Python code to load the TUT Acoustic scenes 2016 dataset.

tutacousticscenes2016.py

"""TUT Acoustic scenes 2016 dataset.
"""
import os
import sys
import numpy as np
import zipfile
import librosa
from tqdm import tqdm
from keras.utils.data_utils import get_file


def __load_files(fileset, fileset_path, labels, n=None):
    # source_sample_rate = 44100
    # source_len = 30 * 44100 + 1 # 30 seconds audio files
    load_sample_rate = 16000
    n_samples_per_file = 30 * load_sample_rate + 1

    if n is None:
        n = len(fileset)

    print("Loading wav files to " + fileset_path)

    # Allocate return buffers
    x = np.empty([n, n_samples_per_file], dtype='float32', order='C')
    y = np.empty(n, dtype='int16')

    for i in tqdm(range(n)):
        meta = fileset[i]
        wav_filename = os.path.join(fileset_path, meta[0])
        label = meta[1]
        # Load audio file as a floating point time series.
        # Data is converted from:
        # int, 44.1kHz stereo to
        # normalized float, 16kHz, mono
        # @note: resampling takes time!! (~1.4s per call)
        x[i], _ = librosa.load(wav_filename, sr=16000, mono=True, dtype=np.float32)
        # x[i], _  = librosa.load(wav_filename, sr=44100, mono=True, dtype=np.float32) # no resampling
        y[i] = labels[label]

    return (x, y)


def __content_valid(infolist, destination):
    for file in infolist:
        # ignore directories
        if file.filename.endswith('/'):
            continue
        # check filesize match for speed (or check CRC)
        path = os.path.join(destination, file.filename)
        try:
            size = os.path.getsize(path)
            if size != file.file_size:
                return False
        except OSError:
            return False

    return True


def load_data():
    """Loads the TUT Acoustic scenes 2016 dataset.
    # Returns
    Tuple of Numpy arrays: `(x_dev, y_dev), (x_eval, y_eval)`.
    """
    dirname = os.path.join('datasets', 'TUT-acoustic-scenes-2016')
    datadir_base = '.'  # current directory
    datadir = os.path.join(datadir_base, dirname)
    dev_fileset_path = os.path.join(datadir, 'TUT-acoustic-scenes-2016-development')
    eval_fileset_path = os.path.join(datadir, 'TUT-acoustic-scenes-2016-evaluation')
    npz_dev_filepath = os.path.join(datadir, 'development_set.npz')
    npz_eval_filepath = os.path.join(datadir, 'evaluation_set.npz')
    dev_meta_filepath = os.path.join(dev_fileset_path, 'meta.txt')
    eval_meta_filepath = os.path.join(eval_fileset_path, 'meta.txt')

    # Development dataset (7.5GB)
    dev_base = 'https://zenodo.org/record/45739/files/'
    dev_files = [
        'TUT-acoustic-scenes-2016-development.audio.1.zip',  # 1.1GB
        'TUT-acoustic-scenes-2016-development.audio.2.zip',  # 1.1GB
        'TUT-acoustic-scenes-2016-development.audio.3.zip',  # 1.1GB
        'TUT-acoustic-scenes-2016-development.audio.4.zip',  # 1.1GB
        'TUT-acoustic-scenes-2016-development.audio.5.zip',  # 1.1GB
        'TUT-acoustic-scenes-2016-development.audio.6.zip',  # 1.1GB
        'TUT-acoustic-scenes-2016-development.audio.7.zip',  # 1.1GB
        'TUT-acoustic-scenes-2016-development.audio.8.zip',  # 528.5MB
        'TUT-acoustic-scenes-2016-development.doc.zip',      # 69.7KB
        'TUT-acoustic-scenes-2016-development.error.zip',    # 1.3KB
        'TUT-acoustic-scenes-2016-development.meta.zip'      # 28.8KB
    ]

