit a.unique(sorted=True, return_inverse=True); torch.cuda.synchronize()
```

```
1.0.1
226 µs ± 929 ns per loop (mean ± std. dev. of 7 runs, 1000 loops each)
302 µs ± 7.06 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
```

After
-------

```python
print(torch.__version__)
%timeit a.unique(dim=0, sorted=True, return_inverse=False); torch.cuda.synchronize()
%timeit a.unique(dim=0, sorted=True, return_inverse=True); torch.cuda.synchronize()
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted=True, return_inverse=False, return_counts=True); torch.cuda.synchronize()
%timeit torch._unique_dim2_temporary_will_remove_soon(a, dim=0, sorted=True, return_inverse=True, return_counts=True); torch.cuda.synchronize()
```

```
1.1.0a0+83ab8ac
190 µs ± 2.14 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
237 µs ± 1.23 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
219 µs ± 2.3 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
263 µs ± 1.15 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
```

```python
print(torch.__version__)
%timeit a.unique(sorted=True, return_inverse=False); torch.cuda.synchronize()
%timeit a.unique(sorted=True, return_inverse=True); torch.cuda.synchronize()
%timeit torch._unique2_temporary_will_remove_soon(a, sorted=True, return_inverse=False, return_counts=True); torch.cuda.synchronize()
%timeit torch._unique2_temporary_will_remove_soon(a, sorted=True, return_inverse=True, return_counts=True); torch.cuda.synchronize()
```

