Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/33779
This should eliminate random warnings and print spew from test_jit.
It also fixes a bug where we weren't properly comparing captured outputs
(!)
Test Plan: Imported from OSS
Differential Revision: D20124224
Pulled By: suo
fbshipit-source-id: 9241d21fdf9470531b0437427b28e325cdf08d3a
Summary:
We want to run ONNX checker only when selected operator type is ONNX, and nowhere else. This PR updates the logic in the exporter.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/33522
Reviewed By: hl475
Differential Revision: D19983954
Pulled By: houseroad
fbshipit-source-id: 15db726321637a96fa110051cc54e9833e201133
Summary:
There are large models such as GPT2-large which cannot be exported with the current exporter because of the 2GB protobuf limit (e.g. see https://github.com/pytorch/pytorch/issues/19277). ONNX spec specifies a special format for large (> 2GB) models. This PR adds support for exporting large models in ONNX large model format in the PyTorch-ONNX exporter.
This is the first PR for this feature that enables the end-to-end execution. Tests for large model export have been added. We may need follow-up PRs to refine this workflow based on user feedback.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/33062
Reviewed By: hl475
Differential Revision: D19782292
Pulled By: houseroad
fbshipit-source-id: e972fcb066065cae6336aa91c03023d9c41c88bd
Summary:
Currently, custom ops are registered for a specific opset version.
For example, all torchvision custom ops are registered for opset 11, and cannot be exported into higher opset versions. This PR extends op registration to higher opset versions.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/32943
Reviewed By: hl475
Differential Revision: D19739406
Pulled By: houseroad
fbshipit-source-id: dd8b616de3a69a529d135fdd02608a17a8e421bc
Summary:
The PR https://github.com/pytorch/pytorch/pull/31791 adds support for float[] constant, which affects some cases of ONNX interpolate support.
This PR adds float[] constants support in ONNX, updates interpolate in ONNX, and re-enable the disabled tests.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/32554
Reviewed By: hl475
Differential Revision: D19566596
Pulled By: houseroad
fbshipit-source-id: 843f62c86126fdf4f9c0117b65965682a776e7e9
Summary:
Included the ONNX model checker code in the ONNX export
this will force onnx checker to run for all models that get exported.
This should help with validating exported models.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/32298
Reviewed By: hl475
Differential Revision: D19538251
Pulled By: houseroad
fbshipit-source-id: eb20b124fe59200048f862ddaf20f6c59a0174d5
Summary:
While ONNX does not currently directly support the Dim operation on a
tensor, we can provide the same functionality with two ONNX operations.
This allows us to support Dim for all opsets. It may be adventageous to
add support for Dim into a future ONNX opset, and use that for more
efficient code.
While testing dim op found that there is an issue with empty blocks
withing if statements. Modified graph generation to prevent generation
of empty if blocks.
Fixes https://github.com/pytorch/pytorch/issues/27569
Pull Request resolved: https://github.com/pytorch/pytorch/pull/31928
Reviewed By: hl475
Differential Revision: D19376602
Pulled By: houseroad
fbshipit-source-id: 111682b058a5341f5cca6c1a950c83ae412a4c6c
Summary:
Updated to export API:
When calling this API, a dict containing the custom opsets (domain and version) used to export the model could be provided.
We allow registering one custom opset (domain, version) per ONNX opset. So, when exporting an operator from a custom domain, users need to pass this pair. Default custom opset version is 1.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/29752
Reviewed By: hl475
Differential Revision: D18703662
Pulled By: houseroad
fbshipit-source-id: 84d22557d132b526169051193d730761798fce60
Summary:
- [x] Add more comments and refactor the logic of `ReshapeToAdvancedIndexingFormat`
- [x] Add more description here. Cases that are/aren't supported, and how they are supported.
- [x] Need to merge this PR https://github.com/pytorch/pytorch/issues/27186 to enable testing inplace operators.
We are now supporting exporting aten::copy_ and aten::index_put to ONNX.
Here's a breakdown of the different cases in PyTorch code.
```
# Case 1: Scalar Indices
x[0, 1, 2] = data
# Case 2: Slice Indices
x[1:3, :, ::2] = data
# Case 3: Ellipsis Indices
x[..., 0] = data
# Case 4: Tensor Indices
ind1 = torch.tensor([0, 2])
ind2 = torch.tensor([1, 1])
x[ind1, ind2] = data
# Case 5: Mixing all the above cases
ind1 = torch.tensor([0, 2])
ind2 = torch.tensor([1, 1])
x[1:3, ind1, ind2, ..., 3] = data
```
Limitations:
Tensor indices must be consecutive, and 1-d tensors.
```
# Supported
ind1 = torch.tensor([0, 2])
ind2 = torch.tensor([1, 1])
x[ind1, ind2] = data
# Not supported
ind1 = torch.tensor([0, 2])
ind2 = torch.tensor([1, 1])
ind3 = torch.tensor([[0], [1]])
x[ind1, :, ind2] = data
x[ind3] = data
```
Negative indices are not supported.
```
# Not supported
x[-1] = data
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26941
Differential Revision: D17951030
Pulled By: houseroad
fbshipit-source-id: 4357777072f53aa0bc4b297aa1ee53457a7f8dec
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/30892
Fixes all outstanding lints and actually installs a properly configured
flake8
Test Plan: Imported from OSS
Differential Revision: D18862825
Pulled By: suo
fbshipit-source-id: 08e9083338a7309272e17bb803feaa42e348aa85
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/30679
Caffe2 expects quantized ops to be in NHWC format while pytorch inputs are in NCHW.
