search

fgnt / sed_scores_eval

MIT License
26 stars 4 forks source link

get stuck while testing on a larger sed set #5

Open Ming-er opened 9 months ago

Ming-er commented 9 months ago

Hi, it works well on desed dataset or urbansed dataset. However, when I use it to evaluate models on audioset_strong_eval (around 15000+ wav files with 300 + sound classes) with psds2 score, it gets stuck with huge memory cost (> 150 GB). Are there any solutions or suggestions? Thanks.

JanekEbb commented 9 months ago

Hi, thanks for reporting. My guess is that it is because of cross-triggers which are tracked for all class pairs which may blow up memory when there are hundreds of classes. Does computing psds without cross triggers (alpha_ct=0,cttc_threshold=None) work?