turpaultn/desed

Repo associated to the DESED dataset, download and creation of data

desed's Issues

Two problems when downloading

OS

Linux

Environment:

python 3.6.6
desed 1.3.7

1. The first problem when running

desed.download_real("./data/dataset")

There comes an info:
INFO - desed.download/_download_audioset_file - Input signal length=0 is too small to resample from 48000->44100

2. The second when running

desed.download_desed_soundbank("./data/soundbank")

There comes a error:
shutil.ReadError: .../DESED/tmp_zer1y99k/tmp_w69ygmt..zip is not a zip file
This error may because a filename contains '..' that leads a failure when recognizing its file type.

conda list

This file may be used to create an environment using:
$ conda create --name --file
platform: linux-64
_libgcc_mutex=0.1=main
_openmp_mutex=5.1=1_gnu
appdirs=1.4.4=pypi_0
attrs=22.1.0=pypi_0
audioread=3.0.0=pypi_0
brotli=1.0.9=pypi_0
ca-certificates=2022.10.11=h06a4308_0
certifi=2021.5.30=py36h06a4308_0
cffi=1.15.1=pypi_0
charset-normalizer=2.0.12=pypi_0
cycler=0.11.0=pypi_0
dcase-util=0.2.20=pypi_0
decorator=5.1.1=pypi_0
desed=1.3.7=dev_0
future=0.18.2=pypi_0
idna=3.4=pypi_0
importlib-metadata=4.8.3=pypi_0
importlib-resources=5.4.0=pypi_0
jams=0.3.4=pypi_0
joblib=1.1.1=pypi_0
jsonschema=3.2.0=pypi_0
kiwisolver=1.3.1=pypi_0
libedit=3.1.20210910=h7f8727e_0
libffi=3.2.1=hf484d3e_1007
libgcc-ng=11.2.0=h1234567_1
libgomp=11.2.0=h1234567_1
librosa=0.9.2=pypi_0
libstdcxx-ng=11.2.0=h1234567_1
llvmlite=0.36.0=pypi_0
matplotlib=3.3.4=pypi_0
mir-eval=0.7=pypi_0
mutagen=1.45.1=pypi_0
ncurses=6.3=h5eee18b_3
numba=0.53.1=pypi_0
numpy=1.19.5=pypi_0
openssl=1.0.2u=h7b6447c_0
packaging=21.3=pypi_0
pandas=1.1.5=pypi_0
pillow=8.4.0=pypi_0
pip=21.2.2=py36h06a4308_0
pooch=1.6.0=pypi_0
pycparser=2.21=pypi_0
pycryptodomex=3.15.0=pypi_0
pydot-ng=2.0.0=pypi_0
pyloudnorm=0.1.0=pypi_0
pyparsing=3.0.7=pypi_0
pyrsistent=0.18.0=pypi_0
python=3.6.6=h6e4f718_2
python-dateutil=2.8.2=pypi_0
python-magic=0.4.27=pypi_0
pytz=2022.6=pypi_0
pyyaml=6.0=pypi_0
readline=7.0=h7b6447c_5
requests=2.27.1=pypi_0
resampy=0.4.2=pypi_0
scaper=1.6.5=pypi_0
scikit-learn=0.24.2=pypi_0
scipy=1.5.4=pypi_0
setuptools=58.0.4=py36h06a4308_0
six=1.16.0=pypi_0
sortedcontainers=2.4.0=pypi_0
soundfile=0.11.0=pypi_0
sox=1.4.0=pypi_0
soxbindings=1.2.3=pypi_0
sqlite=3.33.0=h62c20be_0
threadpoolctl=3.1.0=pypi_0
tk=8.6.12=h1ccaba5_0
tqdm=4.64.1=pypi_0
typing-extensions=4.1.1=pypi_0
urllib3=1.26.12=pypi_0
validators=0.20.0=pypi_0
websockets=9.1=pypi_0
wheel=0.37.1=pyhd3eb1b0_0
xz=5.2.6=h5eee18b_0
yt-dlp=2022.7.18=pypi_0
zipp=3.6.0=pypi_0
zlib=1.2.13=h5eee18b_0

Validation data has overlaps, can't compute PSDS. Some labels are incorrect.

Hey there,
so I rewritten most of the baseline code for the DCASE2021 Challenge by myself and couldn't compute any PSDS related score.

The minimal code would be:

from metrics import compute_psds_from_operating_points
psds_score_scenario1 = compute_psds_from_operating_points(
      None, # Will throw nontheless an erro since Ground truth is incorrect.
      'metadata/validation.tsv,
      'metadata/validation_duration.tsv',
      dtc_threshold=0.7,
      gtc_threshold=0.7,
      alpha_ct=0,
      alpha_st=1,
      save_dir='my_output_dir'.
)

the function above is defined in here

Executing this throws an error:

psds_eval.psds.PSDSEvalError: The ground truth dataframe provided has intersecting events/labels for the same class.

