Fix for derivative of sinc(x) when x is positive but very very small #56986

kbrose · 2021-04-27T03:49:28Z

Problem arises for sinc'(x) where x != 0, but x ** 2 == 0, which happens for some very small floats.

I realized that my solution from #56763 was incomplete when I did a quick implementation using torch.autograd.Function and still got a NaN from my derivative.

facebook-github-bot · 2021-04-27T03:49:34Z

💊 CI failures summary and remediations

As of commit 0f4269a (more details on the Dr. CI page):

2/2 failures possibly* introduced in this PR
- 1/2 non-scanned failure(s)

🕵️ 1 new failure recognized by patterns

The following CI failures do not appear to be due to upstream breakages:

binary_linux_libtorch_3_7m_cpu_gcc5_4_cxx11-abi_shared-with-deps_build (1/1)

Step: "Build" (full log | diagnosis details | 🔁 rerun)

FAILED: lib/libtorch_cpu.so

[5077/5411] Building CXX object test_cpp_rpc/CMakeFiles/test_cpp_rpc.dir/test_e2e_process_group.cpp.o
[5078/5411] Building CXX object test_api/CMakeFiles/test_api.dir/fft.cpp.o
[5079/5411] Building CXX object test_cpp_rpc/CMakeFiles/test_cpp_rpc.dir/test_e2e_tensorpipe.cpp.o
[5080/5411] Building CXX object test_api/CMakeFiles/test_api.dir/misc.cpp.o
[5081/5411] Building CXX object test_tensorexpr/CMakeFiles/test_tensorexpr.dir/test_memdependency.cpp.o
[5082/5411] Building CXX object test_api/CMakeFiles/test_api.dir/support.cpp.o
[5083/5411] Building CXX object dist_autograd/CMakeFiles/test_dist_autograd.dir/__/common/main.cpp.o
[5084/5411] Building CXX object test_api/CMakeFiles/test_api.dir/any.cpp.o
[5085/5411] Building CXX object test_api/CMakeFiles/test_api.dir/tensor_cuda.cpp.o
[5086/5411] Linking CXX shared library lib/libtorch_cpu.so
FAILED: lib/libtorch_cpu.so 
: && /usr/bin/c++ -fPIC -Wno-deprecated-declarations -D_GLIBCXX_USE_CXX11_ABI=1 -Wno-deprecated -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -fopenmp -DUSE_KINETO -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -O2 -fPIC -Wno-narrowing -Wall -Wextra -Werror=return-type -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wno-sign-compare -Wno-unused-parameter -Wno-unused-variable -Wno-unused-function -Wno-unused-result -Wno-unused-local-typedefs -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-psabi -Wno-error=pedantic -Wno-error=redundant-decls -Wno-error=old-style-cast -fdiagnostics-color=always -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -DHAVE_AVX_CPU_DEFINITION -DHAVE_AVX2_CPU_DEFINITION -O2 -g -DNDEBUG  -Wl,--no-as-needed -rdynamic -shared -Wl,-soname,libtorch_cpu.so -o lib/libtorch_cpu.so caffe2/quantization/server/CMakeFiles/caffe2_dnnlowp_avx2_ops.dir/elementwise_sum_dnnlowp_op_avx2.cc.o caffe2/quantization/server/CMakeFiles/caffe2_dnnlowp_avx2_ops.dir/fully_connected_fake_lowp_op_avx2.cc.o caffe2/quantization/server/CMakeFiles/caffe2_dnnlowp_avx2_ops.dir/group_norm_dnnlowp_op_avx2.cc.o caffe2/quantization/server/CMakeFiles/caffe2_dnnlowp_avx2_ops.dir/pool_dnnlowp_op_avx2.cc.o caffe2/quantization/server/CMakeFiles/caffe2_dnnlowp_avx2_ops.dir/relu_dnnlowp_op_avx2.cc.o caffe2/quantization/server/CMakeFiles/caffe2_dnnlowp_avx2_ops.dir/spatial_batch_norm_dnnlowp_op_avx2.cc.o caffe2/quantization/server/CMakeFiles/caffe2_dnnlowp_avx2_ops.dir/transpose.cc.o caffe2/quantization/server/CMakeFiles/caffe2_dnnlowp_avx2_ops.dir/norm_minimization_avx2.cc.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/BatchedFallback.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/BatchedTensorImpl.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/BatchingRegistrations.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/CPUGeneratorImpl.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/Context.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/DLConvertor.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/DynamicLibrary.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/ExpandUtils.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/LegacyTHFunctionsCPU.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/MemoryOverlap.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/NamedTensorUtils.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/ParallelCommon.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/ParallelNative.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/ParallelNativeTBB.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/ParallelOpenMP.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/ParallelThreadPoolNative.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/ScalarOps.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/SequenceNumber.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/SparseCsrTensorImpl.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/SparseTensorImpl.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/SparseTensorUtils.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/TensorGeometry.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/TensorIndexing.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/TensorIterator.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/TensorMeta.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/TensorNames.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/TensorUtils.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/ThreadLocalState.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/Utils.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/Version.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/VmapMode.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/VmapModeRegistrations.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/VmapTransforms.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/autocast_mode.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/cpu/FlushDenormal.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/detail/CPUGuardImpl.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/detail/CUDAHooksInterface.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/detail/HIPHooksInterface.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/detail/MetaGuardImpl.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/record_function.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/core/ATenGeneral.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/core/BackendSelectFallbackKernel.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/core/DeprecatedTypeProperties.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/core/DeprecatedTypePropertiesRegistry.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/aten/src/ATen/core/Dict.cpp.o caffe2/CMakeFiles/torch_cpu.dir/__/at
collect2: fatal error: ld terminated with signal 9 [Killed]
compilation terminated.
[5087/5411] Building CXX object test_api/CMakeFiles/test_api.dir/parameterlist.cpp.o
[5088/5411] Building CXX object test_api/CMakeFiles/test_api.dir/parameterdict.cpp.o
[5089/5411] Building CXX object caffe2/lib_c10d/CMakeFiles/c10d.dir/ParamCommsUtils.cpp.o
[5090/5411] Building CXX object test_api/CMakeFiles/test_api.dir/integration.cpp.o
[5091/5411] Building CXX object test_api/CMakeFiles/test_api.dir/moduledict.cpp.o
[5092/5411] Building CXX object test_api/CMakeFiles/test_api.dir/ordered_dict.cpp.o
[5093/5411] Building CXX object test_api/CMakeFiles/test_api.dir/modulelist.cpp.o