Add a jit pass to insert permutes to convert from nchw2nhwc before each conv op and add nhwc2nchw permute after the conv op.
Using graph rewriter to find consecutive redundant permutes and remove them from the graph
Test Plan:
python test/onnx/test_pytorch_onnx_caffe2_quantized.py TestQuantizedOps
Imported from OSS
Differential Revision: D18790518
fbshipit-source-id: 4dd39cf0b31b21f5586c0edfdce2260d4e245112
Summary:
The PyTorch exporter does not add any name to the ONNX operators in the exported graph. A common request is to add names to op nodes by default. This helps the readability of the graph in visualization tools such a Netron, or when the ONNX graph is printed as a string. Also, it helps with the debuggability of the ONNX graph.
Therefore this PR adds name to operators in the exporters. The names follow a simple format, <op_type>_<index>. Expect files for tests in `test/onnx/test_operators.py` have been updated.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/27342
Reviewed By: hl475
Differential Revision: D17790979
Pulled By: houseroad
fbshipit-source-id: 1eaae88b5f51f152735a2ff96e22827837e34d9d
Summary:
This should resolve https://github.com/pytorch/pytorch/issues/29008. This flag has two effects on the tracer.
- Remove the underscroll for inplace operators. E.g.: index_put_ ==> index_put. This is handled in utils.py separately as well.
- Add out as input for backward computation.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/29466
Reviewed By: hl475
Differential Revision: D18422815
Pulled By: houseroad
fbshipit-source-id: 317b6a3c8a5751fe6fe49d7543e429d281ed0d6d
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/29694
This PR adds preliminary support required to be able to run quantized pytorch models on a C2 backend.
For quantized ops we use a custom domain name 'caffe2' to register the ops if they are in the "quantized" namespace.
The change also adds JIT pass to unpack the quantized weights and insert the unpacked values into the graph.
The actual tensor values are looked up from the params dict.
Test Plan:
python test/onnx/test_pytorch_onnx_caffe2.py TestQuantizedOps
Imported from OSS
Reviewed By: houseroad
Differential Revision: D18467130
fbshipit-source-id: 53ebd8c43935f7d7e74305dad6c231a2247df176
Summary:
Currently, `keep_initializers_as_input` argument in `torch.onnx.export` API can be used to choose whether to export an ONNX model with IR v3 or v4 semantics. Currently, the implementation does not check for which opset is being used for export. This is an issue because ONNX IR v4 is valid only for opset 9 and above (as listed [here](https://github.com/onnx/onnx/releases/tag/v1.4.0)), and opset 8 or lower export with `keep_initializers_as_input=False` will create a illegal ONNX graph.
This change fixes this by introducing a check on opset version when deciding whether to export ONNX IR v3 or v4.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/28990
Reviewed By: hl475
Differential Revision: D18352523
Pulled By: houseroad
fbshipit-source-id: 7e9055d405c3faf52b80a8de0d04186d4c350c15
Summary:
Currently ONNX constant folding (`do_constant_folding=True` arg in `torch.onnx.export` API) supports only opset 9 and 10 of ONNX. Opset 11 support was recently introduced in the ONNX exporter. For opset 11, it is currently a no-op. This change enables ONNX constant folding for opset 11. Specifically there are three main changes:
1) Turn on constant folding ONNX pass for opset 11.
2) Enable constant folding tests in `test/onnx/test_utility_funs.py` and `test/onnx/test_pytorch_onnx_onnxruntime.py` for opset 11.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/29011
Reviewed By: hl475
Differential Revision: D18306998
Pulled By: houseroad
fbshipit-source-id: afeed21ca29e01c278612e51dacd93397dd6e2d8
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/28242
There is no reason to have it in a general API of Module/Method - it's
just another graph pass. It was there because some time ago modules were
not first class and all graphs were lowered. After that changed, this
API was added for easier transition, but now we don't need it anymore.
Test Plan: Imported from OSS
Differential Revision: D17986724
Pulled By: ZolotukhinM
fbshipit-source-id: 279a1ec450cd8fac8164ee581515b09f1d755630
Summary:
We currently support exporting traced interpolate ops to ONNX.
Scripting interpolate op invokes aten::__interpolate in the Torch IR (instead of aten::upsample_[mode][dim]d), which we do not support yet.
This PR implements the ONNX symbolic for __interpolate() to support exporting interpolate in scripting scenarios.
Related open issue: https://github.com/pytorch/pytorch/issues/25807
Pull Request resolved: https://github.com/pytorch/pytorch/pull/27566
Reviewed By: hl475
Differential Revision: D17817731
Pulled By: houseroad
fbshipit-source-id: e091793df503e2497f24821cf2954ff157492c75
Summary:
ONNX does not support dictionaries for inputs and output. The reason is that the arg flattening and unflattening does not handle Dictionary types.
This PR adds flattening/unflattening support for dictionaries and strings.
However this feature should be handled with caution for input dictionaries; and users need to verify their dict inputs carefully, and keep in mind that dynamic lookups are not available.
This PR will allow exporting cases where models have dictionnary outputs (detection and segmentation models in torchvision), and where dictionary inputs are used for model configurations (MultiScaleRoiAlign in torchvision).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25889
Reviewed By: hl475
Differential Revision: D17613605
Pulled By: houseroad
fbshipit-source-id: c62da4f35e5dc2aa23a85dfd5e2e11f63e9174db
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26738
someone may use torch._export directly. Here we change the onnx_export_type's default value to None,
and if it's pytorch onnx caffe2 bundle, we set it to ONNX_ATEN_FALLBACK, otherwise, it's ONNX.
