Issues with baseline PeakFP

[shenberg]

Hi, I'm trying to use PeakFP to the best of my understanding. I took 35 songs (roughly 120 minutes of audio), ran extractor.py on each one, then ran indexer.py in order to generate an index (I provided a txt file containing the paths to all the fingerprint files). The generated index is ~20MB. Next, I tried to run matcher.py on an 8-second query, however the script never finishes.

If I break it after a long time (e.g. >10 minutes), it's busy adding to the interval tree itree.addi(match['start'], match['end'], match). I let it run for >3h and it still did not finish a single 8-second query.

I reduced sensitivity from 5 matches minimum to 12 matches, which reduces the amount of matches by a factor of x10, and even then it gets stuck at itree.split_overlaps() - I gave up after some 10s of minutes, as from what was reported in the paper, I was expecting a runtime that's only, say, a x10 factor slower than real-time, but I'm getting well over x100.

Any ideas why? Am I using the library incorrectly?

[guillemcortes]

Hi @shenberg,
Thanks for your interest in baf, right now I am a bit busy with other things but I will try to take a look into it.
In the meantime can you provide some information to reproduce your error? like Os platform, python and package versions, exact commands you are running, etc.
Many thanks!

[shenberg]

Sure, thanks for the response!

I'm running on an M1 max Macbook pro with MacOS Ventura 13.6.2, installed PeakFP in a virtualenv with Python 3.8.18.

i had to change the requirements a bit in order to get a functional install - older versions of numpy don't work smoothly with ARM-based macs so I changed the requirements file to be:

intervaltree==3.1.0
librosa==0.8.1
scikit-image

And then everything installed smoothly.

I don't know if it's normal, but for me, run_matching() in matcher.py would end up with on the order of 300k matches in processed_matches with the default settings, then spend an incredibly long time in the interval tree. I tried modifying the code to create an interval tree per reference id, and then to coalesce matches between references, but I still ended up having to reduce the sensitivity by a ton (min_peak_threshold 13 instead of 5 where necessary) and results were pretty bad...

[guillemcortes]

PeakFP takes a lot of time to compute the matching. Please refer to the paper for more information on that: https://zenodo.org/records/7372162. Basically it took 98hours to do the matching on BAF (single thread). The results do not match the metrics reported in the paper, you mean? In the end PeakFP was designed to be a naive, non-scalable baseline for background music identification.