A simple time and memory Python profiler

Installation

pip install tmprofile

Usage

1. As a context manager

from tmprofile import Profile

with Profile() as profiler:
    f(...) # some intensive code
# profile is automatically saved as a json file.
# display elapsed time, RAM & VRAM consumption evolution
profiler.print()
profiler.plot()

2. As a decorator

from tmprofile import profile

@profile()
f(...) # some intensive code
# profile is automatically saved as a json file.

Parameters

• name (str): Name of the profiler. Will be used to name the saved file. Required for context manager. For decorator, will use fonction' name if no name is provided.
• verbose (bool): If True, will print logged values at the end of profiling (i.e. on exit as a context manager, on function's return as a decorator). Defaults to False.
• gpu_idx (int): Which gpu's VRAM consumption has to be monitored. Default to 0.
• time_delta (float): Time to wait (in second) between two RAM / VRAM consumption logging. Defaults to 0.2.
• output_dir (str | Path): Where to write the logged values (as json). Defaults to .tmprofile.
• filename (str | Path, optional): Name of the saved json file. If no name is given, will used the profiler name (see name above).

Saved logs

Each time it is called, the profiler will saved a file.
The filepath will be <output_dir>/filename.json. By default, <output_dir> is .tmprofile/ and <filename> is:

• <name>.json if name was specified.
• profile.json if the profiler is a context manager and name was NOT specified.
• <module>::<function>.json if the profiler is a decorator and name was NOT specified.

Interrupt

Since this profiler runs in a separate process, it will gracefully shutdown if the main programm is interrupted.

Disclaimer

The core of this code (asynchronous RAM/VRAM logging) is taken from Thomas Rouch's code here.