Test Plan: ci
Reviewed By: hl475
Differential Revision: D17546452
fbshipit-source-id: 38e53926e2b101484bbbce7b58ebcd6af8c42438
Summary:
This is a follow-up PR for https://github.com/pytorch/pytorch/pull/23284. In that PR we had removed changing the default behavior for `keep_initializers_as_input` argument to the export API. With this PR we are enabling that change in that if `keep_initializers_as_input` is not specified then value/behavior for this argument is chosen automatically depending on whether the export type is ONNX or not.
This was part of the earlier PR was removed for further review. The test points have also been updated.
This change may fail some internal tests which may require explicitly setting `keep_initializers_as_input=True` to preserve old behavior.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26146
Reviewed By: hl475
Differential Revision: D17369677
Pulled By: houseroad
fbshipit-source-id: 2aec2cff50d215714ee8769505ef24d2b7865a11
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/26487
The way it is implemented currently is bad because while we're inlining
to a graph G, we are also mutating all the graphs that are being
inlined. The problem is that the graphs we're inlining are usually the
original graphs of functions, so we're silently changing them behind the
scenes, and we don't have a way to recover 'unoptimized' graphs
afterwards.
Test Plan: Imported from OSS
Differential Revision: D17485748
Pulled By: ZolotukhinM
fbshipit-source-id: 6094ef56077240e9379d4c53680867df1b6e79ef
Summary:
This pass tries to resolve scalar type mismatch issues between input tensors introduced by the implicit type conversions on scalars.
e.g. https://github.com/pytorch/pytorch/issues/23724
Pull Request resolved: https://github.com/pytorch/pytorch/pull/24378
Reviewed By: hl475
Differential Revision: D17088682
Pulled By: houseroad
fbshipit-source-id: 3de710f70c3b70b9f76fd36a7c4c76e168dbc756
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/25052
Previously we would not inline nested functions, now we do.
Test Plan: Imported from OSS
Differential Revision: D16973848
Pulled By: suo
fbshipit-source-id: 94aa0b6f84a2577a663f4e219f930d2c6396d585
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23799
Before, we inlined as part of the initial IR generation process, which
has a few disadvantages:
1. It loses information about what nodes came from which function/method
calls. Other parties who want to implement transformations on the
function/module level don't have a reliable way of doing so.
2. It duplicates a ton of code if we are inlining the same
function/method a tons of times.
After this PR: inline is deferred to the optimization stage, so
optimizations that rely on inlining will still work. But things get
serialized with the function/method calls in.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23799
Differential Revision: D16652819
Test Plan: Imported from OSS
Reviewed By: jamesr66a
Pulled By: suo
fbshipit-source-id: a11af82aec796487586f81f5a9102fefb6c246db
Summary:
Existing code adds two enumerators to the set instead of forming their union.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23974
Differential Revision: D16732762
Pulled By: ezyang
fbshipit-source-id: 787737b7cf4b97ca4e2597e2da4a6ade863ce85c
Summary:
Starting ONNX IR version 4, the initializers in the ONNX graph do not have to be inputs of the graphs. This constraint, which existed in IR version 3 and earlier, was relaxed in IR version 4. This PR provides an API level argument to allow ONNX export with the relaxed constraint of IR version 4, i.e. provides the option to not include initializers as inputs. This allows backends/runtimes to do certain optimizations, such as constant folding, better.
*Edit*: After discussion with houseroad we have the following behavior. For any OperatorExportType, except OperatorExportTypes.ONNX, the current status of export is maintained in this PR by default. However, the user can override it by setting the `keep_initializers_as_inputs` argument to the export API. But when exporting to ONNX, i.e. OperatorExportType is OperatorExportTypes.ONNX, the current status is changed in that by default the initializers are NOT part of the input. Again, the default can be overridden by setting the `keep_initializers_as_inputs` argument.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23284
Differential Revision: D16459961
Pulled By: bddppq
fbshipit-source-id: b8f0270dfaba47cdb8e04bd4cc2d6294f1cb39cf
Summary:
Don't automatically unwrap top layer DataParalllel for users. Instead, we provide useful error information and tell users what action to take.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23365
Reviewed By: zrphercule
Differential Revision: D16514273
Pulled By: houseroad
fbshipit-source-id: f552de5c53fb44807e9d9ad62126c98873ed106e
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/23180
This pass needs to be run later because it breaks jit graph invariants and the lower_all_tuples pass still needs a valid jit graph.
Reviewed By: houseroad
Differential Revision: D16427680
fbshipit-source-id: 427c7e74c59a3d7d62f2855ed626cf6258107509
Summary:
This is an extension to the original PR https://github.com/pytorch/pytorch/pull/21765
1. Increase the coverage of different opsets support, comments, and blacklisting.
2. Adding backend tests for both caffe2 and onnxruntime on opset 7 and opset 8.
3. Reusing onnx model tests in caffe2 for onnxruntime.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/22421
Reviewed By: zrphercule
Differential Revision: D16225518
Pulled By: houseroad
fbshipit-source-id: 01ae3eed85111a83a0124e9e95512b80109d6aee
Summary:
Currently ONNX constant folding (`do_constant_folding=True` arg in `torch.onnx.export` API) supports only opset 9 of ONNX. For opset 10, it is a no-op. This change enables ONNX constant folding for opset 10. Specifically there are three main changes:
1) Turn on constant folding ONNX pass for opset 10.
2) Update support for opset 10 version of `onnx::Slice` op for backend computation during constant folding.
3) Enable constant folding tests in `test/onnx/test_utility_funs.py` for multiple opsets (9 and 10).
