Android - No OpKernel was registered to support Op 'Cos' with these attrs. Registered devices: [CPU], Registered kernels: <no registered kernels> [[Node: stft/hann_window/Cos = Cos[T=DT_FLOAT, _device="/device:CPU:0"](stft/hann_window/truediv)]]

[eli99999]

I've exported a model to Android that uses :
stfts = tf.contrib.signal.stft(transwav, frame_length=2048, frame_step=512,
fft_length=2048,window_fn=functools.partial(tf.contrib.signal.hann_window, periodic=False), pad_end=True)
the model works properly in pyhton
but when i load it in Android using the downloaded compiled tensorflow - compile 'org.tensorflow:tensorflow-android:+'
I get this error -
FATAL EXCEPTION: main
Process: org.tensorflow.demo, PID: 25836
java.lang.IllegalArgumentException: No OpKernel was registered to support Op 'Cos' with these attrs. Registered devices: [CPU], Registered kernels:
[[Node: stft/hann_window/Cos = CosT=DT_FLOAT, _device="/device:CPU:0"]]

which comes from the hann_window

any recommended work around?

[jart]

This is a very common issue with Android development. You need to make sure all the kernels you need are in your build. Please check StackOverflow and search Google for similar issues.

[jart]

Please note that C++ kernels use a registration pattern.

[eli99999]

I have recompiled libtensorflow_inference.so with print_selective_registration_header
and i see in the ops_to_register.h file the Cos OP -

...
constexpr inline bool ShouldRegisterOp(const char op[]) {
return false
|| isequal(op, "Add")
|| isequal(op, "BatchToSpaceND")
|| isequal(op, "CTCBeamSearchDecoder")
|| isequal(op, "Cast")
|| isequal(op, "Complex")
|| isequal(op, "ComplexAbs")
|| isequal(op, "ConcatV2")
|| isequal(op, "Const")
|| isequal(op, "Conv2D")
|| isequal(op, "Cos")
|| isequal(op, "DecodeWav")
|| isequal(op, "Exp")
|| isequal(op, "ExpandDims")
|| isequal(op, "Fill")...
there for it should be compiled into the libtensorflow_inference.so

but i still get the same error message when trying to run the model in my android app

[bobeo]

@eli99999 Please let me know if this help.

[eli99999]

@bobeo thanks,
I looked into it, and it seems that there is a problem with the bazel compilation
i changed the models a few time and then did the ops_to_register.h process,
each time the ops_to_register.h file changed but the libtensorflow_inference.so does not.
I tried all sort of bazel clean and bazel dump, and not chaned the resulted libtensorflow_inference.so
so i think that either there is some issue with the bazel caching or that the bug is in the print_selective_registration_header process

i'm working on ubuntu 16.04
bazel 0.7.0
tf 1.4

[naktinis]

I've answered a similar question on StackOverflow: https://stackoverflow.com/questions/49116833/no-opkernel-was-registered-to-support-op-cos-when-running-inference-on-android

Apparently, you might need to "hack" the BUILD file, according to one of the TensorFlow developers: #11804 (comment)