Merge branch 'master' into krovatkin/cpp_export

Author: brianjo

.jenkins/build.sh

@@ -16,13 +16,18 @@ rm -rf src
 pip install -r $DIR/../requirements.txt
 
 export PATH=/opt/conda/bin:$PATH
-conda install -y sphinx==1.8.2 pandas
+pip install sphinx==1.8.2 pandas
+
+# install awscli
+# pip uninstall awscli
+# pip install awscli==1.16.35
+
 # PyTorch Theme
 rm -rf src
 pip install -e git+git://github.com/pytorch/pytorch_sphinx_theme.git#egg=pytorch_sphinx_theme
 # pillow >= 4.2 will throw error when trying to write mode RGBA as JPEG,
 # this is a workaround to the issue.
-pip install sphinx-gallery tqdm matplotlib ipython pillow==4.1.1
+pip install sphinx-gallery==0.3.1 tqdm matplotlib ipython pillow==4.1.1
 
 # Install torchaudio from source
 git clone https://github.com/pytorch/audio --quiet
@@ -36,9 +41,9 @@ aws configure set default.s3.multipart_threshold 5120MB
 export NUM_WORKERS=20
 if [[ "${JOB_BASE_NAME}" == *worker_* ]]; then
   # Step 1: Remove runnable code from tutorials that are not supposed to be run
-  python $DIR/remove_runnable_code.py beginner_source/aws_distributed_training_tutorial.py beginner_source/aws_distributed_training_tutorial.py
+  python $DIR/remove_runnable_code.py beginner_source/aws_distributed_training_tutorial.py beginner_source/aws_distributed_training_tutorial.py || true
   # TODO: Fix bugs in these tutorials to make them runnable again
-  python $DIR/remove_runnable_code.py beginner_source/audio_classifier_tutorial.py beginner_source/audio_classifier_tutorial.py
+  python $DIR/remove_runnable_code.py beginner_source/audio_classifier_tutorial.py beginner_source/audio_classifier_tutorial.py || true
 
   # Step 2: Keep certain tutorials based on file count, and remove runnable code in all other tutorials
   # IMPORTANT NOTE: We assume that each tutorial has a UNIQUE filename.

_static/imagenet_class_index.json

@@ -13,6 +13,6 @@ Advanced Tutorials
 	Custom C Extensions for PyTorch
 	https://pytorch.org/tutorials/advanced/c_extension.html
 
-4. super_resolution_with_caffe2.py
-	Transfering a Model from PyTorch to Caffe2 and Mobile using ONNX
-	https://pytorch.org/tutorials/advanced/super_resolution_with_caffe2.html
+4. super_resolution_with_onnxruntime.py
+    Exporting a Model from PyTorch to ONNX and Running it using ONNXRuntime
+	https://pytorch.org/tutorials/advanced/super_resolution_with_onnxruntime.html

_static/img/cat_output1.png

@@ -147,23 +147,22 @@ For the "ahead of time" flavor, we build our C++ extension by writing a
 ``setup.py`` script that uses setuptools to compile our C++ code. For the LLTM, it
 looks as simple as this::
 
-  from setuptools import setup
-  from torch.utils.cpp_extension import CppExtension, BuildExtension
+  from setuptools import setup, Extension
+  from torch.utils import cpp_extension
 
   setup(name='lltm_cpp',
-        ext_modules=[CppExtension('lltm', ['lltm.cpp'])],
-        cmdclass={'build_ext': BuildExtension})
-
+        ext_modules=[cpp_extension.CppExtension('lltm_cpp', ['lltm.cpp'])],
+        cmdclass={'build_ext': cpp_extension.BuildExtension})
 
 In this code, :class:`CppExtension` is a convenience wrapper around
 :class:`setuptools.Extension` that passes the correct include paths and sets
 the language of the extension to C++. The equivalent vanilla :mod:`setuptools`
 code would simply be::
 
-  setuptools.Extension(
+  Extension(
      name='lltm_cpp',
      sources=['lltm.cpp'],
-     include_dirs=torch.utils.cpp_extension.include_paths(),
+     include_dirs=cpp_extension.include_paths(),
      language='c++')
 
 :class:`BuildExtension` performs a number of required configuration steps and
@@ -413,7 +412,7 @@ see::
 If we call ``help()`` on the function or module, we can see that its signature
 matches our C++ code::
 
-  In[4] help(lltm.forward)
+  In[4] help(lltm_cpp.forward)
   forward(...) method of builtins.PyCapsule instance
       forward(arg0: torch::Tensor, arg1: torch::Tensor, arg2: torch::Tensor, arg3: torch::Tensor, arg4: torch::Tensor) -> List[torch::Tensor]
 
@@ -473,6 +472,8 @@ small benchmark to see how much performance we gained from rewriting our op in
 C++. We'll run the LLTM forwards and backwards a few times and measure the
 duration::
 
+  import time
+
   import torch
 
   batch_size = 16

_static/img/cat_superres_with_ort.jpg

@@ -35,13 +35,14 @@
 
 
 class BadFFTFunction(Function):
-
-    def forward(self, input):
+    @staticmethod
+    def forward(ctx, input):
         numpy_input = input.detach().numpy()
         result = abs(rfft2(numpy_input))
         return input.new(result)
 
-    def backward(self, grad_output):
+    @staticmethod
+    def backward(ctx, grad_output):
         numpy_go = grad_output.numpy()
         result = irfft2(numpy_go)
         return grad_output.new(result)
@@ -51,7 +52,7 @@ def backward(self, grad_output):
 
 
 def incorrect_fft(input):
-    return BadFFTFunction()(input)
+    return BadFFTFunction.apply(input)
 
 ###############################################################
 # **Example usage of the created layer:**