Pull Request resolved: https://github.com/pytorch/pytorch/pull/22515
Reviewed By: zrphercule
Differential Revision: D16189336
Pulled By: houseroad
fbshipit-source-id: 3e2e748a06e4228b69a18c5458ca71491bd13875
Summary:
- Fix typo in ```torch/onnx/utils.py``` when looking up registered custom ops.
- Add a simple test case
1. Register custom op with ```TorchScript``` using ```cpp_extension.load_inline```.
2. Register custom op with ```torch.onnx.symbolic``` using ```register_custom_op_symbolic```.
3. Export model with custom op, and verify with Caffe2 backend.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/21321
Differential Revision: D16101097
Pulled By: houseroad
fbshipit-source-id: 084f8b55e230e1cb6e9bd7bd52d7946cefda8e33
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/22499
Another place where onnx export is running dead code elimination after making the jit graph invalid. Fixing it.
Reviewed By: houseroad
Differential Revision: D16111969
fbshipit-source-id: 5ba80340c06d091988858077f142ea4e3da0638c
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/22476
Dead code elimination assumes a valid jit graph because it checks if operators have side effects.
The onnx export path destroys the jit graph right before calling dead code elimination, but it actually doesn't care about side effects.
We can just call dead code elimination and disable side effect lookup and things should work.
Reviewed By: houseroad
Differential Revision: D16100172
fbshipit-source-id: 8c790055e0d76c4227394cafa93b07d1310f2cea
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/22413
_jit_pass_erase_number_types invalidates the jit graph but parts of _jit_pass_onnx rely on having a valid jit graph.
This splits _jit_pass_onnx into _jit_pass_onnx_remove_print and _jit_pass_onnx_preprocess_caffe2 (which rely on the valid jit graph), runs these before _jit_pass_erase_number_types,
and then runs the rest of _jit_pass_onnx after _jit_pass_erase_number_types
Reviewed By: houseroad
Differential Revision: D16079890
fbshipit-source-id: ae68b87dced077f76cbf1335ef3bf89984413224
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/22319
The onnx pass replacing ints with Tensors produces an invalid JIT graph. It should only be called right before the onnx pass.
Also, it should only be called if we actually export to onnx.
Reviewed By: houseroad
Differential Revision: D16040374
fbshipit-source-id: e78849ee07850acd897fd9eba60b6401fdc4965b
Summary:
- [x] Add tests after https://github.com/pytorch/pytorch/pull/20256 is merged
- Support exporting ScriptModule with inputs/outputs of arbitrarily constructed tuples.
- Moved the assigning of output shapes to after graph conversion to ONNX is completed. By then all tuples in the IR has already been lowered by the pass ```_jit_pass_lower_all_tuples```. If assigning output shapes is required to happen before that, we'll need to hand parse the tuple structures in the graph, and repeat the same logic in ```_jit_pass_lower_all_tuples```. Handling inputs is easier because all tuple information is encoded within the input tensor type.
- Swap the order of ```_jit_pass_lower_all_tuples``` and ```_jit_pass_erase_number_types```. Ops like ```prim::TupleIndex``` relies on index being a scalar. ```_jit_pass_erase_number_types``` will convert these kind of scalars to tensors.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/20784
Reviewed By: zrphercule
Differential Revision: D15484171
Pulled By: houseroad
fbshipit-source-id: 4767a84038244c929f5662758047af6cb92228d3
Summary:
Input argument `f` in `_model_to_graph()` method in `torch/onnx/utils.py` is unused. This PR removes it. If there's a reason to keep it around, please let me know.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/19647
Reviewed By: dzhulgakov
Differential Revision: D15071720
Pulled By: houseroad
fbshipit-source-id: 59e0dd7a4d5ebd64d0e30f274b3892a4d218c496
Summary:
Strip the doc_string by default from the exported ONNX models (this string has the stack trace and information about the local repos and folders, which can be confidential).
The users can still generate the doc_string by specifying add_doc_string=True in torch.onnx.export().
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18882
Differential Revision: D14889684
Pulled By: houseroad
fbshipit-source-id: 26d2c23c8dc3f484544aa854b507ada429adb9b8
Summary:
Almost there, feel free to review.
these c10 operators are exported to _caffe2 domain.
TODO:
- [x] let the onnx checker pass
- [x] test tensor list as argument
- [x] test caffe2 backend and converter
- [x] check the c10 schema can be exported to onnx
- [x] refactor the test case to share some code
- [x] fix the problem in ONNX_ATEN_FALLBACK
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18210
Reviewed By: zrphercule
Differential Revision: D14600916
Pulled By: houseroad
fbshipit-source-id: 2592a75f21098fb6ceb38c5d00ee40e9e01cd144
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
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/18598
ghimport-source-id: c74597e5e7437e94a43c163cee0639b20d0d0c6a
Stack from [ghstack](https://github.com/ezyang/ghstack):
* **#18598 Turn on F401: Unused import warning.**
This was requested by someone at Facebook; this lint is turned
on for Facebook by default. "Sure, why not."
I had to noqa a number of imports in __init__. Hypothetically
we're supposed to use __all__ in this case, but I was too lazy
to fix it. Left for future work.
Be careful! flake8-2 and flake8-3 behave differently with
respect to import resolution for # type: comments. flake8-3 will
report an import unused; flake8-2 will not. For now, I just
noqa'd all these sites.
All the changes were done by hand.
Signed-off-by: Edward Z. Yang <ezyang@fb.com>
Differential Revision: D14687478
fbshipit-source-id: 30d532381e914091aadfa0d2a5a89404819663e3
Summary:
So, we will keep the names of ONNX initializers the same as the names in PyTorch state dict.
