Installing Bazel is easy:
sudo apt-get install openjdk-8-jdk wget https://github.com/bazelbuild/bazel/releases/download/0.4.3/bazel-0.4.3-installer-linux-x86_64.sh ./bazel-0.4.3-installer-linux-x86_64.sh --user rm bazel-0.4.3-installer-linux-x86_64.sh
For building Tensorflow, the created ~/bin directory should be added to the path (e.g. in .bashrc):
# Bazel export PATH=$PATH:~/bin
Another interesting error I came across using Tensorflow is a result of having multiple GPUs. For example, one GPU dedicated for heavy computation, and another, cheaper GPU to handle multiple displays and provide a smooth user interface. The latter is obviously not meant to be used with CUDA. Still, Tensorflow tries to grab all available graphics cards (as described here) which results in the following error:
W tensorflow/stream_executor/cuda/cuda_driver.cc:590] creating context when one is currently active; existing: 0x3fd7d50 E tensorflow/core/common_runtime/direct_session.cc:136] Internal: failed initializing StreamExecutor for CUDA device ordinal 1: Internal: failed call to cuDevicePrimaryCtxRetain: CUDA_ERROR_INVALID_DEVICE # ... tensorflow.python.framework.errors_impl.InternalError: Failed to create session.
The solution is to tell CUDA which graphics cards to consider:
export CUDA_VISIBLE_DEVICES=0 # use nvidia-smi to check the correct index
Alternatively, the full GPU id can also be used — these can be used using nvidia-smi -L.
What is your opinion on this article? Let me know your thoughts on Twitter @davidstutz92 or LinkedIn in/davidstutz92.