Skip to content

README.md

Latest commit

 

History

History
83 lines (65 loc) · 3.5 KB

README.md

File metadata and controls

83 lines (65 loc) · 3.5 KB
Raw

gasm

Python rewrite of RARE (ch. 2) and RASE (ch. 3) from the thesis' equations, not a translation of matlab/. See the repo README for context and the thesis for the equations referenced in each module's docstring.

Install

pip install -e .[dev]

Tests

pytest
mypy --strict

Demos

examples/esc50_classification.py and examples/musdb18_separation.py run on public audio, since the thesis' own A-Volute training data no longer exists. The classification demo does reproduce the published score to within its uncertainty (see below); the separation demo only checks that DM-GMM and Def-MAP beat a naive "mixture as estimate" baseline.

Reproducing the thesis' RARE score

chapter2RTASC.tex, table:q reports 64.7 (2.9) correct classification on ESC-10 at quantization=257. esc50_classification.py measures 0.592 (0.023) over its 3 draws, about 1.5 combined standard deviations away.

Getting there needs three things that the manuscript does not spell out, each measured separately on ESC-10 at the thesis' parameters:

correct classification
kernel="multinomial", ESC-50's curated folds, no additive noise 0.290
kernel="cross_entropy" (what the MATLAB ran) 0.521 (0.031)
+ random 80/20 split over sounds (what split_dataset_folds.m does) 0.579 (0.028)
+ create_dataset.m's additive noise (gasm.rare.dataset) 0.592 (0.023)
thesis, table:q 0.647 (0.029)
  1. The kernel. The published code never quantizes: the rounding is commented out in compute_feature.m l.16-18 and identification.m l.51, so it scores a cross-entropy KDE rather than the multinomial PMF of eq. def_xi. The two agree in the quantization -> infinity limit, which is the convergence the manuscript itself invokes for table:q's plateau, and table:q reports the same 64.7 at 1e4 and 1e5. RareClassifier defaults to the manuscript's kernel and takes kernel="cross_entropy" for the other.
  2. The split. A random stratified 80/20 over sounds scores ~6 points above ESC-50's curated 5 folds. This is not a parent-recording leak: grouping the split by parent Freesound recording gives 0.582 (0.012), the same thing.
  3. The additive noise. create_dataset.m l.44-45 adds white noise at 1% of each signal's peak to every signal, train and test. It lifts the noise floor ~34 dB above the -60 dB energy gate (1.5% of frames gated instead of 23.8%) and leaves no zero bin in the spectra. gasm.rare.dataset.prepare_clip reproduces it, quirks included.

Two things the MATLAB gets away with, checked rather than assumed: main_monophonic.m l.120 uses an unnormalized frame count as prior_g and adds it raw to log-likelihoods, but every 5 s clip yields the same frame count (3432 per class here) so the constant cancels exactly; and the energy gate NaNs its rows without removing them, so it never enters that count either.

Datasets are large enough that they're best run outside a local machine (the datasets themselves are not committed to this repo):

pip install -e .[demos]

# ESC-50: https://github.com/karolpiczak/ESC-50 (zip download, ~600 MB)
python examples/esc50_classification.py /path/to/ESC-50-master

# MUSDB18 7s preview: auto-downloaded by musdb on first run (~140 MB)
MUSDB_ROOT=/path/to/musdb18-7s python examples/musdb18_separation.py

musdb depends on ffmpeg/ffprobe (not a Python dependency, install it separately, e.g. apt-get install ffmpeg).