TensorFlow in C++ without the Bazel tax

What it does

CppFlow wraps the TensorFlow 2 C API in modern C++ so you can load Python-trained SavedModels, manipulate tensors, and run inference. You download the prebuilt C API library, link against it, and skip the multi-hour Bazel build entirely. The API handles JPEG decoding, casting, normalization, and eager execution in a few lines of code.

The interesting bit

The raw_ops.h header—covering (mostly) all TensorFlow raw ops—is auto-generated by a Python script. That’s the kind of boring automation that saves you from manually wrapping hundreds of C functions with void pointers and manual memory management.

Key highlights

• Header-only C++ wrapper; no TensorFlow source build required
• Auto-generated facade over TensorFlow’s raw C ops
• RAII-style tensor and model classes (no manual TF_DeleteTensor)
• Eager execution and SavedModel loading supported
• Examples include full image-inference pipeline

Caveats

• Requires separate download of the TensorFlow C API binary; not fully self-contained
• README notes the project is “basically a wrapper”—don’t expect training or graph-building features
• Ops coverage is “(mostly) all” per the README; exact gaps unclear

Verdict

Worth a look if you need to deploy TensorFlow models in a C++ environment and can’t stomach Bazel. Skip it if you need training, custom ops, or a fully hermetic build.