1

I read in this post by HediBy that this ROI_POOLING user_op implementation works: LINK

I used bazel build 

-c opt //tensorflow/core/user_ops:roi_pooling.so to generate the so file (after installing tensorflow)

But when I use tf.load_op_library to load roi_pooling.so, I get this weird error: 

 tensorflow.python.framework.errors.NotFoundError: /home/fishdrop/tensorflow/bazel-bin/tensorflow/core/user_ops/roi_pooling.so: undefined symbol: _Z21ROIPoolForwardLaucherPKffiiiiiiS0_PfPiRKN5Eigen9GpuDeviceE

Any ideas? has anyone else been successful with this user_op?

Community
  • 1
  • 1
Link L
  • 439
  • 7
  • 25
  • 2
    Are you trying to run on a GPU? Your bazel build command line is incomplete, but if you are trying to run on a GPU, I would try compiling with the following command line : `bazel build --config=cuda -c opt //tensorflow/core/user_ops:roi_pooling.so` – keveman Aug 16 '16 at 16:51

1 Answers1

1

thanks for the tip i just found out that the error occured when I built the user_op without GPU support. I reinstalled tensorflow v 0.10, with GPU support, and placed all user_op files inside //tensorflow/core/us‌​er_ops.

If I compile the user_op using bazel build -c opt --config=cuda //tensorflow/core/us‌​er_ops:roi_pooling.so (the addition of --config=cuda isn't in the tensorflow user_op documentation), the user_op now works.. I guess this issue could now be closed

Link L
  • 439
  • 7
  • 25