correlation_cost ./compile.sh error in Feelvos

What is the top-level directory of the model you are using: feelvos
OS Platform and Distribution: Ubuntu (server)
TensorFlow installed from: conda install
TensorFlow version: gpu-1.9.0
Bazel version: 1.9.0
CUDA/cuDNN version: 9.0
GPU model and memory: 2* titan xp 12G
python version = 2.7
Have I written custom code: No. Just directly run the original "sh ./build.sh" in correlation_cost part.

Exact command to reproduce:

1. When i was running original source code sh ./build.sh
error shows compile.sh has unexpected '(' in line 36, then I change it to

2. after that, error shows

so I tried to 
`chmod +x ./compile.sh` 
 and modify the last line in "build.sh" from `./compile.sh $1 ` to `"./compile.sh $1"` 

3. Then, error shows

4.Because until now all methods cannot work for me , I try to analyse every line in compile.sh. with echo.

5.after this, terminal shows the error that

I find something may get wrong in line 55 with echo ${TF_CFLAGS[@]} and this is also related to line 56 'nvcc ....${TF_CFLAGS[@]}...' and line 59 'g++ ....${TF_CFLAGS[@]}....' .

I don't know what the [@] means so I tried to delete this signal and replace line 56 and 59 with
nvcc -std=c++11 --expt-relaxed-constexpr -I ./ -I ${CUB_DIR}/../ -I ${THRUST_DIR} -I ${CUDA_DIR}/ -c -o correlation_cost_op_gpu.o kernels/correlation_cost_op_gpu.cu.cc ${TF_CFLAGS} -D GOOGLE_CUDA=1 -x cu -Xcompiler -fPIC

g++ -std=c++11 -I ./ -L ${CUDA_DIR}/cuda/lib64 -shared -o correlation_cost.so ops/correlation_cost_op.cc kernels/correlation_cost_op.cc correlation_cost_op_gpu.o ${TF_CFLAGS} -fPIC -lcudart ${TF_LFLAGS} -D GOOGLE_CUDA=1

6. Surprisingly, './compile.sh' finished with warnings and I got echo 8 and echo 9
   

7. Then I set USE_CORRELATION_COST = True in feelvos/utils/embedding_utils.py and try to run sh eval.sh. However, error repeatedly occurs.
   

I strongly hesitate whether tensorflow-gpu 1.9.0 installed by conda doesn't work here. I will try to create virtualenv to install tensorflow-gpu by pip and CUDA9.0 one more time outside the anaconda env. Is there any one can completely figure out this issue? Thank you!