apply v1.10, pytorch/tutorials@2571e95 (#387)

PyTorchKorea · Nov 28, 2021 · 1e95c9c · 1e95c9c
1 parent 4f72864
commit 1e95c9c
Show file tree

Hide file tree

Showing 84 changed files with 6,783 additions and 2,441 deletions.
diff --git a/.github/ISSUE_TEMPLATE/1_TRANSLATE_REQUEST.md b/.github/ISSUE_TEMPLATE/1_TRANSLATE_REQUEST.md
@@ -15,4 +15,4 @@ _(반드시 지키셔야 하는 일정이 아닙니다 - 일정이 너무 늦어
 ## 관련 이슈
 _현재 번역 요청 / 진행 내역을 보기 위해 각 버전의 메인 이슈를 참조합니다._ <br />
 _(특별한 일이 없다면 변경하지 않으셔도 됩니다.)_
-* 관련 이슈: #221 (v1.9)
+* 관련 이슈: #387 (v1.10)
diff --git a/README.md b/README.md
@@ -4,7 +4,7 @@
 
 PyTorch에서 제공하는 튜토리얼의 한국어 번역을 위한 저장소입니다.\
 번역의 결과물은 [https://tutorials.pytorch.kr](https://tutorials.pytorch.kr)에서 확인하실 수 있습니다. (번역을 진행하며 **불규칙적으로** 업데이트합니다.)\
-현재 번역 진행 중인 내용은 [#221 이슈](https://github.com/9bow/PyTorch-tutorials-kr/issues/221)를 참고해주세요.
+현재 버전의 번역 / 변경 관련 이슈는 [#387 이슈](https://github.com/9bow/PyTorch-tutorials-kr/issues/387)를 참고해주세요.
 
 ## 기여하기
 
@@ -22,7 +22,7 @@ PyTorch에서 제공하는 튜토리얼의 한국어 번역을 위한 저장소
 
 ## 원문
 
-현재 PyTorch v1.9 튜토리얼 번역이 진행 중입니다. ([pytorch/tutorials@24ad87f](https://github.com/pytorch/tutorials/commit/24ad87fdb9e5cd73ed8a27eed1efed6c4a7a1bb5) 기준)
+현재 PyTorch v1.9 튜토리얼 번역이 진행 중입니다. ([pytorch/tutorials@2571e95](https://github.com/pytorch/tutorials/commit/2571e95df42b8ed46d11ac9827c637fa4e826dfe) 기준)
 
 최신 버전의 튜토리얼(공식, 영어)은 [PyTorch tutorials 사이트](https://pytorch.org/tutorials) 및 [PyTorch tutorials 저장소](https://github.com/pytorch/tutorials)를 참고해주세요.
 
@@ -31,7 +31,7 @@ PyTorch에서 제공하는 튜토리얼의 한국어 번역을 위한 저장소
 ### PyTorch v1.0 이상의 튜토리얼 보기
 
 v1.0 이후 번역은 별도 저장소로 관리하지 않습니다. [이 저장소의 Release 메뉴](https://github.com/9bow/PyTorch-tutorials-kr/releases)를 확인해주세요. \
-`버전-base`(예. `1.6-base`)는 해당 버전을 시작할 때의 릴리즈이고, `버전-latest`(예. `1.6-latest`)는 해당 버전의 마지막 릴리즈입니다.
+`버전-base`(예. `1.9-base`)는 해당 버전을 시작할 때의 릴리즈이고, `버전-latest`(예. `1.9-latest`)는 해당 버전의 마지막 릴리즈입니다.
 
 해당 릴리즈의 문서를 내려받으신 후 빌드하시면 해당 버전의 문서를 확인하실 수 있습니다. \
 빌드 방법은 [기여하기 문서의 `2-5. (내 컴퓨터에서) 결과 확인하기`](https://github.com/9bow/PyTorch-tutorials-kr/blob/master/CONTRIBUTING.md#2-5-내-컴퓨터에서-결과-확인하기) 부분을 참고해주세요.
@@ -46,5 +46,5 @@ v1.0 이후 번역은 별도 저장소로 관리하지 않습니다. [이 저장
   | 0.3.1  | [PyTorch-tutorials-kr-0.3.1](https://9bow.github.io/PyTorch-tutorials-kr-0.3.1) | [GitHub 저장소](https://github.com/9bow/PyTorch-tutorials-kr-0.3.1) |
 
