Failed to build tensorflow iOS binaries with selective registration - ios

Question

I am a new user of tensorflow. I wish to integrate a specific graph.pb file into an ios device, and have followed all the compilation instructions carefully as described in tensorflow git.
yet when I try to run the camera example app I get the error:
E /tensorflow/tensorflow/examples/ios/camera_copy/tensorflow_utils.mm:140] Could not create TensorFlow Graph: Invalid argument: No OpKernel was registered to support Op 'Conv2D' with these attrs. Registered devices: [CPU], Registered kernels:
<no registered kernels>
[[{{node pyramid_regression_0/convolution}} = Conv2D[T=DT_FLOAT, data_format="NHWC", dilations=[1, 1, 1, 1], padding="SAME", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true](input_2, pyramid_regression_0/kernel/read)]]
2018-08-22 16:54:00.303760: F tensorflow/examples/ios/camera_copy/tensorflowUtil.mm:68] Couldn't load model: Invalid argument: No OpKernel was registered to support Op 'Conv2D' with these attrs. Registered devices: [CPU], Registered kernels:
<no registered kernels>
[[{{node pyramid_regression_0/convolution}} = Conv2D[T=DT_FLOAT, data_format="NHWC", dilations=[1, 1, 1, 1], padding="SAME", strides=[1, 1, 1, 1], use_cudnn_on_gpu=true](input_2, pyramid_regression_0/kernel/read)]]
I also followed the instructions according to the solution offered in here, but I got there are some differences (probably due to the versions diffs between a year ago and now (Aug 2018))
bazel build --copt="-DUSE_GEMM_FOR_CONV" tensorflow/python/tools/print_selective_registration_header compiled without errors
then created the ops_to_register.h file:
when used the bazel-bin command, it adds a strange print of tf.estimator package not installed to the first line.
In addition, during the file creation this warning appears multiple times:
RuntimeWarning: numpy.dtype size changed, may indicate binary incompatibility. Expected 96, got 88
when I create the ops_to_register.h file using python3 , it creates the file without strange prints
in both cases there are no occurrences of strcmp or strcmpc at all, and got Conv2DUsingGemmOp op in the bazel created file instead of Conv2DOp by python3
I tried to use the ops_to_register.h file for the tensorflow/contrib/makefile/compile_ios_tensorflow.sh compilation, but it fails during the compilation:
Undefined symbols for architecture arm64:
"nsync::nsync_cv_signal(nsync::nsync_cv_s_)", referenced from:
tensorflow::condition_variable::notify_one() in libtensorflow-core-arm64.a(mutex.o)
"nsync::nsync_cv_wait_with_deadline(nsync::nsync_cv_s_, nsync::nsync_mu_s_, timespec, nsync::nsync_note_s_)", referenced from:
tensorflow::condition_variable::wait_until_system_clock(tensorflow::mutex_lock&, std::__1::chrono::time_point > >) in libtensorflow-core-arm64.a(mutex.o)
"nsync::nsync_cv_init(nsync::nsync_cv_s_)", referenced from:
tensorflow::condition_variable::condition_variable() in libtensorflow-core-arm64.a(mutex.o)
"nsync::nsync_from_time_point_(std::__1::chrono::time_point > >)", referenced from:
tensorflow::condition_variable::wait_until_system_clock(tensorflow::mutex_lock&, std::__1::chrono::time_point > >) in libtensorflow-core-arm64.a(mutex.o)
"nsync::nsync_cv_wait(nsync::nsync_cv_s_, nsync::nsync_mu_s_)", referenced from:
tensorflow::condition_variable::wait(tensorflow::mutex_lock&) in libtensorflow-core-arm64.a(mutex.o)
"nsync::nsync_mu_runlock(nsync::nsync_mu_s_)", referenced from:
tensorflow::mutex::unlock_shared() in libtensorflow-core-arm64.a(mutex.o)
"nsync::nsync_mu_rlock(nsync::nsync_mu_s_)", referenced from:
tensorflow::mutex::lock_shared() in libtensorflow-core-arm64.a(mutex.o)
"nsync::nsync_mu_unlock(nsync::nsync_mu_s_)", referenced from:
tensorflow::mutex::unlock() in libtensorflow-core-arm64.a(mutex.o)
"nsync::nsync_mu_lock(nsync::nsync_mu_s_)", referenced from:
tensorflow::mutex::lock() in libtensorflow-core-arm64.a(mutex.o)
"nsync::nsync_cv_broadcast(nsync::nsync_cv_s_)", referenced from:
tensorflow::condition_variable::notify_all() in libtensorflow-core-arm64.a(mutex.o)
"nsync::nsync_mu_init(nsync::nsync_mu_s_*)", referenced from:
tensorflow::mutex::mutex() in libtensorflow-core-arm64.a(mutex.o)
ld: symbol(s) not found for architecture arm64
clang: error: linker command failed with exit code 1 (use -v to see invocation)
make:
[/tensorflow/tensorflow/contrib/makefile/gen/bin/ios_ARM64/benchmark] Error 1
'[' 2 -ne 0 ']'
echo 'arm64 compilation failed.'
arm64 compilation failed.
exit 1
environment details:
tensorflow 1.9.0 (there is tensorflow 1.10.0, but is currently not supported for the ios compilations)
python 3.6.6
all the compilations are done with -a arm64 only
does anyone have an idea what went wrong here?
thanks a lot

Answer 1

I once had a similar error when compiling for the raspberry pi, I had to manually export the environment variables HOST_NSYNC_LIB and TARGET_NSYNC_LIB
First check if you have them already exported: echo $HOST_NSYNC_LIB
if that is empty you should look where your nsync library is compiled (just look for libnsync.a)
Then export both paths:
export HOST_NSYNC_LIB='your/path/to/libnsync.a'
export TARGET_NSYNC_LIB="$HOST_NSYNC_LIB"
Hope this works for you!