NMS CPU extension does not compile with pytorch 1.0.1 / nightly

[miguelvr]

🐛 Bug

A compilation error is thrown when building with the latest version of pytorch or pytorch-nightly

To Reproduce

nvidia-docker build -t maskrcnn-benchmark docker/

Expected behavior

maskrcnn-benchmark should build correctly

Environment

• Docker

Additional context

building 'maskrcnn_benchmark._C' extension
creating build/temp.linux-x86_64-3.6
creating build/temp.linux-x86_64-3.6/maskrcnn-benchmark
creating build/temp.linux-x86_64-3.6/maskrcnn-benchmark/maskrcnn_benchmark
creating build/temp.linux-x86_64-3.6/maskrcnn-benchmark/maskrcnn_benchmark/csrc
creating build/temp.linux-x86_64-3.6/maskrcnn-benchmark/maskrcnn_benchmark/csrc/cpu
creating build/temp.linux-x86_64-3.6/maskrcnn-benchmark/maskrcnn_benchmark/csrc/cuda
gcc -pthread -B /miniconda/envs/py36/compiler_compat -Wl,--sysroot=/ -Wsign-compare -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -DWITH_CUDA -I/maskrcnn-benchmark/maskrcnn_benchmark/csrc -I/miniconda/envs/py36/lib/python3.6/site-packages/torch-1.0.1.post2-py3.6-linux-x86_64.egg/torch/lib/include -I/miniconda/envs/py36/lib/python3.6/site-packages/torch-1.0.1.post2-py3.6-linux-x86_64.egg/torch/lib/include/torch/csrc/api/include -I/miniconda/envs/py36/lib/python3.6/site-packages/torch-1.0.1.post2-py3.6-linux-x86_64.egg/torch/lib/include/TH -I/miniconda/envs/py36/lib/python3.6/site-packages/torch-1.0.1.post2-py3.6-linux-x86_64.egg/torch/lib/include/THC -I/usr/local/cuda/include -I/miniconda/envs/py36/include/python3.6m -c /maskrcnn-benchmark/maskrcnn_benchmark/csrc/vision.cpp -o build/temp.linux-x86_64-3.6/maskrcnn-benchmark/maskrcnn_benchmark/csrc/vision.o -DTORCH_API_INCLUDE_EXTENSION_H -DTORCH_EXTENSION_NAME=_C -D_GLIBCXX_USE_CXX11_ABI=0 -std=c++11
cc1plus: warning: command line option '-Wstrict-prototypes' is valid for C/ObjC but not for C++
gcc -pthread -B /miniconda/envs/py36/compiler_compat -Wl,--sysroot=/ -Wsign-compare -DNDEBUG -g -fwrapv -O3 -Wall -Wstrict-prototypes -fPIC -DWITH_CUDA -I/maskrcnn-benchmark/maskrcnn_benchmark/csrc -I/miniconda/envs/py36/lib/python3.6/site-packages/torch-1.0.1.post2-py3.6-linux-x86_64.egg/torch/lib/include -I/miniconda/envs/py36/lib/python3.6/site-packages/torch-1.0.1.post2-py3.6-linux-x86_64.egg/torch/lib/include/torch/csrc/api/include -I/miniconda/envs/py36/lib/python3.6/site-packages/torch-1.0.1.post2-py3.6-linux-x86_64.egg/torch/lib/include/TH -I/miniconda/envs/py36/lib/python3.6/site-packages/torch-1.0.1.post2-py3.6-linux-x86_64.egg/torch/lib/include/THC -I/usr/local/cuda/include -I/miniconda/envs/py36/include/python3.6m -c /maskrcnn-benchmark/maskrcnn_benchmark/csrc/cpu/nms_cpu.cpp -o build/temp.linux-x86_64-3.6/maskrcnn-benchmark/maskrcnn_benchmark/csrc/cpu/nms_cpu.o -DTORCH_API_INCLUDE_EXTENSION_H -DTORCH_EXTENSION_NAME=_C -D_GLIBCXX_USE_CXX11_ABI=0 -std=c++11
cc1plus: warning: command line option '-Wstrict-prototypes' is valid for C/ObjC but not for C++
In file included from /miniconda/envs/py36/lib/python3.6/site-packages/torch-1.0.1.post2-py3.6-linux-x86_64.egg/torch/lib/include/ATen/ATen.h:9:0,
                 from /miniconda/envs/py36/lib/python3.6/site-packages/torch-1.0.1.post2-py3.6-linux-x86_64.egg/torch/lib/include/torch/csrc/api/include/torch/types.h:3,
                 from /miniconda/envs/py36/lib/python3.6/site-packages/torch-1.0.1.post2-py3.6-linux-x86_64.egg/torch/lib/include/torch/csrc/api/include/torch/data/dataloader_options.h:4,
                 from /miniconda/envs/py36/lib/python3.6/site-packages/torch-1.0.1.post2-py3.6-linux-x86_64.egg/torch/lib/include/torch/csrc/api/include/torch/data/dataloader.h:3,
                 from /miniconda/envs/py36/lib/python3.6/site-packages/torch-1.0.1.post2-py3.6-linux-x86_64.egg/torch/lib/include/torch/csrc/api/include/torch/data.h:3,
                 from /miniconda/envs/py36/lib/python3.6/site-packages/torch-1.0.1.post2-py3.6-linux-x86_64.egg/torch/lib/include/torch/csrc/api/include/torch/all.h:4,
                 from /miniconda/envs/py36/lib/python3.6/site-packages/torch-1.0.1.post2-py3.6-linux-x86_64.egg/torch/lib/include/torch/extension.h:4,
                 from /maskrcnn-benchmark/maskrcnn_benchmark/csrc/cpu/vision.h:3,
                 from /maskrcnn-benchmark/maskrcnn_benchmark/csrc/cpu/nms_cpu.cpp:2:
/maskrcnn-benchmark/maskrcnn_benchmark/csrc/cpu/nms_cpu.cpp: In lambda function:
/maskrcnn-benchmark/maskrcnn_benchmark/csrc/cpu/nms_cpu.cpp:71:46: error: invalid initialization of reference of type 'const at::Type&' from expression of type 'c10::ScalarType'
   AT_DISPATCH_FLOATING_TYPES(dets.scalar_type(), "nms", [&] {
                                              ^
/miniconda/envs/py36/lib/python3.6/site-packages/torch-1.0.1.post2-py3.6-linux-x86_64.egg/torch/lib/include/ATen/Dispatch.h:15:32: note: in definition of macro 'AT_DISPATCH_FLOATING_TYPES'
     const at::Type& the_type = TYPE;                                         \
                                ^
error: command 'gcc' failed with exit status 1

[LeviViana]

Sorry, I read too quickly. The torch-1.0.1.post2 version is too old. The nightly version should be newer than the PyTorch's PR that brought this change.

[miguelvr]

isn't it a better solution to fix the NMS extension with the correct type?

[miguelvr]

@LeviViana the pytorch nightly from conda is not working as well

[fmassa]

@miguelvr there was a breaking change in PyTorch that made the extensions only work on pre 1.0.1 or on the nightly. There was no way of making it work for both.

This has now been fixed in PyTorch, so I believe #555 could be reverted.

Can you try locally reverting this patch and see if it works with PyTorch nightly?

[miguelvr]

@fmassa will do! thanks