Hello Pytorch lovers!
I am facing some problems porting c++ torch code to low-level embedded devices for very level on-board training on RIOT platform. It would be nice if somebody in the community help me in addressing following issues:
-
PyTorch is using a lot of exceptions. For RIOT this is an issue as a CPP exception needs dynamic allocation and on embedded this causes issues with the low amount of memory available. RIOT compiles CPP code with -fno-exceptions because of this and pytorch doesn’t compile with this compiler flag enabled.
-
Multi-threading is heavily used in PyTorch with the threadpool. On the bare metal target this is not available and usually not useful because of the single core nature of the system.
Hardware used: Arduino ble 33
Thanks in advance!