```
1.1.0a0+83ab8ac
232 µs ± 2.21 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
301 µs ± 1.65 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
264 µs ± 7.67 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
339 µs ± 9.2 µs per loop (mean ± std. dev. of 7 runs, 1000 loops each)
```

Differential Revision: D14730905

fbshipit-source-id: 10026b4b98628a8565cc28a13317d29adf1225cc

5 years agoSupport replicating multi-GPU modules (#18687)
Shen Li [Wed, 3 Apr 2019 21:37:54 +0000 (14:37 -0700)]
Support replicating multi-GPU modules (#18687)

Summary:
If the input `network` resides on multiple GPUs, `devices` must be a 2D list with `devices[0]` matching `network`'s devices. See  #18591
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18687

Differential Revision: D14706162

Pulled By: mrshenli

fbshipit-source-id: dca630d3308f2dbcf8b75629c452d7a64092ba42

5 years agoflake8 fix
Wanchao Liang [Wed, 3 Apr 2019 21:07:31 +0000 (14:07 -0700)]
flake8 fix

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/18810

Differential Revision: D14758293

Pulled By: wanchaol

fbshipit-source-id: 975abe4fc5dc0dc4d43af61ec0f987e2c5670874

5 years agoRemove `device_guard: False` from native_functions that don't have a … (#18803)
Gregory Chanan [Wed, 3 Apr 2019 20:55:03 +0000 (13:55 -0700)]
Remove `device_guard: False` from native_functions that don't have a … (#18803)

Summary:
…tensor.

There's nothing to device_guard on.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18803

Reviewed By: ezyang

Differential Revision: D14748091

Pulled By: gchanan

fbshipit-source-id: ed6f16d6f4d3f07b6d5ad9696f71a14333c228b8

5 years agoSwitch our Linux machine AMI to a newer image. (#18433)
Edward Yang [Wed, 3 Apr 2019 20:38:56 +0000 (13:38 -0700)]
Switch our Linux machine AMI to a newer image. (#18433)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18433
ghimport-source-id: 1c92f98b091232c0045a2e1db75d19c1f258ac1f

Differential Revision: D14748827

Pulled By: ezyang

fbshipit-source-id: a459451058cf5560811403bafb96c6ff083d7e3a

5 years agoQTensor (#18230)
Jerry Zhang [Wed, 3 Apr 2019 20:13:26 +0000 (13:13 -0700)]
QTensor (#18230)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18230

Implementing minimum qtensor API to unblock other workstreams in quantization

Changes:
- Added Quantizer which represents different quantization schemes
- Added qint8 as a data type for QTensor
- Added a new ScalarType QInt8
- Added QTensorImpl for QTensor
- Added following user facing APIs
  - quantize_linear(scale, zero_point)
  - dequantize()
  - q_scale()
  - q_zero_point()

Reviewed By: dzhulgakov

Differential Revision: D14524641

fbshipit-source-id: c1c0ae0978fb500d47cdb23fb15b747773429e6c

5 years agoEnforce import order to make protobuf cpp implementation in python work (#18560)
Dmytro Dzhulgakov [Wed, 3 Apr 2019 20:12:28 +0000 (13:12 -0700)]
Enforce import order to make protobuf cpp implementation in python work (#18560)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18560

We have to import python protobuf here **before** we load cpp extension.
Otherwise it breaks under certain build conditions if cpp implementation of
protobuf is used. Presumably there's some registry in protobuf library and
python side has to initialize the dictionary first, before static
initialization in python extension does so. Otherwise, duplicated protobuf
descriptors will be created and it can lead to obscure errors like

  Parameter to MergeFrom() must be instance of same class: expected caffe2.NetDef got caffe2.NetDef.

I think it also fixes https://github.com/facebookarchive/caffe2/issues/1573

Reviewed By: ezyang, iroot900

Differential Revision: D14622054

fbshipit-source-id: 2499eb88ecdee85ff8d845859048f7ae5da2a480

5 years agoPin onnx ir_version to 4 (#18768)
Lu Fang [Wed, 3 Apr 2019 20:08:21 +0000 (13:08 -0700)]
Pin onnx ir_version to 4 (#18768)

Summary:
to make test_operators.py more stable. in future, we will bump this up manually, and I think it's acceptable, since ir_version should be bumped too often.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18768

Reviewed By: zrphercule

Differential Revision: D14741514

Pulled By: houseroad

fbshipit-source-id: 0369dbc55424e345a113e49fc104a441ea290d58

5 years agofix nccl compilation to make sure it compiles for architectures that pytorch compiles...
Soumith Chintala [Wed, 3 Apr 2019 19:44:35 +0000 (12:44 -0700)]
fix nccl compilation to make sure it compiles for architectures that pytorch compiles for (#18739)

Summary:
resubmit of https://github.com/pytorch/pytorch/pull/18704 with additional fixes

Fixes https://github.com/pytorch/pytorch/issues/18359
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18739

Differential Revision: D14737274

Pulled By: soumith

fbshipit-source-id: cfbbbf68b098594bd045861d1b2c085da693ea51

5 years agopush magma init into lazyInitCUDA (#18527)
Soumith Chintala [Wed, 3 Apr 2019 19:27:19 +0000 (12:27 -0700)]
push magma init into lazyInitCUDA (#18527)

Summary:
Tries to fix C++ API's usage of MAGMA-based functions.

Attempts to Fix https://github.com/pytorch/pytorch/issues/18074
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18527

Differential Revision: D14691694

Pulled By: soumith

fbshipit-source-id: dd04e74418e486d73ea4a92193ddf79352ed71ba

5 years agoFor some files that are touched by the QTensor diff (#18765)
Jerry Zhang [Wed, 3 Apr 2019 19:01:57 +0000 (12:01 -0700)]
For some files that are touched by the QTensor diff (#18765)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18765

att

Reviewed By: ZolotukhinM

Differential Revision: D14733442

fbshipit-source-id: 525002034e6dccc2045da645e1193671fd0474b3

5 years agoFix contiguous AD and Autogradzero inconsistency (#18633)
Wanchao Liang [Wed, 3 Apr 2019 18:18:05 +0000 (11:18 -0700)]
Fix contiguous AD and Autogradzero inconsistency (#18633)

Summary:
Fixes #17962
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18633

Differential Revision: D14700449

Pulled By: wanchaol

fbshipit-source-id: 3d15d67c01b69b28394a0f2f001db90ed9fd31dc

5 years agoAdded indexing for bool tensors and bool Indices (#18583)
Iurii Zdebskyi [Wed, 3 Apr 2019 17:53:11 +0000 (10:53 -0700)]
Added indexing for bool tensors and bool Indices (#18583)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18583
ghimport-source-id: 2b1941449827f4ab632fa0f5c8cf0791a6be0845

Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#18583 Added indexing for bool tensors and bool Indices**
* #18505 Added numpy conversion
* #18166 Bool Tensor for CUDA

-----------
This PR enables bool tensor indexing and indexing with bool indices. This is a part of Bool Tensor feature implementation work. The whole plan looks like this:
1. Storage Implementation [Done]
2. Tensor Creation.
    a) CPU [Done]
    b) CUDA [In review]
3. Tensor Conversions. [In review]
4. Tensor Indexing. [This PR]
5. Tensor Operations.
6. Back compatibility related changes.

TODO:
as a follow up, we should move nonzero method from TH to Aten to make code cleaner.

Change:
```
v = torch.tensor([True, False, True], dtype=torch.bool)
boolIndices = torch.tensor([True, False, False], dtype=torch.bool)
v[boolIndices]
-> tensor([True], dtype=torch.bool)

