Feature Request: C++ logging from non-TensorFlow frameworks

TensorBoard logging can be done from a variety of ML frameworks, including PyTorch with torch.utils.tensorboard. The FileWriter and protos exposed from TensorBoard provide an API contract that can be easily accessed in Python code to log everything from scalars and histograms to model graphs.

However, things get complicated with C++. Any proto specifications or code would need to be built as a static or dynamic library and referenced by frameworks at build as well as runtime.

A couple of questions:

1. How much interest is there in the community for TensorBoard logging from non-Python languages and non-TensorFlow libraries?
2. Should proto files exist in a separate TensorFlow/TensorBoard repository so they can be pinned, updated, etc. across logging implementations or just be copied as needed into other code bases?
3. Is there value in a shared C++ logging implementation and a precompiled library to match?

We are exploring C++ logging for PyTorch with libtorch and so wanted to get community involvement on this topic. Comments welcome!

cc @nfelt @manivaradarajan @GalOshri @natalialunova @lanpa @sanekmelnikov @caraya10 @jspisak

[jackshi0912]

I work in RL with various search algorithms (including beam search and MCTS) and use C++ for performance. We currently use custom hacked together frameworks. TensorBoard logging for Libtorch would be absolutely awesome!

[jackshi0912]

This is really good feedback. I feel like there are two motivations here. 1) Integration within an already implemented C++ workflow; and 2) Performance. Sounds like both are important and critical but the former is the must have?

“Performance” in this context refers to python’s terrible performance. There is not really a fix for that. Even python wrappers for C++ code is much slower than cache optimised / e2e compile time optimised native C/C++. For example, writing a MCTS algorithm in python using python data structures in my case is ~3-4 orders of magnitude slower than the most naive C++ implementation (heap allocate everything, don’t pack struct or cache lines, no vectorised operations). If you optimise C++ code and use thread parallelisation (can’t do in python with GIL), the performance gap widens even more.

I think exposing a C++ or proto interface would be really helpful.