Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9905
This diff improves lars operator in Caffe2 by applying clipping to the computed learning rate
Reviewed By: pjh5
Differential Revision: D9020606
fbshipit-source-id: b579f1d628113c09366feac9406002f1ef4bd54f
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/10035
This is an initial diff which refactors some of the components in the Seq2SeqModelCaffe2EnsembleDecoder class.
Reviewed By: jmp84
Differential Revision: D9026372
fbshipit-source-id: 449635208f24494209ae2fb78a19fca872970ea8
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/10044
The test was subtly broken! This transform wasn't writing to the correct blob and the test did not catch that because it was looking at the old version.
thanks @[100022211048576:kerenzhou] for catching this
Reviewed By: Jokeren
Differential Revision: D9075520
fbshipit-source-id: c31ff0afcd78dd2dc7ffc240e2e89eeda87f1fb4
Summary:
The goal of this PR is to update the hip files to reflect relevant changes in cuda source files.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9826
Differential Revision: D9032840
Pulled By: bddppq
fbshipit-source-id: 504e55c46308eebfee3c9a7beea1f294fe03470f
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9747
Currently the ctc_greedy_decoder op initializes the `merge_repeated` argument only if it has been provided by the user. Change to initialize in all cases.
Reviewed By: houseroad
Differential Revision: D8963635
fbshipit-source-id: 18955c7c26a77d9d7f5137e4dec085252ffabfeb
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9939
Pull Request resolved: https://github.com/facebookresearch/weakly-supervised-action-detection/pull/13
Pull Request resolved: https://github.com/pytorch/translate/pull/166
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9125
Closes https://github.com/pytorch/pytorch/pull/9125
Use inheritance for polymorphism, and remove template parameter
This is to change the templating in call sites, the core implementations will change later
Before Caffe2 Tensor class was compile-time fixed to bind to a particular device/context. With this change, we're making it a runtime property (stored inside the tensor), but preserve the same semantics. For example, one has to specify device type in order to create a Tensor - there are no uninitialized tensors. More specifically the changes are:
1. We added an extra argument *DeviceType* to most of the constructors of the tensor, e.g. (Tensor(DeviceType type)),
2. Semantics of constructor Tensor(const Tensor<SrcContext>& src, ContextForCopy* context); is changed, in this constructor, the second context is passed in to enable us to call the templated Copy function, it could be in a different context as source and target previously, now we'll enforce that the context should have same device type as src, if it is provided.
3. To preserve 'get-or-construct' semantics of Blob, we added specialized getter Blob::GetMutableTensor that verifies both that Blob contains a Tensor and that it's of a correct type
4. Specifically, Tensor type is not default-constructible any more (as we don't have unknown device tensors) and thus some of the code handling STL containers needs to change
Note: Some changes are postponed just to keep this diff a bit smaller. Please see `TODO`s.
Reviewed By: ezyang, houseroad
Differential Revision: D9024330
fbshipit-source-id: e0b8295d2dc6ebe2963383ded5af799ad17164ba
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9581
Mostly to simplify code. Should also improve performance but order switch ops
don't take much time anyway.
Reviewed By: viswanathgs
Differential Revision: D8909766
fbshipit-source-id: 17a302d5bf4aba2755d88223fc01a41fd72c5919
Summary:
And let Gemm conversion to inspect the input `C` to try converting to FC.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9870
Reviewed By: houseroad
Differential Revision: D9013198
Pulled By: bddppq
fbshipit-source-id: b4c509cfccca238262e1c406b004e66cef256321
Summary:
Pull Request resolved: https://github.com/facebookresearch/weakly-supervised-action-detection/pull/13
Pull Request resolved: https://github.com/pytorch/translate/pull/166
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9125
Closes https://github.com/pytorch/pytorch/pull/9125
Use inheritance for polymorphism, and remove template parameter
This is to change the templating in call sites, the core implementations will change later
Before Caffe2 Tensor class was compile-time fixed to bind to a particular device/context. With this change, we're making it a runtime property (stored inside the tensor), but preserve the same semantics. For example, one has to specify device type in order to create a Tensor - there are no uninitialized tensors. More specifically the changes are:
1. We added an extra argument *DeviceType* to most of the constructors of the tensor, e.g. (Tensor(DeviceType type)),
2. Semantics of constructor Tensor(const Tensor<SrcContext>& src, ContextForCopy* context); is changed, in this constructor, the second context is passed in to enable us to call the templated Copy function, it could be in a different context as source and target previously, now we'll enforce that the context should have same device type as src, if it is provided.
