cannnot locate cuda/cuda_config.h with binary installed tensorflow. about tf_deformable_net HOT 3 CLOSED

Yes this would resolve the problem:

cd `python -c 'import tensorflow as tf; print(tf.sysconfig.get_include())'`
cd tensorflow/stream_executor/cuda
curl -O https://raw.githubusercontent.com/tensorflow/tensorflow/master/third_party/toolchains/gpus/cuda/cuda/cuda_config.h

But I'm not sure if this is the best practice since it seems to be a hack.. For now it's enough.
@Zardinality well your modification for make.sh in a0f9e94 it didn't resolve the problem entirely since you used both CUDA_HOME and CUDA_PATH in the script.

from tf_deformable_net.

Sorry for my neglection on the header stuff, last day in another repo someone file an issue about it, so I change README of it to add a reminder. But I forgot this repo. How I wish could get rid of this include! But I have to use cublas and allocate gpu memory in deform_conv.cc.

If I am right, cuda_config.h should be placed in $TF_INC/tensorflow/stream_executor/cuda and the simplest method to repair it is to copy the header to this path in make.sh. What do you think?

from tf_deformable_net.

I copied https://github.com/tensorflow/tensorflow/blob/master/third_party/toolchains/gpus/cuda/cuda/cuda_config.h to $ROOT_PYTHON/site-packages/tensorflow/include/cude/cuda_config.h. It works for me.

from tf_deformable_net.