Later, we will make this as the default behavior.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17551
Reviewed By: dzhulgakov
Differential Revision: D14491920
Pulled By: houseroad
fbshipit-source-id: f355c02e1b90d7ebbebf4be7c0fb6ae208ec795f
Summary:
1) The changes in the new opset won't affect internal pipeline.
2) The CI won't be affected by the ONNX changes.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17736
Reviewed By: zrphercule
Differential Revision: D14358710
Pulled By: houseroad
fbshipit-source-id: 4ef15d2246b50f6875ee215ce37ecf92d555ca6a
Summary:
Similar to `nn.Parameter`s, this PR lets you store any `IValue` on a module as an attribute on a `ScriptModule` (only from the Python front-end currently). To mark something as an attribute, it should wrapped in `jit.Attribute(value, type)` (ex. `self.table = torch.jit.Attribute(table, Dict[str, torch.Tensor])`)
Followup Work:
* (de)serializing for use in C++
* change `self.training` to be a `bool` attribute instead of a buffer
* mutable attributes
* string frontend support
* documentation
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17309
Differential Revision: D14354316
Pulled By: driazati
fbshipit-source-id: 67e08ab5229366b67fbc837e67b58831a4fb3318
Summary:
Currently, serialization of model parameters in ONNX export depends on the order in which they are stored in a container (`list` on Python side and `std::vector` on C++ side). This has worked fine till now, but if we need to do any pass on that graph that mutates the parameter list, then strictly order-based serialization may not work.
This PR is the first in a set to bring in more passes (such as constant folding) related to ONNX export. This PR lays the groundwork by moving the serialization in ONNX export from order-based to name based approach, which is more amenable to some of the passes.
houseroad - As discussed this change uses a map for export, and removes the code from `export.cpp` that relies on the order to compute initializer names.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17420
Differential Revision: D14361993
Pulled By: houseroad
fbshipit-source-id: da93e945d55755c126de06641f35df87d1648cc4
Summary:
Still wip, need more tests and correct handling for opset 8 in symbolics.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16068
Reviewed By: zrphercule
Differential Revision: D14185855
Pulled By: houseroad
fbshipit-source-id: 55200be810c88317c6e80a46bdbeb22e0b6e5f9e
Summary:
Trying to land again, make prim::None into a case of prim::Constant. Reverted the previous landing because it broke an important onnx export test.
https://github.com/pytorch/pytorch/pull/16160
Pull Request resolved: https://github.com/pytorch/pytorch/pull/17186
Differential Revision: D14115304
Pulled By: eellison
fbshipit-source-id: 161435fc30460b4e116cdd62c7b2e5b94581dcb7
Summary:
This change simplifies analysis done on constants since prim::None does not need to be handled separately now. To check if a constant node is None, use node->isNone().
Next step will be to remove prim::Undefined.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/16160
Differential Revision: D14109636
Pulled By: eellison
fbshipit-source-id: d26fd383976163a2ddd4c24984bd672a541cc876
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
Summary:
* With the update of split output to dynamic list it breaks the export to onnx.
Now split ir becomes two ops: 1. Dynamic[] <= Split(), and 2. out1, out2, out3
<= Prim::ListUnpack. In this fix these two consecutive ops get fused when being
exported to onnx.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/15092
Reviewed By: dzhulgakov
Differential Revision: D13583832
Pulled By: houseroad
fbshipit-source-id: 3eb18c871e750921ad6d5cc179254bee9bcf4c99
Summary:
[ note: stacked on expect files changes, will unstack once they land ]
This adds DeviceObjType (cannot use DeviceType it is already an enum)
to the type hierarchy and an isDevice/toDevice pair to IValue.
Previous hacks which used an int[] to represent Device are removed
and at::Device is used instead.
Note: the behavior or .to is only a subset of python, we need to
fix the aten op so that it accepts Option[Device] and Optional[ScalarType].
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14666
Reviewed By: suo
Differential Revision: D13290405
Pulled By: zdevito
fbshipit-source-id: 68b4381b292f5418a6a46aaa077f1c902750b134
Summary:
Stacked on https://github.com/pytorch/pytorch/pull/14378, only look at the last commit.
This changes the way methods are defined in TorchScript archives to use
PythonPrint rather than ONNX protobufs.
It also updates torch.proto to directly document the tensor data
structure actually being serialized.
Notes:
* because PythonPrint prints all the methods at once per module, this
removes MethodDef in favor of a single torchscript_area and a separate
caffe2_graphs entry. Note that NetDef's already have method names,
so there is no need or a separate method name entry.
* This switches cpp/pickle area to RecordRef (references to a file in
the container format) since it is possible the data in these arenas
may be large and not suited to json ouput.
* Removes 'annotations' -- annotations should be re-added on the first
commit that actually has a practical use for them. In the current state
it is unlikely they are representing the right information.
* Some expect files have changed because PythonPrint is preserving more
debug name information for parameter names.
* MethodEncoder (the ONNX output format) has been deleted. There is still
some cleanup possible combining EncoderBase and GraphEncode now that there
is only a single pathway using EncoderBase.
* This incorporates the changes from #14397
to define TensorDef
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14400
Reviewed By: suo
Differential Revision: D13231800
Pulled By: zdevito
fbshipit-source-id: af5c1152d0bd6bca8b06c4703f59b161bb19f571
Summary:
This PR adds a `try_outplace` option to the tracer. When `try_outplace` is true, the tracer will attempt to out-of-place ops (similar to how things are done today). When it's false, the correct in-place op is emitted.
I made `try_outplace` false by default, but flipped it to true for ONNX export utils. zdevito jamesr66a, anywhere else I should preserve the existing behavior?
