Closed friissoren closed 2 years ago
@friissoren as you pointed out, it's actually not necessary to have the posenet decoder custom ops for classification example. However MakeEdgeTpuInterpreter
is not made specifically for classification. It's actually just a convenience for initializing an interpreter that can be used for all purposes, including posenet, as you could see here: https://github.com/google-coral/libcoral/blob/master/coral/pose_estimation/test_utils.cc#L111
When I follow the code for the classify_image example, it setups the interpreter here: https://github.com/google-coral/libcoral/blob/master/coral/examples/classify_image.cc#L35
As far as I can see, that results in a call to tflite_utils MakeEdgeTpuInterpreter. Why does that function add custom operations related to the pose net? That shouldn't be necessary when doing image classification: https://github.com/google-coral/libcoral/blob/master/coral/tflite_utils.cc#L72
Of course, I might have misunderstood how this is supposed to work.