Skip to content

Latest commit

 

History

History
39 lines (33 loc) · 1.12 KB

README.md

File metadata and controls

39 lines (33 loc) · 1.12 KB
Raw

Data Preparation Pipeline

In this work, we use two datasets: VGGSound and AudioSet. Here we provide the pipeline to pre-process VGGSound data. The process for AudioSet is similar.

On VGGSound, pre-process includes:

  • Convert .mp4 file to .wav file
  • Extract waveform (numpy.array) into hdf5 file
  • Create index on hdf5 file and distribute data

Pipeline scripts

  • Assume .mp4 files are store at vggsound/videos and .wav will be stored in vggsound/audios
cd vggsound
./tools/convert_mp4_to_wav.sh  \
    ./videos/ \
    ./audios/ \
    32000
  • Extract waveform and store in hdf5 file, with vggsound.csv downloaded here:
./tools/extract_waveform_in_h5.sh \
    ./audios/ \
    train \
    32000 \ # sample rate
    10 \ # 10 seconds
    vggsound.csv \
    features/vggsound_train.hdf5
  • Build index and distribute data, data will be in features/Train
python tools/create_index_distribute_data.py \
    --input_h5 features/vggsound_train.hdf5 \
    --index_file features/vggsound_train_index.hdf5 \
    --data_dir Train \
    -n 20