没有OpKernel Tensorflow Mobile Android。 如何调试？

Question

I know this is a very common error to see in Android Studio. I used to see this for 'GatherNd' until I checked my tf_op_files.txt against the one in master and saw that 'gather_nd_op' had been added recently. After building with bazel (though it seems jcenter/'none' still yields 'GatherNd' not found), I don't see this anymore. However, I do see this now with 'StridedSlice'.

java.lang.IllegalArgumentException: No OpKernel was registered to support Op 'StridedSlice' with these attrs.  Registered devices: [CPU], Registered kernels:
                                                                       device='CPU'; T in [DT_BFLOAT16]
                                                                       device='CPU'; T in [DT_BOOL]
                                                                       device='CPU'; T in [DT_FLOAT]
                                                                       device='CPU'; T in [DT_INT32]

                                                                         [[Node: roi_align_classifier/strided_slice = StridedSlice[Index=DT_INT32, T=DT_INT64, begin_mask=1, ellipsis_mask=0, end_mask=1, new_axis_mask=0, shrink_axis_mask=2](roi_align_classifier/Where, roi_align_classifier/strided_slice/stack, roi_align_classifier/strided_slice/stack_1, roi_align_classifier/strided_slice/stack_2)]]
                                                                         at org.tensorflow.Session.run(Native Method)
                                                                         at org.tensorflow.Session.access$100(Session.java:48)
                                                                         at org.tensorflow.Session$Runner.runHelper(Session.java:298)
                                                                         at org.tensorflow.Session$Runner.run(Session.java:248)
                                                                         at org.tensorflow.contrib.android.TensorFlowInferenceInterface.run(TensorFlowInferenceInterface.java:218)

I see that this op exists in tf_op_files.txt and that the error can occur from using the wrong parameters. I have tried changing the input and output types from float[] to int[] but no luck.

I see the type signature and info about strided_slice in array_ops.cc and https://www.tensorflow.org/api_docs/python/tf/strided_slice . However, these don't provide great insight as to whether I'm merely supplying the wrong params or the StridedSlice I am using actually doesn't exist!

My question is: how do I debug this in a more deterministic way than random guessing?

Edit Checking http://www.oreilly.com/data/free/files/building-mobile-applications-with-tensorflow.pdf , it seems this implementation is correct and that it is not supported for mobile currently? ie. this error message is not dependent on the input and output passed. Is this right? I think adding the op would be rather straightforward. I just need to know if that's what the problem is and what the type is supposed to be.

I've also found https://medium.com/joytunes/deploying-a-tensorflow-model-to-android-69d04d1b0cba . That said, any debugging information on the android/type side would be great. ie. what is the type being passed to the op and what is it expecting? It seems I have to use outside resources to determine this

Edit 2

After finding all the ops, quantizing, getting the inputs of input nodes and outputs of output nodes, I tried adding the strided_slice_op.cc to the android_extended_ops_group2 only to see it already existed when building. I think I have to add another implementation. Does anyone know how to get the full type signature of the nodes and does this seem like I'm on the right path? In tensorboard and in the error, the op is strided_slice[0-9] while the build is like so:

`tf_kernel_library(
    name = "strided_slice_op",
    srcs = [
        "strided_slice_op.cc",
        "strided_slice_op_inst_0.cc",
        "strided_slice_op_inst_1.cc",
        "strided_slice_op_inst_2.cc",
        "strided_slice_op_inst_3.cc",
        "strided_slice_op_inst_4.cc",
        "strided_slice_op_inst_5.cc",
        "strided_slice_op_inst_6.cc",
        "strided_slice_op_inst_7.cc",
    ],
    hdrs = [
        "slice_op.h",
        "strided_slice_op.h",
        "strided_slice_op_impl.h",
    ],
    gpu_srcs = [
        "slice_op.h",
        "strided_slice_op.h",
        "strided_slice_op_impl.h",
        "strided_slice_op_gpu.cu.cc",
    ],
    deps = [
        ":bounds_check",
        ":dense_update_functor",
        ":ops_util",
        ":variable_ops",
        "//tensorflow/core:framework",
        "//tensorflow/core:lib",
        "//third_party/eigen3",
    ],
)`

Answer 1

The problem was staring me in the face the whole time. T=DT_INT64 is not one of the registered kernels for strided_slice. Editing the model to cast before slicing solved the issue for me.

Ref: https://github.com/tensorflow/tensorflow/issues/13772

On to registering a custom op!

Edit

Just needed to change the order of operations!