Pull Request resolved: https://github.com/pytorch/pytorch/pull/14254
Reviewed By: eellison
Differential Revision: D13166691
Pulled By: suo
fbshipit-source-id: ce39fdf73ac39811c55100e567466d53108e856b
Summary:
There should really be a single place to erase or do special treatment to the prim::ListConstruct during ONNX export, this will make it consistent across different calls. e.g it will give a correct output graph in the following case:
```python
class Test(torch.nn.Module):
def forward(self, input):
return torch.cat([input, torch.zeros(input.size(0), 1).type_as(input)], dim=1)
```
Before this PR, we have the onnx graph as:
```
graph(%0 : Byte(2, 3)) {
%1 : Long() = onnx::Constant[value={0}](), scope: Test
%2 : Dynamic = onnx::Shape(%0), scope: Test
%3 : Long() = onnx::Gather[axis=0](%2, %1), scope: Test
%4 : Long() = onnx::Constant[value={1}](), scope: Test
%5 : Dynamic = onnx::Unsqueeze[axes=[0]](%3)
%6 : Dynamic = onnx::Unsqueeze[axes=[0]](%4)
%7 : int[] = onnx::Concat[axis=0](%5, %6)
%8 : Float(2, 1) = onnx::ConstantFill[dtype=1, input_as_shape=1, value=0](%7), scope: Test
%9 : Byte(2, 1) = onnx::Cast[to=2](%8), scope: Test
%10 : Byte(2, 4) = onnx::Concat[axis=1](%0, %9), scope: Test
return (%10);
}
```
Which is wrong since onnx does not have a concept of `int[]`, here is the onnx graph after this PR:
```
graph(%0 : Byte(2, 3)) {
%1 : Long() = onnx::Constant[value={0}](), scope: Test
%2 : Dynamic = onnx::Shape(%0), scope: Test
%3 : Long() = onnx::Gather[axis=0](%2, %1), scope: Test
%4 : Long() = onnx::Constant[value={1}](), scope: Test
%5 : Dynamic = onnx::Unsqueeze[axes=[0]](%3)
%6 : Dynamic = onnx::Unsqueeze[axes=[0]](%4)
%7 : Dynamic = onnx::Concat[axis=0](%5, %6)
%8 : Float(2, 1) = onnx::ConstantFill[dtype=1, input_as_shape=1, value=0](%7), scope: Test
%9 : Byte(2, 1) = onnx::Cast[to=2](%8), scope: Test
%10 : Byte(2, 4) = onnx::Concat[axis=1](%0, %9), scope: Test
return (%10);
}
```
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13195
Differential Revision: D12812541
Pulled By: wanchaol
fbshipit-source-id: db6be8bf0cdc85c426d5cbe09a28c5e5d860eb3e
Summary:
This PR changes the compiler to correctly emit in-place operators for augmented assignments (`+=` and friends).
- To better match the Python AST structure, add an `AugAssign` tree view and make `Assign` apply only to `=` assignments.
- Emit those `AugAssign` exprs in the compiler, dispatching to in-place aten ops for tensors and lowering to simple assignments for scalar types.
- In order to preserve (suspect) ONNX export semantics, add a pass to lower the in-place operators to out-of-place operators.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/13364
Differential Revision: D12899734
Pulled By: suo
fbshipit-source-id: bec83be0062cb0235eb129aed78d6110a9e2c146
Summary:
This PR adds optional type to ATen native, autograd, JIT schema and Python Arg parser, closes#9513. It allows us to use optional default values (including None) for function signature and implementations like clamp, etc., and also let us remove the python_default_init hack.
Follow up:
remove python_default_init completely.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/12582
Differential Revision: D10417423
Pulled By: wanchaol
fbshipit-source-id: 1c80f0727bb528188b47c595629e2996be269b89
Summary:
There's some action at a distance issues and not having this is disabling quantization in C2 for prod use cases
ref T34831022
Pull Request resolved: https://github.com/pytorch/pytorch/pull/12538
Differential Revision: D10302931
Pulled By: jamesr66a
fbshipit-source-id: 700dc8c5c4297e942171992266ffb67b815be754
Summary:
- In Python 2, use of `/` (regardless of int/float/Tensor) causes a compiler error if
`from __future__ import division` is not imported in the file.
- The / operator is universally set to do "true" division for integers
- Added a `prim::FloorDiv` operator because it is used in loop unrolling.
The error if users use '/' in python 2 without importing from __future__
occurs when building the JIT AST.
cc apaszke zdevito
Pull Request resolved: https://github.com/pytorch/pytorch/pull/11016
Differential Revision: D9613527
Pulled By: zou3519
fbshipit-source-id: 0cebf44d5b8c92e203167733692ad33c4ec9dac6
Summary:
Operators like aten::chunk used to return a number of tensors, but
now return a list. To make it easier to do shape prop through
aten::chunk and fuse it, I've also introduced prim::ConstantChunk,
which behaves like the previous implementation (has a variable length
output list).
The downside of this PR is that the introduction of more lists to the IR causes the LSTM and MiLSTM graphs to be considered as non-differentiable by the graph executor. I verified that they are still optimize correctly, and my next patch (that changes how the specializations/differentiation works) will restore those.
zdevito
Pull Request resolved: https://github.com/pytorch/pytorch/pull/10949
Reviewed By: zdevito
Differential Revision: D9556823
Pulled By: apaszke
fbshipit-source-id: 33e63b17fc7247cac6cfc05eb7eb9bf069b499ee
Summary:
This was done because it surprising for a decorator to run a function
rather than wrap it, and not simplify the syntax for tracing modules.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/11069
Reviewed By: jamesr66a
Differential Revision: D9583192
Pulled By: zdevito
fbshipit-source-id: b914b7ab4c73c255086465a6576eef3a22de1e13
Summary:
Things like torch.zeros now appear in traces rather than constants.
