No OpKernel was registered to support Op 'RFFT' for CPU (running on android)

[liuzqt]

I'm running tensorflow on android. and got this error:
Caused by: java.lang.IllegalArgumentException: No OpKernel was registered to support Op 'RFFT' with these attrs. Registered devices: [CPU], Registered kernels: <no registered kernels>

I'm using the master branch, where the RFFT CPU verison is already supported. So I'm wondering why this problem show up.

The TF version I used to build the graph(*.pb file) is also the latest master branch.
Here is to code I wrtie .pb graph:

    with tf.Graph().as_default(), tf.Session(config=tf.ConfigProto(
            allow_soft_placement=True)) as session:
        with tf.variable_scope("model"):
            model = DeployModel(config=config)

        print('Graph build finished')
        # variable_names = [n.name for n in
        #                   tf.get_default_graph().as_graph_def().node]
        # for n in variable_names:
        #     print(n)

        saver = tf.train.Saver()
        saver.restore(session, save_path=path_join(self.config.model_path,
                                                   'latest.ckpt'))
        print("model restored from %s" % config.model_path)

        frozen_graph_def = graph_util.convert_variables_to_constants(
            session, session.graph.as_graph_def(),
            ['model/inputX', 'model/softmax', 'model/nn_outputs'])
        tf.train.write_graph(
            frozen_graph_def,
            os.path.dirname(graph_path),
            os.path.basename(graph_path),
            as_text=False,
        )

So maybe RFFT for CPU is still not supported on android in the latest branch?

[aselle]

Android only builds a subset of the ops, You need to add the ops that you need that are not in the "commonly used set" by hacking the build files for your needs.

[andrewharp]

@liuzqt You can follow @aselle's suggestion by adding fft_ops.cc to the //tensorflow/core/kernels:android_extended_ops_group1 filegroup and see if that solves the issue.

[liuzqt]

@andrewharp @aselle adding fft_ops.cc to build file solve the problem, many thanks! But it come up with a new problem, when running fft, raise this exception:
` Caused by: java.lang.IllegalArgumentException: No OpKernel was registered to support Op 'ComplexAbs' with these attrs. Registered devices: [CPU], Registered kernels:

                                                                 	 [[Node: model/fft = ComplexAbs[T=DT_COMPLEX64, Tout=DT_FLOAT, _device="/device:CPU:0"](model/rfft)]]
                                                                     at org.tensorflow.Session.run(Native Method)`

and then I refer to
https://github.com/tensorflow/tensorflow/blob/master/tensorflow/core/kernels/cwise_op_abs.cc
found that
#if !defined(IS_MOBILE_PLATFORM) REGISTER2(UnaryOp, CPU, "ComplexAbs", functor::abs, complex64, complex128); #endif
so it means Complex2real is not supported on android?
If I want to implement this, how should I do? Can I just remove the "if !defined" to make this REGISTER compiled to android lib?

[aselle]

It seems like complexabs is not implemented. You could just try removing the #if !defined and see if it works. I'm not sure why that check was put in, but it is possible that it was for an old version that didn't support the complex type. @petewarden , any ideas?

[liuzqt]

@aselle I remove that #if !defined and it works! many thank!