 ---
-This is a project to translate [pytorch/tutorials@24ad87f](https://github.com/pytorch/tutorials/commit/24ad87fdb9e5cd73ed8a27eed1efed6c4a7a1bb5) into Korean.
+This is a project to translate [pytorch/tutorials@2571e95](https://github.com/pytorch/tutorials/commit/2571e95df42b8ed46d11ac9827c637fa4e826dfe) into Korean.
 For the latest version, please visit to the [official PyTorch tutorials repo](https://github.com/pytorch/tutorials).
diff --git a/_static/img/profiler_callstack.png b/_static/img/profiler_callstack.png
diff --git a/_static/img/profiler_kernel_view.png b/_static/img/profiler_kernel_view.png
diff --git a/_static/img/profiler_memory_curve_selecting.png b/_static/img/profiler_memory_curve_selecting.png
diff --git a/_static/img/profiler_memory_curve_single.png b/_static/img/profiler_memory_curve_single.png
diff --git a/_static/img/profiler_memory_view.png b/_static/img/profiler_memory_view.png
diff --git a/_static/img/profiler_operator_view.png b/_static/img/profiler_operator_view.png
diff --git a/_static/img/profiler_overview1.png b/_static/img/profiler_overview1.png
diff --git a/_static/img/profiler_overview2.png b/_static/img/profiler_overview2.png
diff --git a/_static/img/profiler_trace_view_fwd_bwd.png b/_static/img/profiler_trace_view_fwd_bwd.png
diff --git a/_static/img/profiler_views_list.png b/_static/img/profiler_views_list.png
diff --git a/_static/img/thumbnails/cropped/torchaudio-alignment.png b/_static/img/thumbnails/cropped/torchaudio-alignment.png
diff --git a/_static/img/thumbnails/cropped/torchaudio-asr.png b/_static/img/thumbnails/cropped/torchaudio-asr.png
diff --git a/_static/img/tts_pipeline.png b/_static/img/tts_pipeline.png
diff --git a/advanced_source/dispatcher.rst b/advanced_source/dispatcher.rst
@@ -105,6 +105,57 @@ speaking, the structure of your registrations will look like this:
     that provides implementations for all basic operators on the XLA dispatch
     key.
 
+
+For operators that do not need autograd
+---------------------------------------
+
+Note: This section only applies to versions of PyTorch ``>= 1.10``.
+
+In the next section, we will discuss how to add autograd support to an operator.
+But for the ops that do not need autograd support, the following kernel should be
+registered improve useability and make your op behave like PyTorch's built-in
+operators.
+
+.. code-block:: cpp
+
+  TORCH_LIBRARY_IMPL(myops, Autograd, m) {
+    m.impl(op, autogradNotImplementedFallback());
+  }
+
+The above lines registers an ``Autograd`` kernel that appends a dummy
+``NotImplemented`` node on forward (preserving the ``require_grad``-ness of the inputs).
+On backward, the ``NotImplemented`` node raises an error. This can be helpful
+for debugging in larger models where previously it can be hard to pin-point
+exactly where the ``requires_grad``-ness is lost during the forward pass.
+
+In-place or view ops
+^^^^^^^^^^^^^^^^^^^
+
+To ensure correctness and best possible performance, if your op mutates an input
+in-place or returns a tensor that aliases with one of the inputs, two additional
+steps should be taken:
+
+1. Register an ``ADInplaceOrView`` kernel in addition to the ``Autograd`` kernel
+   above. This kernel handles the necessary bookkeeping to ensure the correctness
+   of in-place or view operations. It is important to note that this ADInplaceOrView
+   kernel should only be used with ``autogradNotImplementedFallback``.
+
+.. code-block:: cpp
+
+  TORCH_LIBRARY_IMPL(myops, Autograd, m) {
+    m.impl(op, autogradNotImplementedFallback());
+  }
+  TORCH_LIBRARY_IMPL(myops, ADInplaceOrView, m) {
+    m.impl(op, autogradNotImplementedInplaceOrViewFallback());
+  }
+
+2. The ``Autograd`` or ``ADInplaceOrView`` boxed kernels registered above
+   rely on operator schema information in their logi. If your op mutates an input
+   in-place or returns a tensor that aliases with one of the inputs it is important to
+   ensure that your schema properly reflects this. See
+   `here <https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/README.md>`_
+   for more information on how to annotate the schema.
+
 .. _autograd-support:
 
 Adding autograd support