where is 2M audioset data and pretrain_audioset2M.sh?

[JHjang223]

Thank you meta for your hard work on the audioMAE implementation.
I want to train with 2M data, but in fact, audioset is only releasing features, so I couldn't get the data. I was finally able to get 20k data from another website. Where do I download the 2M data and I can't find pretrain_audioset2M.sh. Check please.

[Gariscat]

Same issue...... I checked the website and also only found the features instead of the original waveforms. How should we get the raw data or the raw data is not released at all?

[Jingerjia]

My stupid solution is:
Download the html of the class, and you'll find each video has it's youtube-id, start time, end time, and labels.
Then we can download every video we need by analyzing the html of the classes.
Good luck!

[IvanBirkmaier]

You can also use the .wav data which is provided by Huggingface: https://huggingface.co/datasets/confit/audioset-full or Baidu: https://pan.baidu.com/s/13WnzI1XDSvqXZQTS-Kqujg, password: 0vc2 (source: https://github.com/qiuqiangkong/audioset_tagging_cnn).

In the Hugginface dataset (eval) there is one broken file: ID YmW3... (if i remember right) delete this one it can cause headach :D
After downloading the data you have to create an train and eval json like they did in AST (https://github.com/YuanGongND/ast) (see egs/audioset/datafiles/sample_...) don't forget you just need audio an label!!!

must look like this:

    {
        "wav": "your path to wav file (doesn't have to be .flac file -> torchaudio supports both)",
        "labels": "/m/068hy,/m/07q6cd_,/m/0bt9lr,/m/0jbk"
    },

The label mapping for the wav-files/data can be done with the https://github.com/audioset/ontology and the provided CSV files (balanced_train_segments.csv, etc.) given on Audioset website: https://research.google.com/audioset/download.html

Good Luck