Got an itch to run a small model with the C++ side libtorch for practice — partly didn’t want to install a bunch of Python environments, partly thought it’d make deployment easier down the line. Turned out just setting up the environment ate half a day, and the resources I could find were way fewer than on the Python side.
Want to ask the veterans who’ve actually trained something with it: is using it for small-scale training just asking for trouble, would it be more realistic to just train honestly in Python and export for C++ inference.
If it’s purely for practice, is it even worth grinding through the training path?