After a short debug, I checked which files are causing the error. The four files are:

Y2DF8mtDke8k_460.000_470.000.wav
Y5qx1HGIWJww_170.000_180.000.wav
Y6PwBeL3Abl8_60.000_70.000.wav
Y4trGKbbTmC4_30.000_40.000.wav

After checking the metadata, it seems that there are mistakes:

Y2DF8mtDke8k_460.000_470.000.wav	0.000	10.000	Frying
Y2DF8mtDke8k_460.000_470.000.wav	1.218	1.468	Dishes
Y2DF8mtDke8k_460.000_470.000.wav	1.893	2.263	Dishes
Y2DF8mtDke8k_460.000_470.000.wav	2.827	7.980	Speech
Y2DF8mtDke8k_460.000_470.000.wav	3.110	3.592	Dishes
Y2DF8mtDke8k_460.000_470.000.wav	*3.797*	*4.103*	Dishes
Y2DF8mtDke8k_460.000_470.000.wav	*3.797*	*4.492*	Dishes
Y2DF8mtDke8k_460.000_470.000.wav	*4.192*	*4.492*	Dishes
Y2DF8mtDke8k_460.000_470.000.wav	*7.980*	*8.230*	Dishes
Y2DF8mtDke8k_460.000_470.000.wav	*7.980*	*8.617*	Dishes
Y2DF8mtDke8k_460.000_470.000.wav	*8.367*	*8.617*	Dishes
Y2DF8mtDke8k_460.000_470.000.wav	8.797	10.000	Dishes

Y6PwBeL3Abl8_60.000_70.000.wav	0.000	1.318	Speech
Y6PwBeL3Abl8_60.000_70.000.wav	2.794	4.535	Speech
Y6PwBeL3Abl8_60.000_70.000.wav	4.129	4.453	Blender
Y6PwBeL3Abl8_60.000_70.000.wav	*4.724*	*5.335*	Blender
Y6PwBeL3Abl8_60.000_70.000.wav	*4.724*	*6.076*	Blender
Y6PwBeL3Abl8_60.000_70.000.wav	*5.400*	*6.076*	Blender
Y6PwBeL3Abl8_60.000_70.000.wav	6.635	9.118	Blender
Y6PwBeL3Abl8_60.000_70.000.wav	9.435	9.882	Blender

Should I just delete the "segmented" instances and keep the "long" instances i.e. in the example above, keep 4.724 - 6.076 ?

Download real data

Hi.

When I run create_dcase2019_dataset.sh, I got following this errors.

0 Y8lhDSLRfZ8Q_350.000_360.000.wav ERROR: This video is unavailable.\n...
1 Y8Y3du2Iw0v4_60.000_70.000.wav ERROR: This video is unavailable.\n...
2 Y97oxHi3qxs8_0.000_10.000.wav ERROR: This video is unavailable.\n...
3 Y99MfwDS1Vpc_30.000_40.000.wav ERROR: This video is unavailable.\n...
4 Y9aNXDP2rv7o_220.000_230.000.wav ERROR: This video is unavailable.\n...
... ... ...
1163 Y-4pmCrSdMhg_30.000_40.000.wav ERROR: This video is unavailable.\n...
1164 YQb6N3HvbmgQ_580.000_590.000.wav ERROR: This video is unavailable.\n...
1165 Y5iARe0XPQdg_95.000_105.000.wav ERROR: This video is unavailable.\n...
1166 YmAg1_aL-fCY_30.000_40.000.wav ERROR: This video is unavailable.\n...
1167 YxgQjEiwZsAk_146.000_156.000.wav ERROR: This video is unavailable.\n...

How can I fix it?

Downloads extremely slow suddenly

Hello,

downloading the audioset data via youtubeDL within this repo suddenly became very slow. Can you confirm it is an issue on youtube's side?

Best regards

Effective FP Rate not working for single class PSDS

Hello,

as I'm working an a single class sound event detection repository, I encountered the problem that using the effective false positive rate is corrupted and therefor both ROC curve and overall PSDS score cannot be computed (remain zero).

In psds.py, line 939 I changed
psd_roc = self._effective_tp_ratio(tpr_efpr_curve, alpha_st) to psd_roc = self._effective_tp_ratio(tpr_fpr_curve, alpha_st) to sovle the issue. It is possible in a single class problem because there are obviously no cross trigger effect that have to be accounted for, so that the the eFPR shoudl be the FPR actually. This does not seem to be the case.

Edit: This is not solving the bug of course, it's just mitigating it.

pip install -e in python environment links all project

While running
pip install -e .
it creates the line:
/Users/nturpaul/Documents/Seafile/DCASE/Desed/src in "site-packages/easy-install.pth"

It is a problem since we advise to put the data in real/, synthetic/, dataset/at the root of the project.

So then, all files are indexed in an IDE like pycharm from the root folder "Desed"

module 'desed' has no attribute 'download_real

When I wanted to use desed to generate synthetic database, there have some Attribute Error: module 'desed' has no attribute 'download_real' and module 'desed' has no attribute 'download_desed_soundbank'

Wrong Import of FastChildWatcher

Greetings, I'm trying to download the real sounds dataset, but for the life of me I can not , because the following error arises:
cannot import name 'FastChildWatcher' from 'asyncio'
I'm using pycharm as IDE, with python 3.9 and used the project interpreter to install desed library. asyncio version is 3.4.3
The code is in the following image: