warning: calling a constexpr __host__ function from a __host__ __device__ function is not allowed. The experimental flag '--expt-relaxed-constexpr' can be used to allow this. about faster-rcnn_tf

[xxi511]

My environment is CUDA 8.0 cudnn 5, work well (demo and test)
If you think The experimental flag '--expt-relaxed-constexpr' can be used to allow this. is annoyed
you can modify make.sh like below, add --expt-relaxed-constexpr

if [ -d "$CUDA_PATH" ]; then
	nvcc -std=c++11 -c -o roi_pooling_op.cu.o roi_pooling_op_gpu.cu.cc \
		-I $TF_INC -D GOOGLE_CUDA=1 -x cu -Xcompiler -fPIC $CXXFLAGS \
		-arch=sm_37 --expt-relaxed-constexpr

[miraclebiu]

I have the same warning ,but all things can work well. You can run demo and train net.

[emily2he]

what about your environment? is tensorflow installed by pip or make by yourself? cuda version is 7.5, cudnn is v3 in my environment.

[angelfish91]

I meet the same problem and my tensorflow is installed by pip, CUDA 8.0 cudnn 5? can you help me ?

[Nummer42O]

I experienced this problem in some of my own cuda code. I'm currently figuring out how to get rid of it since it doesn't seem to be right. For me it happens when using thrust::exp in a host function. Oddly enough compiling the same code with gcc works perfectly fine but compiling when it with nvcc and gcc it spits out those warnings but still works perfectly fine.

So for now just ignore it - I will try to update you if I found a solution.

EDIT:

It seems like you can't really get rid of those warnings except with --expt-relaxed-constexpr which I'd strongly advise against since you don't know with what else in the code it might interfere. As long as everything still works it happens to be only host code and everything should be fine.
I know it's annoying but just ignore those warnings.