pydra
pydra copied to clipboard
nipype
reduce pydra overhead
in the senselab package which processes audio (mostly through python functions), we are seeing significant pydra overhead. to try and debug a bit of this, i looked at scalene (github.com/plasma-umass/scalene) and it seems like we could incorporate it into the codebase, or use it to optimize bits of it. there is an ai model integration into scalene.
As part of this issue, one of the other maintainers of senselab tried to do some performance testing and found the following when using 535 EMO-DB short clips
| Task | With Pydra (s) | Without (s) |
|---|---|---|
| Read audios | 12.5 | 0.5 |
| Save audios | 10.63 | 0.2 |
| Downmix to mono | 15.0 | 0.1 |
| Resample | 16.0 | 0.1 |
| OpenSMILE features | 33.0 | 20.0 |
| Praat/parselmouth features | 93.7* | 246.1 |
| Torchaudio features | 37.8 | 16.2 |
| Audiomentations (gain) | 13.5 | 0.3 |
| Torchaudiomentations (gain) | 5.3 | 0.2 |
*on Pydra 0.23 it was even faster (~70s)
