So in summary: subprocess calls appear to break clearML tracking, even if I do Task.init() in both main.py and train.py. However the script does run end to end successfully. If I remove the subprocess calls, I only need Task.init() in main.py for everything to work (scalars, reporting, etc).