This comment was automatically generated by Dr. CI (expand for details).

Follow this link to opt-out of these comments for your Pull Requests.

Please report bugs/suggestions to the (internal) Dr. CI Users group.

Click here to manually regenerate this comment.

albanD

Thanks for the update.
That looks good!

albanD · 2021-04-27T17:05:35Z

Could you rebase on top of master to properly trigger the CI please?

Problem arises for sinc'(x) where x > 0, but x ** 2 == 0 (which happens for some very small floats).

kbrose · 2021-04-27T17:31:52Z

Rebased

kbrose · 2021-04-28T00:13:06Z

@albanD seems like all the tests still didn't run. Let me know if I'm doing something wrong.

albanD · 2021-04-28T22:18:44Z

Let me look into it.

kbrose · 2021-04-29T03:38:15Z

I can also open a new PR from a clean branch if that's easier than debugging CI.

albanD · 2021-04-29T13:02:18Z

No need, this is actually because we're getting rate limited by circle CI.

albanD · 2021-04-29T13:06:04Z

Ok, opening another PR with the same commit did the trick :D Should be good now.

kbrose · 2021-04-29T14:00:36Z

Apologies, I'm a little confused. Should I open a new MR?

albanD · 2021-04-29T14:05:49Z

No need to do anything.
You should see the CI in this PR now.

Note that the libtorch fail seems unrelated.

facebook-github-bot · 2021-04-29T15:22:37Z

@albanD has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

kbrose · 2021-04-29T15:58:51Z

Thanks for keeping this moving along @albanD, I appreciate it!

facebook-github-bot · 2021-04-29T18:17:48Z

@albanD merged this pull request in ec86f96.