v = torch.randn(5, 7, 3)
boolIndices = torch.tensor([True, False, True, True, False], dtype=torch.bool)
v[boolIndices]
->
tensor([[[ 0.5885, -0.3322,  0.7388],
         [ 1.1182,  0.7808, -1.1492],
         [-0.7952,  0.5255, -0.0251],
         [ 0.7128,  0.8099,  1.2689],
         [-0.7018, -1.4733, -0.3732],
         [ 0.4503,  0.4986, -1.1605],
         [ 0.3348, -1.3767, -0.2976]],

        [[-2.0303, -0.4720, -0.1448],
         [-0.1914, -0.6821,  2.0061],
         [-1.0420, -0.1872, -0.3438],
         [ 1.7587, -0.4183, -0.7577],
         [ 1.0094, -0.1950, -0.2430],
         [ 0.1174,  0.3308, -0.5700],
         [ 0.1110, -0.2714,  1.3006]],

        [[-0.1946, -1.4747, -0.4650],
         [-1.0567,  1.0110, -0.2809],
         [ 0.3729, -0.5699,  0.0815],
         [-0.7733, -0.8316,  0.1674],
         [ 1.2000, -0.3745, -1.1679],
         [ 1.7105,  0.9851, -0.1907],
         [-1.1077,  0.2086, -0.0548]]])
```

Differential Revision: D14673403

fbshipit-source-id: 2b88ec2c7eb26a4f5ef64f8707fb68068d476fc9

5 years agoadd an assertion to check the param num (#18145)
Lu Fang [Wed, 3 Apr 2019 17:51:41 +0000 (10:51 -0700)]
add an assertion to check the param num (#18145)

Summary:
Introduce this check to see whether it will break any existing workflow
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18145

Reviewed By: dzhulgakov

Differential Revision: D14511711

Pulled By: houseroad

fbshipit-source-id: a7bb6ac84c9133fe94d3fe2f1a8566faed14a136

5 years agoadd Android NDK param to CI docker build script (#18782)
Jiakai Liu [Wed, 3 Apr 2019 17:42:30 +0000 (10:42 -0700)]
add Android NDK param to CI docker build script (#18782)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18782
ghimport-source-id: 6c4bde7dc835b59209c1d5f7b243f00c9fe99de2

Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#18782 [pytorch] add Android NDK param to CI docker build script**

Inspired by discussion: https://github.com/pytorch/pytorch/pull/16242

Reviewed By: dreiss

Differential Revision: D14739471

fbshipit-source-id: 0a081045186cbf359eb3cdadee722741cd8cd62f

5 years agoUpgrade mkldnn-bridge for dnnlowp support (#16308)
Gu, Jinghui [Wed, 3 Apr 2019 17:29:19 +0000 (10:29 -0700)]
Upgrade mkldnn-bridge for dnnlowp support (#16308)

Summary:
The mkldnn-bridge is upgraded in this PR to support DNNLOWP operators.
Meanwhile, APIs have been updated in caffe2 to use latest version.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16308

Differential Revision: D14697018

Pulled By: yinghai

fbshipit-source-id: ca952589098accb08295fd5aa92924c61e74d69c

5 years agoadd 'abs' builtin
Michael Kösel [Wed, 3 Apr 2019 17:11:33 +0000 (10:11 -0700)]
add 'abs' builtin

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/18502

Differential Revision: D14750173

Pulled By: eellison

fbshipit-source-id: 359cf08938ada442ca1a3b3ea14022ce10229499

5 years agoFix dense Embedding to work with double backward (#9078)
kshitij12345 [Wed, 3 Apr 2019 16:16:29 +0000 (09:16 -0700)]
Fix dense Embedding to work with double backward (#9078)

Summary:
Fixes : #6469

1. `ATen/native/native_functions.yml` had [dispatch](https://github.com/pytorch/pytorch/blob/03e7953a98875c0164cb8e2c19b45800e85f4347/aten/src/ATen/native/native_functions.yaml#L451-L455) variants for for `embedding_dense_backward` , however `embedding_backward` explicitly made [call](https://github.com/pytorch/pytorch/blob/03e7953a98875c0164cb8e2c19b45800e85f4347/aten/src/ATen/native/Embedding.cpp#L35-L45) to it, thus leading to error.

2. In case of CUDA type tensor, the function crashed used to crash on dereferencing of indices's data [pointer](https://github.com/pytorch/pytorch/blob/03e7953a98875c0164cb8e2c19b45800e85f4347/aten/src/ATen/native/Embedding.cpp#L93).

Both have been solved and checked against (on CUDA and CPU)

1.  As mentioned in the issue
```
import torch

