Implement Custom Ops in build

[ragavendranbala]

Hi @terryky , I want to load the hairsegmentation model in python but when i try to initialize the tflite model in interpreter, it always showing MaxPoolingWithArgmax2D ops error. Can you tell me how to build this custom ops with bazel when building tensorflow. if I'm wrong please tell me how to implement custom ops. Thanks in Advance

[terryky]

Hi @ragavendranbala,

I'm not familier with Tensorflow Lite python API, but It's easy to implement custom Op in C++ API.
Please refer follows:

• https://www.tensorflow.org/lite/guide/inference#write_a_custom_operator

• https://github.com/terryky/tflite_gles_app/blob/master/gl2hair_segmentation/tflite_hair_segmentation.cpp#L25-L26

[ragavendranbala]

@terryky thanks for your reply, actually i wanted to build the plugin for android, i think , to make it work on android, i need to add this custom ops while building the plugin itelf. I have searched for links to reference the plugin in build. but no luck. i guess the build of plugin working on c++ only. I''m expecting your ideas for this

[ragavendranbala]

hi @terryky , I have built the plugin which custom ops. now i need to know, is this running on CPU or GPU? is there is any possibilities to make the ops to run on GPU?

[terryky]

Tensorflow Lite GPU Delegate supports following custom operations.
If your implementation enables GPU Delegate, the custom ops run on GPU.

https://github.com/tensorflow/tensorflow/blob/master/tensorflow/lite/delegates/gpu/common/model_builder.cc#L2715-L2753

[ragavendranbala]

@terryky the link you have given having the ops functions, but i think it's not building with my plugin

bazel build -c opt --config android_arm64 --copt -Os --copt -DTFLITE_GPU_BINARY_RELEASE --copt -fvisibility=hidden --linkopt -s --strip always //tensorflow/lite/delegates/gpu:libtensorflowlite_gpu_delegate.so

This is command I'm using to build my GPU delegate.

Should I add any files to make it work on GPU?

[terryky]

Does your app enable GPU Delegate ?

delegate = TfLiteGpuDelegateV2Create(&options);
interpreter->ModifyGraphWithDelegate(delegate)

https://github.com/terryky/tflite_gles_app/blob/master/common/util_tflite.cpp#L167-L174

[ragavendranbala]

@terryky The plugin I have created running deeplab v3 on smoothly but this tflite is hicking up , based on the deeplab performance, I think gpu delegate is enabled

[soham24]

@ragavendranbala can you share the android implementation