Build fails to copy all required linux shared objects for TensorFlow version 1.15.

[craigwalton]

Describe the bug
Build fails to copy all required linux shared objects for TensorFlow version 1.15, TensorFlowSharp version 1.15.1.

To Reproduce
Run using TensorFlowSharp 1.15.1 NuGet packages on linux.

Expected behavior
Required dependencies to be available.

Exception
Unhandled Exception: System.DllNotFoundException: Unable to load shared library 'libtensorflow' or one of its dependencies.

Desktop (please complete the following information):
Debian GNU/Linux 9 (stretch) within Docker container.

Additional context
Only libtensorflow.so and libtensorflow_framework.so are copied from libtensorflow\lib to runtimes\linux\native\ as part of the build specified in TensorFlowSharp.csproj. The following additional libraries from libtensorflow are also required for running TensorFlow 1.15 on linux:

libtensorflow.so.1
libtensorflow.so.1.15.0
libtensorflow_framework.so.1
libtensorflow_framework.so.1.15.0

I suspect the same issue exists for the .dylib files required by OS X, though have not tested this myself.

I propose replacing the following lines in TensorFlowSharp.csproj:

<None Include="..\native\libtensorflow.so" Link="nuget\runtimes\linux\native\libtensorflow.so" PackagePath="runtimes\linux\native\libtensorflow.so" Pack="true" />
<None Include="..\native\libtensorflow_framework.so" Link="nuget\runtimes\linux\native\libtensorflow_framework.so" PackagePath="runtimes\linux\native\libtensorflow_framework.so" Pack="true" />

With the following:

<None Include="..\native\libtensorflow*.so*">
  <Link>nuget\runtimes\linux\native\%(Filename)%(Extension)</Link>
  <PackagePath>runtimes\linux\native\%(Filename)%(Extension)</PackagePath>
  <Pack>True</Pack>
</None>

Libtensorflow binaries
https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-cpu-linux-x86_64-1.15.0.tar.gz
https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-cpu-darwin-x86_64-1.15.0.tar.gz
https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-cpu-windows-x86_64-1.15.0.zip

Please let me know if you'd like me to create a pull request to resolve this issue.

[nateagr]

Hi @craigwalton,

I'm facing the same issue as you. In the meantime, I've created my own NuGet which is working. But it is not productized. It seems this project is not maintained anymore. How did you overcome the issue on your side ? Did you fork the project ? Did you create a new NuGet as well ? If yes, did you publish it in public repo ?

Thanks.

[craigwalton]

Hi @nateagr,

I am using TensorFlowSharp within a Docker container. So I added a RUN instruction to the Dockerfile which downloads (https://storage.googleapis.com/tensorflow/libtensorflow/libtensorflow-cpu-linux-x86_64-1.15.0.tar.gz) and unpacks the missing libraries into the required directory. I did not fork the project or publish anything on NuGet.

Sorry - this probably is not much help to you.

[nateagr]

Thanks for your answer 👍