Labeling and extracting data

Warning: This is still under development and not very extensible. Currently, only three labels are supported: A, B, and background.

Use this tool to create labels to train a custom model on any .wav data. After labeling is done, extract labels for training with beetles extract.

Using:

beetles label --wav_file <> --output_csv_path <>

Try it with

beetles label --wav_file resources/example.wav --output_csv_path test_labels.csv

Extracting

beetles extract --data_dir data_dir --output_data_path <> <OPTS>

This assumes a directory structure like below, where the .csv files contain labels for the .wav file in the same directory.

data_dir
   ├── M12F31_8_24
   │   ├── 2_M12F31_8_24.WAV
   │   └── 2_M12F31_8_24.csv
   ├── M14F15_8_7
   │   ├── 1_M14F15_8_7.WAV
   │   └── 1_M14F15_8_7.csv

The extraction script finds all first-level subdirectories in data_dir, then pairs together the .wav and .csv files in each subdirectory for extraction, whether or not their names match. An object that contains labels and features is then created in memory for each .wav/.csv pair. The entire object is shuffled and then split into test/train/validation data (7.5/85/7.5% splits, respectively). The test/train/validation data are saved in three separate directories: --data_dir/test, --data_dir/train, --data_dir/validation. The model training tool relies on this structure.

beetles extract can accept a few other arguments, like the number of ffts to use when calculating the spectrogram and whether or not to use a MelSpectrogram. Sensible defaults are provided.