Fast-Audioset-Download

Download audioset data super fastly with youtube-dl and Python Multiprocessing. Major difference with other repos is that my repo utilizes ffmpeg which supports "Partial" youtube download when provided to youtube-dl as an external downloader, while others rely on default downloader that downloads full video and then clip it.

Audioset(Weak) is basically a multi-labelled dataset. Weirdly enough, many repositories tend to store them in a class-folder fashion which is more suitable for single-labelled dataset. Thus, I chose to take a different approach where this repo stores all of the downloaded data in a single folder and create metadata json file containing information. Note that the csv files provided by google already contain metadata, but one has to match that information to the downloaded files every time.

This repo only downloads audio(dataset size issue)!! If you want to download as a video, Follow instructions below.

After the download, files will be stored under the 'wavs' folder.

One can acquire each file's information(including url, stored path(relative), tags, labels interpreted from tags ...) by parsing the metadata json file.

It would be very nice of you to put a star★ if you liked my code :)

Required Packages

youtube-dl(outdated), yt-dlp(new!), tqdm, ffmpeg(*Allows partial youtube download)

Cookies.txt

Cookies are required to suppress the auth related warnings(+erros) of youtube-dl(reference) You can download a cookies.txt by

1. Run firefox web browser on the desired machine(PC, or Server)
2. Install firefox extension
3. Go to youtube
4. Make sure you are logged out(Creating cookie while logged in might cause severe instability to your youtube streaming elsewhere)
5. Download cookie to desired location.
6. Modify donwload.py specifying the path of your downloaded cookie file

How to Use

1. First download unbalanced_train_segments.csv from http://storage.googleapis.com/us_audioset/youtube_corpus/v1/csv/unbalanced_train_segments.csv
2. Move the downloaded csv file under csvs folder
3. Run download.py

foo@bar:/path/to/this/repo $ python download.py

** Resume supported(kind of) in case of crash or keyboard interrupt: Only acquire video meta-information(for creating final metadata.json) without downloading actual audio.

Downloading in a video(mp4) format

youtube-dl and ffmpeg also support video downloading. Naturally, my repo can do that with a simple change in the code.

1. Change the 'ext' global variable from m4a(default) to mp4 in the download.py
2. Change the ydl_opts on the line 81 like below.

ext = 'mp4'
...
   ydl_opts = {
               "logger": ytdl_logger,
               'cookiefile' : f'temps/id_{ids}/cookies.txt',
               'ignoreerrors': True,
               'outtmpl':"temps/id_%(id)s/audio.%(ext)s",
               'quiet' : True,
               'format': 'bestvideo[ext=mp4]+bestaudio[ext=m4a]/bestvideo+bestaudio',
               'postprocessor_args': ['-ar', str(sample_rate)],
               'external_downloader':'ffmpeg',
               'external_downloader_args':['-ss',st, '-to',dur, '-loglevel', 'quiet']
           }
...