…ytorch#56986) Summary: Problem arises for sinc'(x) where x != 0, but x ** 2 == 0, which happens for some very small floats. I realized that my solution from pytorch#56763 was incomplete when I did a quick implementation using `torch.autograd.Function` and still got a `NaN` from my derivative. Pull Request resolved: pytorch#56986 Reviewed By: gchanan Differential Revision: D28093507 Pulled By: albanD fbshipit-source-id: 2a30e1065b08c5c60de843a0778dedeb0fb295f4

facebook-github-bot added the cla signed label Apr 27, 2021

kbrose requested review from albanD and soulitzer as code owners April 27, 2021 03:54

pytorchbot added the open source label Apr 27, 2021

H-Huang added the triaged This issue has been looked at a team member, and triaged and prioritized into an appropriate module label Apr 27, 2021

albanD approved these changes Apr 27, 2021

View reviewed changes

kbrose added 4 commits April 27, 2021 12:31

Test identifying problem w/ sinc' close to 0

6a31f06

Problem arises for sinc'(x) where x > 0, but x ** 2 == 0 (which happens for some very small floats).

Handle numbers very close to 0 in sinc derivative

765a00f

remove unneeded parens

37849ec

semicolons

0f4269a

albanD mentioned this pull request Apr 29, 2021

[DO NOT REVIEW] Sinc derivative numeric fix #57277

Closed

facebook-github-bot closed this in ec86f96 Apr 29, 2021

facebook-github-bot added the Merged label Apr 29, 2021

kbrose deleted the sinc-derivative-numeric-fix branch April 29, 2021 22:21

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Fix for derivative of sinc(x) when x is positive but very very small #56986

Fix for derivative of sinc(x) when x is positive but very very small #56986

Uh oh!

kbrose commented Apr 27, 2021 •

edited

Loading

Uh oh!

facebook-github-bot commented Apr 27, 2021 •

edited

Loading

Uh oh!

albanD left a comment

Uh oh!

albanD commented Apr 27, 2021

Uh oh!

kbrose commented Apr 27, 2021

Uh oh!

kbrose commented Apr 28, 2021

Uh oh!

albanD commented Apr 28, 2021

Uh oh!

kbrose commented Apr 29, 2021

Uh oh!

albanD commented Apr 29, 2021

Uh oh!

albanD commented Apr 29, 2021

Uh oh!

kbrose commented Apr 29, 2021

Uh oh!

albanD commented Apr 29, 2021 •

edited

Loading

Uh oh!

facebook-github-bot commented Apr 29, 2021

Uh oh!

kbrose commented Apr 29, 2021

Uh oh!

facebook-github-bot commented Apr 29, 2021

Uh oh!

Uh oh!

Fix for derivative of sinc(x) when x is positive but very very small #56986

Fix for derivative of sinc(x) when x is positive but very very small #56986

Uh oh!

Conversation

kbrose commented Apr 27, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

facebook-github-bot commented Apr 27, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

💊 CI failures summary and remediations

🕵️ 1 new failure recognized by patterns

binary_linux_libtorch_3_7m_cpu_gcc5_4_cxx11-abi_shared-with-deps_build (1/1)

Uh oh!

albanD left a comment

Choose a reason for hiding this comment

Uh oh!

albanD commented Apr 27, 2021

Uh oh!

kbrose commented Apr 27, 2021

Uh oh!

kbrose commented Apr 28, 2021

Uh oh!

albanD commented Apr 28, 2021

Uh oh!

kbrose commented Apr 29, 2021

Uh oh!

albanD commented Apr 29, 2021

Uh oh!

albanD commented Apr 29, 2021

Uh oh!

kbrose commented Apr 29, 2021

Uh oh!

albanD commented Apr 29, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

facebook-github-bot commented Apr 29, 2021

Uh oh!

kbrose commented Apr 29, 2021

Uh oh!

facebook-github-bot commented Apr 29, 2021

Uh oh!

Uh oh!

kbrose commented Apr 27, 2021 •

edited

Loading

facebook-github-bot commented Apr 27, 2021 •

edited

Loading

albanD commented Apr 29, 2021 •

edited

Loading