class Test(torch.nn.Module):

    def __init__(self):
        super(Test,self).__init__()
        self.embd = torch.nn.Embedding(1000, 100)
        self.dense = torch.nn.Linear(100, 1)

    def forward(self, inp):
        inp = self.embd(inp)
        return self.dense(inp)

test = Test()
inp = torch.tensor([0,1,2,1,1])
out = test(inp)
raw_loss = out.mean(dim=0)

loss_grad = torch.autograd.grad(outputs=raw_loss,
                         inputs=list(test.parameters()),
                         retain_graph=True, create_graph=True, only_inputs=True)
norm = sum([param.norm()**2 for param in loss_grad])
loss = raw_loss + norm

loss.backward(retain_graph=True)

print(test.embd.weight.grad)

```

2. Test Script
```
import torch
import time
start = time.time()
l = [1,1]*100
input = torch.tensor([[1,0],[1,0]],device='cpu')
embedding_matrix = torch.tensor([[1.0,3.0],[2.0,4]],requires_grad=True,device='cpu')

sq = embedding_matrix * embedding_matrix
emb = torch.nn.functional.embedding(input, sq,scale_grad_by_freq=False)

print('Embedding Matrix')
print(embedding_matrix)
print('-----------------')

sum_ = emb.sum()#prod.sum()

loss_grad, = torch.autograd.grad(outputs=sum_,inputs=embedding_matrix,create_graph=True)

print('Gradient')
print(loss_grad)
print('-----------------')

sum2_ = sum_ + loss_grad.sum()
print(sum2_)
sum2_.backward()

print(embedding_matrix.grad)
print(time.time() - start)
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9078

Reviewed By: ezyang

Differential Revision: D14691901

Pulled By: soumith

fbshipit-source-id: 78e2612ba39080be564c876311671eb5a0119a0f

5 years agoHighlight NCCL all_reduce and all_gather requirements (#18741)
Shen Li [Wed, 3 Apr 2019 16:06:09 +0000 (09:06 -0700)]
Highlight NCCL all_reduce and all_gather requirements (#18741)

Summary:
See #18689
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18741

Differential Revision: D14726874

Pulled By: mrshenli

fbshipit-source-id: a92404c653e3c62fc23fa3ccacfb3b2959b2e307

5 years agoUpdating submodules
svcscm [Wed, 3 Apr 2019 15:25:14 +0000 (08:25 -0700)]
Updating submodules

Reviewed By: zpao

fbshipit-source-id: ea0b06ce68d3fd6092eaea7c835a8b51c1120ea0

5 years agoMake it possible for users for select /Zi or /ZI over /Z7 when using MSVC (#18790)
peter [Wed, 3 Apr 2019 15:19:45 +0000 (08:19 -0700)]
Make it possible for users for select /Zi or /ZI over /Z7 when using MSVC (#18790)

Summary:
Fixes https://github.com/pytorch/pytorch/issues/18701.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18790

Differential Revision: D14748195

Pulled By: ezyang

fbshipit-source-id: e50df1b5ca199a88d7b5ea3ea45d25d23cd31a27

5 years agouse optimization in D14020675 (#16945)
Jongsoo Park [Wed, 3 Apr 2019 14:55:02 +0000 (07:55 -0700)]
use optimization in D14020675 (#16945)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16945

As title

Reviewed By: jianyuh

Differential Revision: D14020769

fbshipit-source-id: fc0f05fcc57bfe9b4aa0c5750060d7b2ba57dd7a

5 years agoAdd device and dtype to storage. (#18749)
Gregory Chanan [Wed, 3 Apr 2019 14:52:54 +0000 (07:52 -0700)]
Add device and dtype to storage. (#18749)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18749
ghimport-source-id: 9026a037f5e11cdb9ccd386f4b6b5768b9c3259b

Stack from [ghstack](https://github.com/ezyang/ghstack):
* #18751 Disallow changing the device of a tensor via set_.
* #18750 Use non-legacy constructors for tensor deserialization.
* **#18749 Add device and dtype to storage.**

The goal here is to fix our serialization, which currently depends on the legacy constructors.  Having dtype and device on Storage allows us to use the non-legacy constructors.

