SigmoidFocalLoss [Bug]

[HKEa]

Prerequisite

• I have searched Issues and Discussions but cannot get the expected help.
• The bug has not been fixed in the latest version(https://github.com/open-mmlab/mmcv).

Environment

Hello

thank you for a great library.
But looks like a bug in SigmoidFocalLoss:

OrderedDict([('sys.platform', 'linux'), ('Python', '3.10.12 (main, Jul 5 2023, 18:54:27) [GCC 11.2.0]'), ('CUDA available', True), ('numpy_random_seed', 2147483648), ('GPU 0,1,2,3,4,5,6,7', 'NVIDIA A40'), ('CUDA_HOME', '/usr/local/cuda'), ('NVCC', 'Cuda compilation tools, release 12.2, V12.2.91'), ('GCC', 'gcc (Ubuntu 9.4.0-1ubuntu1~20.04.1) 9.4.0'), ('PyTorch', '2.0.1'), ('PyTorch compiling details', 'PyTorch built with:\n - GCC 9.3\n - C++ Version: 201703\n - Intel(R) oneAPI Math Kernel Library Version 2023.1-Product Build 20230303 for Intel(R) 64 architecture applications\n - Intel(R) MKL-DNN v2.7.3 (Git Hash 6dbeffbae1f23cbbeae17adb7b5b13f1f37c080e)\n - OpenMP 201511 (a.k.a. OpenMP 4.5)\n - LAPACK is enabled (usually provided by MKL)\n - NNPACK is enabled\n - CPU capability usage: AVX2\n - CUDA Runtime 11.8\n - NVCC architecture flags: -gencode;arch=compute_37,code=sm_37;-gencode;arch=compute_50,code=sm_50;-gencode;arch=compute_60,code=sm_60;-gencode;arch=compute_61,code=sm_61;-gencode;arch=compute_70,code=sm_70;-gencode;arch=compute_75,code=sm_75;-gencode;arch=compute_80,code=sm_80;-gencode;arch=compute_86,code=sm_86;-gencode;arch=compute_90,code=sm_90;-gencode;arch=compute_37,code=compute_37\n - CuDNN 8.7\n - Magma 2.6.1\n - Build settings: BLAS_INFO=mkl, BUILD_TYPE=Release, CUDA_VERSION=11.8, CUDNN_VERSION=8.7.0, CXX_COMPILER=/opt/rh/devtoolset-9/root/usr/bin/c++, CXX_FLAGS= -D_GLIBCXX_USE_CXX11_ABI=0 -fabi-version=11 -Wno-deprecated -fvisibility-inlines-hidden -DUSE_PTHREADPOOL -DNDEBUG -DUSE_KINETO -DLIBKINETO_NOROCTRACER -DUSE_FBGEMM -DUSE_QNNPACK -DUSE_PYTORCH_QNNPACK -DUSE_XNNPACK -DSYMBOLICATE_MOBILE_DEBUG_HANDLE -O2 -fPIC -Wall -Wextra -Werror=return-type -Werror=non-virtual-dtor -Werror=bool-operation -Wnarrowing -Wno-missing-field-initializers -Wno-type-limits -Wno-array-bounds -Wno-unknown-pragmas -Wunused-local-typedefs -Wno-unused-parameter -Wno-unused-function -Wno-unused-result -Wno-strict-overflow -Wno-strict-aliasing -Wno-error=deprecated-declarations -Wno-stringop-overflow -Wno-psabi -Wno-error=pedantic -Wno-error=redundant-decls -Wno-error=old-style-cast -fdiagnostics-color=always -faligned-new -Wno-unused-but-set-variable -Wno-maybe-uninitialized -fno-math-errno -fno-trapping-math -Werror=format -Werror=cast-function-type -Wno-stringop-overflow, LAPACK_INFO=mkl, PERF_WITH_AVX=1, PERF_WITH_AVX2=1, PERF_WITH_AVX512=1, TORCH_DISABLE_GPU_ASSERTS=ON, TORCH_VERSION=2.0.1, USE_CUDA=ON, USE_CUDNN=ON, USE_EXCEPTION_PTR=1, USE_GFLAGS=OFF, USE_GLOG=OFF, USE_MKL=ON, USE_MKLDNN=ON, USE_MPI=OFF, USE_NCCL=ON, USE_NNPACK=ON, USE_OPENMP=ON, USE_ROCM=OFF, \n'), ('TorchVision', '0.15.2'), ('OpenCV', '4.8.0'), ('MMEngine', '0.8.2'), ('MMCV', '2.0.1'), ('MMCV Compiler', 'GCC 9.3'), ('MMCV CUDA Compiler', '11.8')])

I've checked https://github.com/open-mmlab/mmcv/blob/main/mmcv/ops/csrc/common/cuda/sigmoid_focal_loss_cuda_kernel.cuh
for num_classes=1=> c=0, the result is that flag_p and flag_n just swap
T flag_p = (t == c);
T flag_n = (t != c);
for sigmoid like target, the num_classes=1 from https://github.com/open-mmlab/mmcv/blob/main/mmcv/ops/csrc/pytorch/cuda/focal_loss_cuda.cu
(int num_classes = input.size(1);)

BR

Reproduces the problem - code sample

loss=SigmoidFocalLoss(alpha=-1, gamma=2)
loss(torch.tensor([[-1000.]]).cuda(), torch.tensor([1]).cuda())
tensor(0., device='cuda:0')
loss(torch.tensor([[1000.]]).cuda(), torch.tensor([1]).cuda())
tensor(174.6731, device='cuda:0')

Reproduces the problem - command or script

none

Reproduces the problem - error message

none

Additional information

No response

[qingpeng9802]

I am currently working on fixing softmax focal algorithm in #2893 as an outside contributor.
Per my understanding, the current sigmoid focal loss implementation is correct. It is very similar to the implementation of
https://github.com/pytorch/pytorch/blob/main/modules/detectron/sigmoid_focal_loss_op.cu

If I understand your issue correctly, you are confusing with the binary classification task here. This is actually a very tricky part. The pytorch's implementaion mentioned above actually drops the background class (you can see a weird d+1). However, mmcv's implementation dose not drop it (so you may need to drop it manually if you need). See pytorch/vision#3250 for more discussion.

In short, num_classes=1 is invalid here, and you should have a background class and a foreground class. Feel free to add any comment.

[HKEa]

as you can see the results between torchvision implementation and this are different
the num_classes=1 because size of the logits last dim=1 for binary/focal classification

[qingpeng9802]

as you can see the results between torchvision implementation and this are different the num_classes=1 because size of the logits last dim=1 for binary/focal classification

torchvision implementation is correct for both num_classes>1 and num_classes=1, but mmcv's implementation is only correct for num_classes>1.
This is the natural limitation of mmcv's implementation. In object detection tasks, we are usually assigning classes to anchors. For example, the classes are [cat, dog]. For your case, the classes should be [cat, notCat(background)], that is, you have to give 2 classes.

You can imagine that the label with values 0 and 1 is forced to expand into a 2-channel tensor (one-hot) so you have to give a 2-channel input to match the 2-channel label tensor.

I would strongly recommend using a 2-channel input for your case if you are in object detection tasks since 2-channel input solution is common in object detection tasks. However, it is also possible to add a specialized kernel function for num_classes=1 or a d+1 version like detectron to mmcv.