To continue to support our current level of ONNX export, we run
constant prop to turn these back into constants where possible before
export.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/10935
Differential Revision: D9527427
Pulled By: zdevito
fbshipit-source-id: 552a8bcc01b911251dab7d7026faafdd7a3c758a
Summary:
* Fix the necessary pathways so that tuples and lists can be inputs to the script.
* prevent linear algebra functions from being run in shape prop because
they frequently will error out for nonsense data.
* favor schema-driven python input conversion where possible.
remaining cases where we directly create Stacks without schema are
only for debugging
* Make the error messages when calling script/trace functions more pythonic
* Simplify FlattenTuples -- now that tuples are supported we can choose to only flatten tuples when needed. This may have to be revisited pending onnx test results, but is necessary for making tuple io work.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/10812
Differential Revision: D9477982
Pulled By: zdevito
fbshipit-source-id: ed06fc426e6ef6deb404602a26c435a7fc40ea0c
Summary:
Now, run `python test/onnx/test_operators.py --no-onnx`, we won't introduce any onnx python dependence. (No onnx/protobuf python packages needs to be installed)
The major changes:
- output pbtxt from C++ exporter directly, so the floating format may be slightly different. (This should be fine, since it's just to guard ONNX exporting.)
- ONNX python packages are only imported if we run the ONNX related checks. Those checks are disabled when using `--no-onnx` flag.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/10151
Reviewed By: jamesr66a
Differential Revision: D9130706
Pulled By: houseroad
fbshipit-source-id: ea28cf5db8399929179698ee535137f209e9ce6f
Summary: Minor fix for a bug introduced by D9004285
Reviewed By: anderspapitto
Differential Revision: D9028762
fbshipit-source-id: 9b9c5eef30e61d7ae19784e0418fa29bad2b5564
Summary:
Based on top of #9763 (first 3 commits belong to that PR). The first commits from this PR are "Stop using attributes ..."
I tried to separate the changes into fairly meaningful commits. I can't split them up into smaller PRs, because everything starts working and all tests pass only after the whole sequence, but hopefully this will make reviewing somewhat easier.
Known issues/regressions/future tasks:
- `aten::lerp` and `aten::clamp` are no longer fusable
- `CreateAutodiffSubgraphs` needs a rewrite
- It is much more strict now, and will miss a lot of opportunities, especially when viewing ops are involved. Our previous approach was "ignore the assumption on shape availability in gradient formulas to determine differentiability, and hope that shape prop will be robust enough to actually deliver them before we differentiate", which obviously doesn't scale well to more complex cases. We should either work on reducing the size dependency of grad formulas (feasible e.g. for `view`/`reshape`, unfeasible for `squeeze`/`unsqueeze`), or make `CreateAutodiffSubgraphs` integrate some kind of "I could integrate this node into an AD subgraph, but will I be able to infer the shape of its input" reasoning (kind of like a limited shape prop, that doesn't infer anything, and only tells if it *could* infer something).
- It sometimes creates constant-only (or constants + one node) graphs, which is useless
- Broken `aten::add` in auto-batching, because it gained a non-tensor input. I changed the test for pointwise operations to use `aten::mul` instead, but I needed to disable the LSTM cell test. I'm not sure how scalar constants should be implemented in this case, because I don't fully understand our format. cc: ChunliF
- Graph import does some hacks to recover type of constants. This code should be removed once we'll gain the ability to export the IR along with value types.
- There's still a fair amount of dead code that can be removed. I didn't want to make this diff any bigger, and removing it is an easy task.
- Graph fuser could be improved to use signature matching (possibly using `OperatorSet`) instead of basing on node kinds.
- Manual constant propagation for the `ListConstruct` node in `torch/onnx/utils.py` should be replaced with a proper constant propagation pass (or we should ensure that the one we have handles at least this case before we remove this code).
zdevito
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9807
Reviewed By: ezyang
Differential Revision: D9004285
Pulled By: apaszke
fbshipit-source-id: fe88026a765f6b687354add034c86402362508b7
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9718
This patch switches the interpreter to use IValue's primitive numbers rather than tensors for computing on integers and floats. In addition to preparing the interpreter for first-class support of other types, this cleans up the handling of primitive numbers, making it possible to just use the normal operator overloading dispatch to find the right implementation for numbers. As a result of this change, a lot of other functionality needed to be updated since it was the first time we use non-tensors in a lot of places in the code base.
Notes:
* Fixes code_template.py so that multi-line strings are indented correctly when used on a standalone line
* Cast operators (`int(x)`) now are functional. Some tests have addition conversions to integers because
we no longer allow implicit tensor -> integer conversions following the same convention as in python
* prim::ListConstruct/createList has been added to the interpreter for creating lists and this has
replaced aten::stack for integers lists
* gen_jit_dispatch.py has been refactored so that non-tensor types use operators on IValues to extract
the primitives
* IValue gains a .to<T> method that is the equivalent of tensor_as but for IValue instead of at::Tensor
* `constant_as<T>` is switched over to using IValues's `.to<T>` method, to make conversion from constant->IValue->C++ type
more consistent. This functionality combined with `toIValue(Value*)` replaces the `tensor_as` and `as_tensor` family of functions.
* conditional expressions (if, loop) and operators related to them are now computed on integers rather than tensors
* IValue gains constructors for constructing from at::Scalar and converting to it. However, IValue itself will always store
the scalars as a double or int64.