This fits somewhat along our goal of removing Storage, my having Storage act like a Tensor.

Differential Revision: D14729516

fbshipit-source-id: bf4a3e8669ad4859931f4a3fa56df605cbc08dcb

5 years agoUse non-legacy constructors for tensor deserialization. (#18750)
Gregory Chanan [Wed, 3 Apr 2019 14:51:15 +0000 (07:51 -0700)]
Use non-legacy constructors for tensor deserialization. (#18750)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18750
ghimport-source-id: f1475cfb67841c41d9867d4429ba9125d5c7dd07

Stack from [ghstack](https://github.com/ezyang/ghstack):
* #18751 Disallow changing the device of a tensor via set_.
* **#18750 Use non-legacy constructors for tensor deserialization.**
* #18749 Add device and dtype to storage.

Deserialization currently uses legacy constructors.  This is bad because we need to maintain them, but there is a more immediate problem:
1) We are trying to implement device caching on TensorImpl to get rid of a virtual dispatch
2) This doesn't work if one is able to change the device of a Tensor underlying a Variable.
3) Deserialization does 2)

So the plan is to change deserialization, then enforce that we don't change the device out from underneath a Variable.

Differential Revision: D14729513

fbshipit-source-id: 090d6cdb375b94dc1bf4f554b2df243952b8cdc6

5 years agoAdded numpy conversion (#18505)
Iurii Zdebskyi [Wed, 3 Apr 2019 14:22:38 +0000 (07:22 -0700)]
Added numpy conversion (#18505)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18505
ghimport-source-id: f3c9b9251e5793f9e192f587194ddfebb45facc1

Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#18505 [WIP]Added numpy conversion**
* #18166 Bool Tensor for CUDA

Differential Revision: D14646403

fbshipit-source-id: 79d39d692c778ce1981c1d35b1c33e3d93111041

5 years agoRemove THTensor_(newUnfold). (#18773)
Gregory Chanan [Wed, 3 Apr 2019 14:05:16 +0000 (07:05 -0700)]
Remove THTensor_(newUnfold). (#18773)

Summary:
It's not used and unfold's use of `device_guard: False` is scary.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18773

Differential Revision: D14736526

Pulled By: gchanan

fbshipit-source-id: 6281a284bee45fa5038783e4c1ed4d1ed7ca81ab

5 years agotemp fix for flake8 error (#18788)
mingzhe0908 [Wed, 3 Apr 2019 05:49:49 +0000 (22:49 -0700)]
temp fix for flake8 error (#18788)

Summary:
Fix lint error
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18788

Reviewed By: houseroad

Differential Revision: D14741840

Pulled By: mingzhe09088

fbshipit-source-id: 1fa630e3c6e606e3d78fe8293e5b0e7ea1b78da3

5 years agoFix flake8 issues
Igor Fedan [Wed, 3 Apr 2019 04:10:22 +0000 (21:10 -0700)]
Fix flake8 issues

Summary: Pull Request resolved: https://github.com/pytorch/pytorch/pull/18762

Reviewed By: houseroad

Differential Revision: D14734152

Pulled By: ifedan

fbshipit-source-id: 5adf123f88273895ad34ee9041896358d686de08

5 years agoChange ReinitializeTensor to use C10_LOG_FIRST_N (#18531)
Jerry Zhang [Wed, 3 Apr 2019 03:54:28 +0000 (20:54 -0700)]
Change ReinitializeTensor to use C10_LOG_FIRST_N (#18531)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18531

Currently we use C10_LOG_EVERY_MS to log the data type change, but it pollutes the log of some service,
we would like to change it to C10_LOG_FIRST_N to prevent that.

