On 10/17/20 8:10 AM, gevisz wrote:
Any thoughts on how to proceed with this obstacle other than to depclean tensorflow and try to reinstall it again? > The problem is that all versions of tensorflow and almost all of its dependencies are masked by default. So it is almost impossible to guess which versions of them should be unmasked for the successful compilation.
My strategy was to run emerge with --autounmask=y and then to add the dependencies to package.accept_keywords one-by-one. If any of the dependencies that it suggests to you are actually incompatible, it's a bug in the tensorflow ebuild (or the ebuild of one of its dependencies).
This is what I've currently got, but I originally generated the list for tensorflow-2.2 and then upgraded, so it may not be perfect. (It takes me three days to compile, so I'm not going to experiment!)
# tensorflow =dev-python/pandas-1.1.2 ~amd64 =dev-python/gast-0.3.3 ~amd64 =dev-python/astunparse-1.6.3 ~amd64 =net-libs/grpc-1.28.1-r1 ~amd64 =sci-libs/tensorflow-2.3.1 ~amd64 =net-libs/google-cloud-cpp-0.10.0-r1 ~amd64 =sci-visualization/tensorboard-2.3.0 ~amd64 =dev-util/bazel-3.2.0 ~amd64 =sci-libs/tensorflow-estimator-2.3.0 ~amd64 =sys-apps/hwloc-2.1.0 ~amd64 =dev-python/grpcio-tools-1.28.1 ~amd64 =dev-cpp/abseil-cpp-20200225.2 ~amd64 =dev-python/google-auth-oauthlib-0.4.1 ~amd64 =dev-python/google-pasta-0.2.0 ~amd64
