microWakeWord-Trainer-Nvidia-Docker/cli/setup_audioset

#!/bin/bash
set -euo pipefail

PROGPATH=$(realpath "$0")
PROGDIR=$(dirname "${PROGPATH}")

source "${PROGDIR}/shell.functions"

if [ "${HELP}" == "true" ] ; then
    cat <<EOF >&2
Usage: $0 [ --cleanup-archives ] [ --cleanup-input-files ] [ --data-dir=<data_dir> ]

    --cleanup-archives    : Automatically clean up any downloaded archvies after
                            extraction.
    --cleanup-intermediate-files
                          : Automatically clean up the intermediate files after they've
                          : converted to 16k.
    <data_dir>            : Path to the data directory.
                          : Default: ${DATA_DIR}

EOF
    exit 1
fi

mkdir -p "${DATA_DIR}/training_datasets/downloads" || :
cd "${DATA_DIR}/training_datasets"

echo "***** Checking audioset *****"

AUDIO_URL="https://huggingface.co/datasets/agkphysics/AudioSet/resolve"
AUDIO_DIR="./audioset"
mkdir -p "${AUDIO_DIR}"
AUDIO16K_DIR="./audioset_16k"
mkdir -p "${AUDIO16K_DIR}"
AUDIO_FILECOUNT="./downloads/audioset_filecount"
AUDIO_IN_GLOB="*.flac"

declare -A filecounts
for i in {0..9} ; do
    fname="bal_train0${i}.tar"
    filecounts[${fname}]=0
done

get_filecounts filecounts "${AUDIO_FILECOUNT}"


REV_CANDIDATES=(
    "6762f044d1c88619c7f2006486036192128fb07e"
    "0049167e89f259a010c3f070fe3666d9e5242836"
    "ceb9eaaa7844c9ad7351e659c84a572e376ad06d"
    "main"
)

TAR_PATTERNS=(
    "data/bal_train0"
    "data/bal_train/bal_train0"
)

find_rev() {
    for rev in "${REV_CANDIDATES[@]}" ; do
        for pattern in "${TAR_PATTERNS[@]}" ; do
            url="https://huggingface.co/datasets/agkphysics/AudioSet/resolve/${rev}/${pattern}0.tar"
            curl -I -L --fail -s "${url}" > /dev/null && echo "${rev},${pattern}"
        done
    done
    echo ""
}

converter() {
    # shellcheck source=/dev/null
    source "${DATA_DIR}/.venv/bin/activate"

    python - "${AUDIO_DIR}" "${AUDIO16K_DIR}" <<-EOF
import os, sys
from pathlib import Path
from datetime import datetime, timezone

import numpy as np
import scipy.io.wavfile
import librosa

def write_wav(dst: Path, data: np.ndarray, sr: int):
    dst.parent.mkdir(parents=True, exist_ok=True)
    x = np.clip(data, -1.0, 1.0)
    scipy.io.wavfile.write(dst, sr, (x * 32767).astype(np.int16))

audioset_dir = Path(sys.argv[1])
audioset_out = Path(sys.argv[2])

flacs = list(audioset_dir.rglob("*.flac"))
total = len(flacs)
print(f"   FLAC files: {total}")
print("   Converting AudioSet → 16k mono WAV")
print("   Sit tight — this step can take a while.")
print("")

audioset_bad = []
ok = 0
skipped = 0

START = datetime.now(timezone.utc).replace(microsecond=0)

# Heartbeat interval (prints every N files)
HEARTBEAT_EVERY = 500

for idx, p in enumerate(flacs, start=1):
    try:
        outfile = audioset_out / (p.stem + ".wav")
        if outfile.exists():
            skipped += 1
        else:
            y, _ = librosa.load(p, sr=16000, mono=True)
            if y.size == 0:
                raise ValueError("empty audio")
            write_wav(outfile, y, 16000)
            ok += 1
    except Exception as e:
        audioset_bad.append(f"{p}:{e}")

    if idx == 1 or (idx % HEARTBEAT_EVERY) == 0 or idx == total:
        print(f"   Progress: {idx}/{total}  (ok={ok}, skipped={skipped}, failed={len(audioset_bad)})")

if audioset_bad:
    (audioset_out / "audioset_corrupted_files.log").write_text("\n".join(audioset_bad))

END = datetime.now(timezone.utc).replace(microsecond=0)
elapsed = END - START
print("")
print(f"   AudioSet complete ({ok} ok, {skipped} skipped, {len(audioset_bad)} failed)  Elapsed: {elapsed}")
EOF
}

expected_filecount=$(get_total_filecount filecounts)
actual_filecount=$(find "${AUDIO16K_DIR}" -name "*.wav" 2>/dev/null | wc -l) || :
write_filecount=false

# Option B behavior: if we already have output WAVs, don't re-download/re-extract/re-convert
if [ "${actual_filecount}" -ne 0 ] ; then
    echo "   Existing ${AUDIO16K_DIR} present (${actual_filecount} wav); skipping extract/convert"
else
    dl=$(find_rev)
    [ -n "$dl" ] || { echo "   Could not locate an AudioSet revision with FLAC tarballs still present on HF." ; exit 1 ; }
    rev=${dl%%,*}
    pattern=${dl##*,}

    echo "   Checking 10 tarballs"
    for i in {0..9} ; do
        fname="downloads/bal_train0${i}.tar"
        if [ ! -f "${fname}" ] ; then
            echo "   Downloading bal_train0${i}.tar"
            url="${AUDIO_URL}/${rev}/${pattern}${i}.tar"
            curl -L -s --fail "${url}" -o "${fname}" || { echo "Could not fetch ${fname} at rev ${rev}; continuing." ; continue ; }
        fi

        tarball_filecount=$(tar -tvf "${fname}" | wc -l )
        filecounts["bal_train0${i}.tar"]=${tarball_filecount}
        write_filecount=true

        echo "   Untarring bal_train0${i}.tar"
        tar -xf "${fname}" -C "${AUDIO_DIR}"
        if "${CLEANUP_ARCHIVES}" && [ -f "${fname}" ] ; then
            echo "   Cleaning up bal_train0${i}.tar"
            rm -rf "${fname}"
        fi
    done

    rm -rf "${AUDIO16K_DIR}/audioset_corrupted_files.log" || :
    converter

    # Recompute counts and warn (but do not fail)
    expected_filecount=$(get_total_filecount filecounts)
    actual_filecount=$(find "${AUDIO16K_DIR}" -name "*.wav" 2>/dev/null | wc -l) || :
    if [ "${actual_filecount}" -ne "${expected_filecount}" ] ; then
        echo "   Converted file count(${actual_filecount}) != expected file count(${expected_filecount})" >&2
        echo "   WARNING: mismatch is expected if some AudioSet files are corrupted; continuing." >&2
    fi
fi

if ${write_filecount} ; then
    write_filecounts filecounts "${AUDIO_FILECOUNT}"
fi

if "${CLEANUP_ARCHIVES}" ; then
    for i in {0..9} ; do
        fname="downloads/bal_train0${i}.tar"
        if [ -f "${fname}" ] ; then
            echo "   Cleaning up bal_train0${i}.tar"
            rm -rf "${fname}"
        fi
    done
fi

if "${CLEANUP_INTERMEDIATE_FILES}" && [ -d "${AUDIO_DIR}" ] ; then
    echo "   Cleaning up ${AUDIO_DIR}"
    rm -rf "${AUDIO_DIR}"
fi

echo "   Audioset complete"
exit 0