TensorFlow C API examples for Windows, Linux, and macOS, with a small C++ helper library.
- CMake 3.20 or newer and a C++17 compiler.
- Python with pip; CI uses Python 3.12.
- A 64-bit platform supported by the TensorFlow wheel.
git clone --depth 1 https://github.com/Neargye/hello_tf_c_api
cd hello_tf_c_apicmake -S . -B build -A x64
cmake --build build --config Release
ctest --test-dir build --output-on-failure -C Releasecmake -S . -B build -DCMAKE_BUILD_TYPE=Release
cmake --build build
ctest --test-dir build --output-on-failureCMake downloads the TensorFlow 2.21.0 wheel into build/_deps/tensorflow/python and links its native libraries. On Windows, it also copies the runtime DLLs. You do not need to install TensorFlow into your Python environment to build the C++ examples.
The repository includes models/graph.pb. Run examples from the build directory so they can find its copy: ./Release/repeated_inference.exe on Windows or ./repeated_inference on Linux/macOS.
Pass these options to cmake -S . -B build:
-DHELLO_TF_BUILD_EXAMPLES=OFF: skip example executables.-DBUILD_TESTING=OFF: skip tests.-DTENSORFLOW_ROOT=/path/to/tensorflow -DHELLO_TF_FETCH_TENSORFLOW=OFF: use an existing wheel extraction. Headers must be under<root>/python/tensorflow/include, with native libraries under<root>/python/tensorflowor itspythonsubdirectory. CMake does not overwrite an external root.
OpenCV is optional. CMake builds and tests the OpenCV example when it finds the library.
- TensorFlow version, load a graph.
- Create a tensor, allocate a tensor, string tensors.
- Run a session, run a target operation.
- Helper API, batch inference, repeated inference.
- Image tensors, image files with OpenCV.
- Tensor information, graph information.
See tf_utils.hpp for declarations. Within this CMake project, link the helpers with:
target_link_libraries(your_target PRIVATE hello_tf_utils)For raw C API examples, use the project's target_link_tensorflow(your_target) function. There is no installable CMake package.
CreateTensor(dims, values)infers the TensorFlow type from the vector element type. UseCreateStringTensorfor strings.- Reader overloads taking a result reference return
TF_Codeand leave the result unchanged on error. They distinguish errors from valid empty values. - The checked
GetTensorShapewritesnulloptfor unknown rank, an empty vector for a scalar, and-1for unknown dimensions. RunSessionrequires output slots initialized tonullptr. Delete returned tensors and reset the slots before reuse.
- Prepare models: GraphDef, operation names, and checkpoints.
- Runtime and performance: resource reuse and measurement.
- Create a Windows import library: only for manual linking.