I see that it is possible to use libtensorflow
that is built from source, as mentioned in the README
https://github.com/tensorflow/tfjs-node#optional-build-libtensorflow-from-tensorflow-source
I have successfully built from source, but I don't know how to let tfjs-node
use this custom built version, instead of the pre-built one.
I found that the following workaround seems to work:
Copy the file bazel-bin/tensorflow/libtensorflow.so
from the tensorflow source directory to replace node_modules/@tensorflow/tfjs-node/deps/lib/libtensorflow.so
inside your project.
Hopefully there is a better way to do npm install
such that there is no need to download the pre-compiled version.