According to other tutorials installed basically no problem, but a few days ago to upgrade a bit ubuntu kernel, then TensorFlow error
ImportError: libnvidia-fatbinaryloader.so.384.130: cannot open shared object file: No such file or directory
It seems nvidia driver problems that may be caused because the kernel upgrade, so I plan to reinstall the driver.
Reinstall the driver
First try .run installation, the installation fails:
Failed to run `/sbin/dkms build -m nvidia -v 384.90 -k …
Then use apt-get:
sudo apt-get install nvidia-384
The installation was successful, but after the restart the login screen cycle. And error:
nvidia-smi command not found
Suspected driver version is too low, there is the 430 installation:
sudo apt-get install nvidia-430
Restart error:
nvidia driver the system is running in low-graphics mode
Helpless and replace the 384, this adds several packages, there is no problem landing cycles, and nvidia-smi normal:
sudo apt-get install nvidia-384 nvidia-settings nvidia-prime
Probably because my computer is a dual graphics cards, so there will be the problem.
However, tf run the program again, still being given:
import error:libnppi.so.7.5: cannot open shared object file: No such file or directory
Check the / usr local cuda / at / / lib64 did not have this so, and I decided to reinstall cuda
Reloading cuda
First installed cuda 8.0, Bahrain still being given:
ImportError: libcudart.so.9.0: cannot open shared object file: No such file or directory
Although cuda's lib, there libcudart.so.8.0, but why use 9.0? Global search a bit this file:
find / -name libcudart.so.9.*
result:
/home/xxlyu/.cache/bazel/_bazel_xxlyu/8fa709ebf344796d539e4e2cfed28084/execroot/org_tensorflow/bazel-out/k8-opt/bin/tensorflow/cc/ops/candidate_sampling_ops_gen_cc.runfiles/org_tensorflow/_solib_local/_U@local_Uconfig_Ucuda_S_Scuda_Ccudart___Uexternal_Slocal_Uconfig_Ucuda_Scuda_Scuda_Slib/libcudart.so.9.0
/home/xxlyu/.cache/bazel/_bazel_xxlyu/8fa709ebf344796d539e4e2cfed28084/execroot/org_tensorflow/bazel-out/k8-opt/bin/tensorflow/cc/ops/candidate_sampling_ops_gen_cc.runfiles/org_tensorflow/external/local_config_cuda/cuda/cuda/lib/libcudart.so.9.0
Because before I use the cuda 9.0, it is estimated TensorFlow cache is not clear ... too lazy to control, or with the original 9.0 bar. Reloading process has given:
Error: libcudart.so.9.0: cannot open shared object file: No such file or directory
This so exist, the environment variable is also set up, because there is no update dynamic link:
sudo ldconfig /usr/local/cuda/lib64
Reference: https://blog.csdn.net/mumoDM/article/details/79502848
After installing cudnn behind, suggesting no libcudnn.so is the same solution.