* To align with python 3 syntax, TK_INT, TK_FLOAT, and TK_BOOL have been removed from the parser, and int/float/bool are just treated as special identifiers in the compiler,
along with print. These are represented as special sugared values with a `call` method implemented. For int/float/bool this implements casting behavior.
* Dropped shared_from_this from Type/Module. They were not needed and they making debugging harder because they internally throw/catch exceptions.
* Shape propagation has been updated to support running nodes that include floating point primitive types, this required some refactoring of internal functions.
* TensorToNum and NumToTensor have actual implementations as operators now
* regster_prim_ops now contains implementations of math operators for float/int primitive types, and for mixed (prim <+> tensor) versions. This removes the need for special handling in compiler.cpp
* Primitive math is now entirely handled by letting the compiler choose the right overloads. This removes tons of special casing in the compiler.
* incorporates eellison's change to allow casting from return values. Due to the addition of primitive support, the code need slight modifications, so I just pre-merged it here.
* stack.h gains generic vararg versions of push/pop that know how to convert to/from C++ types:
```
at::Tensor a;
at::Scalar b;
pop(stack, a, b);
at::Tensor c = a + b;
push(stack, c);
```
apaszke
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9584
Reviewed By: apaszke
Differential Revision: D8910546
Pulled By: zdevito
fbshipit-source-id: 0f3e60d4d22217f196a8f606549430e43b7e7e30
Addresses #8177
A design doc can be found here: [gist](https://gist.github.com/zou3519/4b7f13f03cc9f3612bd9363e6405fa0a) version or [quip](https://fb.quip.com/azL1AqUckBdo) version
General approach:
- Add NumberType, FloatType, IntType to represent Python numbers, floats and ints.
- Emit these types for python literals
- Change aten_schema such that Scalars are NumberType, int64_t and bool are IntType.
- Emit aten::type_as, prim::NumToTensor, and prim::TensorToNum nodes for tensor-number math. (see examples below)
- Erase NumberType, prim::NumToTensor, and prim::TensorToNum for ONNX export
### Tensor/number math
```
import torch
@torch.jit.script
def fn(x):
return x + 1
```
```
graph(%x : Dynamic) {
%1 : int = prim::Constant[value={1}]()
%2 : Dynamic = prim::NumToTensor(%1)
%3 : Dynamic = aten::type_as(%2, %x)
%4 : Dynamic = aten::add[alpha={1}](%x, %4)
return (%5);
}
```
### Number/Number Math
```
import torch
@torch.jit.script
def fn(zero):
c = 1 + 1
return zero + c
```
```
graph(%zero : Dynamic) {
%1 : int = prim::Constant[value={1}]()
%2 : int = prim::Constant[value={1}]()
%3 : Dynamic = prim::num_to_tensor(%1)
%4 : Dynamic = prim::num_to_tensor(%2)
%5 : Dynamic = aten::add[alpha={1}](%3, %4)
%c : int = prim::TensorToNum(%6) # this is the result of the addition
...
return (%13);
}
```
List of squashed commits:
* Introduce Python Number types
Added: IntType, FloatType, NumberType with
IntType <: NumberType
FloatType <: NumberType
Changed aten_schema so arguments have corresponding types
* Emit a NumberType for python literals.
Also emit a NumberType for Scalar default values.
* Add prim::NumToTensor and prim::TensorToNum
* Add DynamicType -> NumberType implicit cast for bc
* Better ensureTensor error message
* Add ensureTensorOrNumber. Allow passing Number to some functions
Like the range() construct and slices
* Patch IntList to work.
IntList is still a DynamicType in the frontend: a tensor gets built from
a List[int].
Also, IntList[1] is a "union between int and IntList" the way it is
implemented. If the frontend sees an int being passed for an IntList[1]
arg, it converts it to a tensor as well.
* Enforce some order on schemas to avoid overload ambiguity
add(Tensor, Tensor) should appear earlier than add(Tensor, Scalar). This
matches the order in which python_arg_parser parses its arguments.
* Disable std_dim and var_dim tests.
With the new schema information, std(input, keepdim) and std(input, dim)
are ambiguous. This will need to be fixed at a later date.
* Add NumberType erasure pass.
This is used for ONNX export and to ensure that NumberType information
doesn't reach the interpreter
* Add support for mixed tensor/number math ops.
* Tests for new functionality.
Includes:
- Tensor/number math
- number/number math
- EraseNumberTypes pass test
* Patch tests
Update expect tests for:
- decompose_addmm
- loop unrolling tests
Because python numbers are now NumberType, they cannot be returned by
functions anymore. Work around this by using "torch.full", or by adding
a tensor([0]) (taken from FIXME_zerol()). Both approaches are used
because torch.full is more readable, but it is broken in some cases.
* Add erase_number_types to torch/CMakeLists.txt
* Move math back to emitSimpleExpr from emitSugaredExpr
* Remove some dead lines
* Renable some excluded script/trace tests that are fixed.
* Move some tests to expected failure
* Address some comments (more addressing to come)
* Erase relevant aten::type_as nodes in EraseNumberTypes
I also changed it so that EraseNumberTypes is only called for ONNX
export. It is no longer used to prevent
prim::NumToTensor/prim::TensorToNum from reaching shape_analysis or
interpreter.cpp.
shape_analysis infers the type of the output of these nodes to be the
same as their input.
intepreter.cpp treats both of these nodes as no-ops.
* Add reminder to fix std/var
* Call EraseNumberTypes only when exporting a script module
* Update expects after rebase
* [ONNX] Allow specifying only a subset of input/output names
Then we can only specify the "real" names while ignoring the names for all the parameters
* fix
* Update utils.py
