Home

Awesome

hear-preprocess

Dataset preprocessing code for the HEAR Benchmark and for all the tasks used during the 2021 HEAR NeurIPS challenge. To find out more about HEAR please visit https://hearbenchmark.com.

Unless you need to pre-process HEAR benchmark tasks yourself or want to contribute a task, you won't need this repo. Use hear-eval-kit to evaluate your embedding models on these tasks.

Pre-processed datasets (at 48000Hz) for all HEAR Benchmark tasks are available on zenodo. Other sampling rates (16000, 22050, 32000, 44100), are available for download (requester pays) from Google Storage gs://hear2021-archive/tasks/

This preprocessing is slow and disk-intensive but safe and careful.

Cloud Usage

See hear-eval's README.spotty for information on how to use spotty.

Installation

pip3 install hearpreprocess

Tested with Python 3.7 and 3.8. Python 3.9 is not officially supported because pip3 installs are very finicky, but it might work.

Development

Clone repo:

git clone https://github.com/hearbenchmark/hear-preprocess
cd hear-preprocess

Install in development mode:

pip3 install -e ".[dev]"

Make sure you have pre-commit hooks installed:

pre-commit install

Running tests:

python3 -m pytest

Preprocessing

You probably don't need to do this unless you can't use the available pre-processed datasets and need to preprocess the data yourself..

If you want to run preprocessing yourself:

These Luigi pipelines are used to preprocess the evaluation tasks into a common format for downstream evaluation.

To run the preprocessing pipeline for all available tasks, with all available modes for each task:

python3 -m hearpreprocess.runner all --mode all

You can instead just call a specific single task

python3 -m hearpreprocess.runner task1 --mode all

or specific multiple tasks:

python3 -m hearpreprocess.runner task1 task2 --mode all

Tasks

List of available tasks used in HEAR 2021:

Task NameModes
dcase2016_task2full
nsynth_pitch5h, 50h
speech_commands5h, full
beehive_states_fold05h, full
beehive_states_fold15h, full
beijing_operafull
esc50full
fsd50kfull
gunshot_triangulationfull
libricountfull
maestro5h
mridangam_strokefull
mridangam_tonicfull
tfds_crema_dfull
tfds_gtzanfull
tfds_gtzan_music_speechfull
vocal_imitationfull
vox_lingua_top10full

Pipelines

Each pipeline will download and preprocess each dataset according to the following DAG:

In terms of sampling:

These commands will download and preprocess the entire dataset. An intermediary directory defined by the option luigi-dir(default _workdir) will be created, and then a final directory defined by the option tasks-dir (default tasks) will contain the completed dataset.

Options:

Options:
  --num-workers INTEGER  Number of CPU workers to use when running. If not
                         provided all CPUs are used.
  --sample-rate INTEGER  Perform resampling only to this sample rate. By
                         default we resample to 16000, 22050, 44100, 48000.
  --tmp-dir TEXT         Temporary directory to save all the intermediate
                         tasks (will not be deleted afterwords). (default:
                         _workdir/)
  --tasks-dir TEXT       Directory to save the final task output (default:
                         tasks/)
  --tar-dir TEXT         Directory to save the tar'ed output (default: .)
  --mode TEXT            default, all, or small mode for each task.
  --help                 Show this message and exit.

To check the stats of an audio directory:

python3 -m hearpreprocess.audio_dir_stats {input folder} {output json file}

Stats include: audio_count, audio_samplerate_count, mean meadian and certain (10, 25, 75, 90) percentile durations. This is helpful in getting a quick glance of the audio files in a folder and helps in decideing the preprocessing configurations.

The pipeline will also generate some stats of the original and preprocessed data sets, e.g.:

speech_commands-v0.0.2/01-ExtractArchive/test_stats.json
speech_commands-v0.0.2/01-ExtractArchive/train_stats.json
speech_commands-v0.0.2/03-ExtractMetadata/labelcount_test.json
speech_commands-v0.0.2/03-ExtractMetadata/labelcount_train.json
speech_commands-v0.0.2/03-ExtractMetadata/labelcount_valid.json

Faster preprocessing, for development

The small flag runs the preprocessing pipeline on a small version of each dataset stored at Downsampled HEAR Open Tasks. This is used for development and continuous integration tests for the pipeline.

These small versions of the data can be generated deterministically with the following command:

python3 -m hearpreprocess.sampler <taskname>

NOTE : --mode small is used to run the task on a small version of the dataset for development.

Breaking change for hear-eval

If the open tasks have changed enough to break the downstream CI, (for example in the heareval repo), the Preprocessed Downsampled HEAR Open Tasks should be updated. An example of an obvious breaking changes can be modification of the task configuration.

The version should be bumped up in hearpreprocess/__init__.py and the pipeline should be run for the open tasks with --mode small flag

Thereafter, the following command can be used to copy the tarred files produced by running the pipeline for the open tasks to the repo( Please clone the repo )

git clone git@github.com:hearbenchmark/hear2021-open-tasks-downsampled.git
cp hear-LATEST-speech_commands-v0.0.2-small-44100.tar.gz ./hear2021-open-tasks-downsampled/preprocessed/
cp hear-LATEST-nsynth_pitch-v2.2.3-small-44100.tar.gz ./hear2021-open-tasks-downsampled/preprocessed/
cp hear-LATEST-dcase2016_task2-hear2021-small-44100.tar.gz ./hear2021-open-tasks-downsampled/preprocessed/
cp hear-2021.0.6-speech_commands-v0.0.2-small-44100.tar.gz ./hear2021-open-tasks-downsampled/preprocessed/
cp hear-2021.0.6-nsynth_pitch-v2.2.3-small-44100.tar.gz ./hear2021-open-tasks-downsampled/preprocessed/
cp hear-2021.0.6-dcase2016_task2-hear2021-small-44100.tar.gz ./hear2021-open-tasks-downsampled/preprocessed/