libtensorflow_cc

Build TensorFlow C++ APIs with ease. See Releases for prebuilt tarballs.

Use the tarball

1. Extract the contents of the tarball.
2. Add libtensorflow_cc/include to the header searching directory list. (with -I option.)
3. Add libtensorflow_cc/lib to the library searching directory list. (with -L option.)
4. Link against libprotobuf.a, libtensorflow_cc.so, and libtensorflow_framework.so. (with -l option.) Notice that because ProtoBuf has a very struct versioning requirement, you have to compile with and link against the protobuf shipped within the tarball, i.e., the version used to compile TensorFlow.

Makefile example

See this Makefile example to get more sense on how to compile and link your C++ program with TensorFlow. The example can be built and and run with the following commands:

cd examples/mnist

# Build with GPU support
make CUDA=/usr/local/cuda LIBTENSORFLOW_CC=/path/to/libtensorflow_cc-gpu

# Build without GPU support
make CUDA= LIBTENSORFLOW_CC=/path/to/libtensorflow_cc-cpu

# Should prints "8"
./mnist

Build the tarball

usage: env KEY1=VALUE1 KEY2=VALUE2 ./scripts/build.py [args]

positional arguments:
  version               TensorFlow release tag. e.g. `v2.3.1`.
                        See: https://github.com/tensorflow/tensorflow/tags
  {cpu,gpu}             Whether to support NVIDIA GPU.

optional arguments:
  --preset              Presets for the environment variables and docker images.
  --rm                  Remove the builder container after success.

To build from scratch, use scripts/build.py. For example, to build TensorFlow v2.3.1 without CUDA support:

./scripts/build.py v2.3.1 cpu --preset tensorflow-2.3.0

You can override the build options by providing the environment variables. See scripts/build.py for the list of available environment variables. For example,

env DOCKER_IMAGE="nvidia/cuda:11.1-cudnn8-devel-ubuntu18.04" \
    TF_CUDA_COMPUTE_CAPABILITIES="8.0" \
    ./scripts/build.py v2.3.1 gpu --preset tensorflow-2.3.0