My current project entails working with ML models and real-time audio, so when I discovered the C++ frontend to PyTorch my first instinct was to combine it with Cycling74’s Max SDK.

GET: GitHub

Installing LibTorch is easy enough, and compiling a test project in Visual Studio is just a matter of including the right folders and linking the .lib files1. So, you would think that applying this to a Max SDK example would be easy enough, right? Not so fast!

First, you need to set the project to compile as C++ code (obviously). Then, and this took me two nights to figure out, you also need to include “torch.h” before Max’s own headers (“ext.h” et al) and not after.

Anyway, it’s now on GitHub in two flavours: one is the standard C example from the SDK, and the other is the C++ bridge that I regularly use.2

Now, should you use these for development? Probably not, unless you’re some LibTorch and Max SDK wizard (in which case, what are you doing reading this? I should be learning from you!). So far I’ve only encountered one other instance of LibTorch being used with Max, by Philippe Esling &co @ IRCAM3. What I intend to do is design & train my model in good ol’ Python, and then load it into a Max external. And the next step would be doing some online training in Max, but we’re already getting ahead of ourselves.

  1. Also see this Medium post. ↩

  2. (edit Nov ‘20) I discovered the official C++ Development Kit for Max, called Min. Don’t see why it wouldn’t work with LibTorch, too: one more for the TODO pile then. ↩

  3. They too are still using standard PyTorch for main R&D though. ↩