    # Evaluation dataset (2.5 GB)
    eval_base = 'https://zenodo.org/record/165995/files/'
    eval_files = [
        'TUT-acoustic-scenes-2016-evaluation.audio.1.zip',  # 1.1GB
        'TUT-acoustic-scenes-2016-evaluation.audio.2.zip',  # 1.1GB
        'TUT-acoustic-scenes-2016-evaluation.audio.3.zip',  # 538.9MB
        'TUT-acoustic-scenes-2016-evaluation.doc.zip',      # 69.2KB
        'TUT-acoustic-scenes-2016-evaluation.meta.zip'      # 6.0KB
    ]

    # 3 classes : 0 indoor, 1 outdoor, 2 vehicle
    # http://www.cs.tut.fi/sgn/arg/dcase2016/acoustic-scenes
    labels = {
        'cafe/restaurant':  0,  # indoor
        'grocery_store':    0,  # indoor
        'library':          0,  # indoor
        'metro_station':    0,  # indoor
        'office':           0,  # indoor
        'home':             0,  # indoor
        'city_center':      1,  # outdoor
        'forest_path':      1,  # outdoor
        'beach':            1,  # outdoor
        'residential_area': 1,  # outdoor
        'park':             1,  # outdoor
        'car':              2,  # vehicle
        'bus':              2,  # vehicle
        'train':            2,  # vehicle
        'tram':             2,  # vehicle
    }

    # Check if data has already been loaded into NumPy or load from .wav files
    # warning: parameter n (number of loaded files) is not considered
    if os.path.exists(npz_dev_filepath):
        with np.load(npz_dev_filepath) as npzfile:
            x_dev = npzfile['arr_0']
            y_dev = npzfile['arr_1']
            print("Loaded development set from cache")
            dev_set_loaded = 1
    else:
        dev_set_loaded = 0

    if os.path.exists(npz_eval_filepath):
        with np.load(npz_eval_filepath) as npzfile:
            x_eval = npzfile['arr_0']
            y_eval = npzfile['arr_1']
            print("Loaded evaluation set from cache")
            eval_set_loaded = 1
    else:
        eval_set_loaded = 0

    # Download
    if not dev_set_loaded:
        dev_paths = []
        for fname in dev_files:
            dev_paths.append(get_file(fname,
                                origin=os.path.join(dev_base, fname),
                                cache_subdir=dirname,
                                cache_dir=datadir_base))
    if not eval_set_loaded:
        eval_paths = []
        for fname in eval_files:
            eval_paths.append(get_file(fname,
                                origin=os.path.join(eval_base, fname),
                                cache_subdir=dirname,
                                cache_dir=datadir_base))

    # Extract development set
    if not dev_set_loaded:
        for fname in dev_paths:
            with zipfile.ZipFile(fname, 'r') as z:
                infolist = z.infolist()
                # Check if already extracted
                if __content_valid(infolist, datadir) is True:
                    print("Extracted dataset found for " + fname)
                    sys.stdout.flush()
                else:
                    print("Extracting " + fname)
                    sys.stdout.flush()
                    z.extractall(datadir)

    # Extract evaluation set
    if not eval_set_loaded:
        for fname in eval_paths:
            with zipfile.ZipFile(fname, 'r') as z:
                infolist = z.infolist()
                # Check if already extracted
                if __content_valid(infolist, datadir) is True:
                    print("Extracted dataset found for " + fname)
                    sys.stdout.flush()
                else:
                    print("Extracting " + fname)
                    sys.stdout.flush()
                    z.extractall(datadir)

    # load dataset meta data, csv-format, [audio file (string)][tab][scene label (string)]
    if not dev_set_loaded:
        dev_fileset = np.loadtxt(dev_meta_filepath, dtype=str)
        # Load wav data
        x_dev, y_dev = __load_files(dev_fileset, dev_fileset_path, labels)
        # Save data
        np.savez(npz_dev_filepath, x_dev, y_dev)

    if not eval_set_loaded:
        eval_fileset = np.loadtxt(eval_meta_filepath, dtype=str)
        # Load wav data
        x_eval, y_eval = __load_files(eval_fileset, eval_fileset_path, labels)
        # Save data
        np.savez(npz_eval_filepath, x_eval, y_eval)

    return (x_dev, y_dev), (x_eval, y_eval)