Audio classification

Readers and writers for audio classification have the -ac suffix.

Plugins

sub-dir to ADAMS

The following converts an audio classification dataset from the sub-dir format (sub-directory names represent the audio classification labels) into the ADAMS format, which stores the label in an associated .report file (Java properties file):

adc-convert -l INFO \
  from-subdir-ac \
    -l INFO \
    -i ./subdir/ \
  to-adams-ac \
    -l INFO \
    -o ./adams \
    -c classification

sub-dir (randomized train/val/test splits)

By enforcing batch-processing --force_batch and using the randomize-records filter, randomized train/val/test splits (writers typically support generating splits) can be generated like this:

adc-convert -l INFO --force_batch \
  from-subdir-ac \
    -l INFO \
    -i ./subdir/ \
  randomize-records \
    -s 42 \
  to-subdir-ac \
    -l INFO \
    -o ./subdir-split \
    -c classification \
    --split_names train val test \
    --split_ratios 70 15 15