Reviewed By: dzhulgakov

Differential Revision: D14647704

fbshipit-source-id: b84e4002bd4aa94d616133cd1049c3d4ab05386e

5 years agoAdd support for getting TensorProto argument (#18364)
Yinghai Lu [Wed, 3 Apr 2019 03:52:58 +0000 (20:52 -0700)]
Add support for getting TensorProto argument (#18364)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18364

att

Reviewed By: bddppq

Differential Revision: D14584784

fbshipit-source-id: 03f9207d5cf4f7f4b812428a931edbcdcb21ca8d

5 years agomake test module hook use save/load (#18284)
Michael Suo [Wed, 3 Apr 2019 01:06:07 +0000 (18:06 -0700)]
make test module hook use save/load (#18284)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18284
ghimport-source-id: 5a92c03fda19072ffb6afd40e0f56806716c7be6

Stack from [ghstack](https://github.com/ezyang/ghstack):
* #18296 [jit] Add namespacing for ScriptClasses
* **#18284 [jit] make test module hook use save/load**
* #18211 [jit] Turn script_type_parser into a class
* #18148 [jit] python interop for script classes

Instead of python-printing and comparing strings (which does not capture
depdency information, etc.), use save/load on in-memory buffers and
compare the main module contents inside the buffer

Reviewed By: ailzhang

Differential Revision: D14581129

fbshipit-source-id: 52264ae9ce076775ab3fd1a0c32c8d6f6677a903

5 years agoAdd ability to specialize class types to ArgumentSpec (#18314)
Zachary DeVito [Wed, 3 Apr 2019 00:33:06 +0000 (17:33 -0700)]
Add ability to specialize class types to ArgumentSpec (#18314)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18314
ghimport-source-id: 8cecb768d476ab19c9460f39c8f94a764e4cb052

Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#18314 Add ability to specialize class types to ArgumentSpec**
* #18226 Add Slot type to abstract the raw pointers being used for slots.

Differential Revision: D14574395

fbshipit-source-id: cc3af6e56e9ae52990f4a1ad56ecceaa2d493577

5 years agoOperator-level performance microbenchmarks (#18740)
Mingzhe Li [Wed, 3 Apr 2019 00:03:23 +0000 (17:03 -0700)]
Operator-level performance microbenchmarks (#18740)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18740

Test utilities for writing Caffe2/PyTorch performance microbenchmarks. Brief description of the file structure

* benchmark_core.py : core utiltiites for running microbenchmark tests
* benchmark_caffe2.py : Caffe2 specific benchmark utilitites
* benchmark_pytorch.py: PyTorch specific benchmark utilities
* benchmark_runner.py : Main function. Currently it can run the microbenchmark tests in a stand-alone mode. The next step is to have this integrate with AI-PEP.

The utilities are located at https://github.com/pytorch/pytorch/tree/master/test to have access to both Caffe2/PyTorch Python's frontend.

Include two operator microbenchmarks; support both Caffe2/PyTorch:
* MatMul
* Add

Reference: PyTorch benchmarks : https://github.com/pytorch/benchmark/tree/master/timing/python. In this work, we start with two example binary operators MatMul and Add, but eventually we should to cover unary operators like in the PyTorch benchmark repo.

Reviewed By: zheng-xq

Differential Revision: D13887111

fbshipit-source-id: b7a56b95448c9ec3e674b0de0ffb96af4439bfce

5 years agoBool Tensor for CUDA (#18166)
Iurii Zdebskyi [Tue, 2 Apr 2019 23:10:43 +0000 (16:10 -0700)]
Bool Tensor for CUDA (#18166)

Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18166
ghimport-source-id: a8e2ba2d966e49747a55701c4f6863c5e24d6f14

Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#18166 Bool Tensor for CUDA**
* #18165 Resolved comments from Bool Tensor for CPU PR
------

This PR enables bool tensor creation and some basic operations for the CPU backend. This is a part of Bool Tensor feature implementation work. The whole plan looks like this:
1. Storage Implementation [Done]
2. Tensor Creation.
a) CPU [Done]
b) CUDA [This PR]
3. Tensor Conversions.
4. Tensor Indexing.
5. Tensor Operations.
6. Back compatibility related changes.

Change:
Enable bool tensor in CUDA with the following operations:

    torch.zeros
    torch.tensor
    torch.ones
    torch.rand/rand_like/randint/randint_like
    torch.full
    torch.full_like
    torch.empty
    torch.empty_like

Tested via unit tests and local scripts.

Differential Revision: D14605104

fbshipit-source-id: b7d7340a7d70edd03a109222d271e68becba762c