3. To preserve 'get-or-construct' semantics of Blob, we added specialized getter Blob::GetMutableTensor that verifies both that Blob contains a Tensor and that it's of a correct type
4. Specifically, Tensor type is not default-constructible any more (as we don't have unknown device tensors) and thus some of the code handling STL containers needs to change
Note: Some changes are postponed just to keep this diff a bit smaller. Please see `TODO`s.
Reviewed By: xw285cornell
Differential Revision: D8121878
fbshipit-source-id: 4a5e9a677ba4ac82095df959851a054c81eccf81
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9643
Current map interface assumes float data type, which is not always correct.
Reviewed By: kennyhorror
Differential Revision: D8455784
fbshipit-source-id: b94a31267760f7f97c15aa4b03008affc347fd10
Summary:
Not in the same format. Skip at the moment.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9751
Reviewed By: yinghai
Differential Revision: D8965636
Pulled By: houseroad
fbshipit-source-id: 81d39c2f5625c14c0e1ee11408b5f7267b53798f
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9667
MKL-DNN doesn't support 64-bit integger (cfee61bf81/include/mkldnn_types.h (L62-L75)). So force converting from `TensorCPU<long>` to `s32` Ideep tensor will cause memory issue. This diff gives an alternative solution, where we just fall through to TensorCPU. The reasoning is that since MKL-DNN doesn't support 64 bit integer tensor, downstream ops have to be in CPUConext. So there is no reason force converting to ideep tensor and back.
Reviewed By: pjh5
Differential Revision: D8943544
fbshipit-source-id: f514903cda27e34b8887271c9df56c8220895116
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9665
In data_parallel_model, we isolate synchronizing barrier init net into its own from the param_init_net, so that we could have finer granularity of control over the barrier net.
Reviewed By: andrewwdye
Differential Revision: D8375389
fbshipit-source-id: ce0c8c1c8e4bd82b7078a1b07abaced3f149d578
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9637
Adding a method to run plan in background. The intended use is to run BlueWhale's data reading & preprocessing net in background while the GPU is training.
Reviewed By: MisterTea
Differential Revision: D8906439
fbshipit-source-id: b1c73ca7327e2d87a8f873924e05ab3d161a3f1e
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9636
Make sure that the blobs are registered to the net
Reviewed By: pjh5
Differential Revision: D8924883
fbshipit-source-id: f09422a2d4d5ba8bf6cfbfd00172097b5ab1fcd6
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9598
The "max_length" should be passed to UnPackSegmentsOp if "max_length" is given when calling PackSegmentsOp.
Reviewed By: jerryzh168
Differential Revision: D8919799
fbshipit-source-id: 8c97aa717b69177b8a5d5d56892817d488853840
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9594
When the input vector is a zero vector, the previous GPU code will give Nan in backward. We fix this.
Reviewed By: pjh5
Differential Revision: D8849732
fbshipit-source-id: 87b1fb1ee05dfdb0d43bcbe67e36f15896fe1706
Summary:
The goal of this PR is to add an infrastructure; to convert(hipify) CUDA ops into [HIP](https://github.com/ROCm-Developer-Tools/HIP) ops , at **compile** time.
Note that HIP ops, which are portable c++ code, can run on AMD and NVIDIA platform.
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9322
Differential Revision: D8884707
Pulled By: bddppq
fbshipit-source-id: dabc6319546002c308c10528238e6684f7aef0f8
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9438
Current implementation of create_from_proto doesn't work as expected: it
duplicates networks and execution steps by copying original PlanDef first and
adding each step one-by-one later.
Reviewed By: pjh5
Differential Revision: D8850316
fbshipit-source-id: 9b02836d6e6ee1c91cfdd3b4c4804f14137dc22b
Summary:
Pull Request resolved: https://github.com/pytorch/pytorch/pull/9458
The goal is to support count_include_pad in Caffe2 ONNX backend. This commit contains the first step - support 4-D tensor cases.
AveragePool with count_include_pad can be expressed as PadImage + AveragePool.
Reviewed By: houseroad
Differential Revision: D8852180
fbshipit-source-id: 4db00e9771be7a000a2d92850dfd066d9c9c38bf
