def test_execute(self):
sample_rate = 22050
win_size = 256
hop_size = 128
fc = FeatureContainer("fake_audio_path", sample_rate, win_size,
hop_size)
fc.features["energy"]["data"] = np.array([1, 1, 0.2, 0.1])
fc.features["spectral_flatness"]["data"] = np.array(
[0.1, 0.5, 0.12, 0.3])
segment_list = []
for i in range(4):
segment_list.append(
Segment((i * hop_size) / sample_rate,
((i * hop_size) + win_size - 1) / sample_rate))
sc = SegmentContainer("fake_audio_path")
sc.segments = segment_list
act_det = Simple(energy_threshold=0.3, spectral_flatness_threshold=0.2)
act_det.execute(sc, fc)
assert (sc.segments[0].activity and not sc.segments[1].activity
and not sc.segments[2].activity
and not sc.segments[3].activity)
def test_data_multiple_features(self):
"""
Compare data from "manual" constructor to config file-based constructor.
"""
# construct minibatch generator "manually"
sample_rate = 22050
win_size = 256
hop_size = 128
energy_threshold = 0.2
spectral_flatness_threshold = 0.3
seg_duration = 0.2
seg_overlap = 0.5
batch_size = 50
num_features = 64
n_time_bins = 34
af_gen = AudioFrameGen(sample_rate=sample_rate,
win_size=win_size,
hop_size=hop_size)
en_ext = EnergyExtractor()
sf_ext = SpectralFlatnessExtractor()
mel_ext = MelSpectrumExtractor(sample_rate=sample_rate,
fft_size=win_size,
n_mels=64,
min_freq=0,
max_freq=sample_rate / 2)
ff_pro = FrameFeatureProcessor(af_gen, [en_ext, sf_ext, mel_ext],
feature_container_root=FEATURE_ROOT)
ffc_ext = FrameFeatureChunkExtractor("mel_spectrum")
act_det = Simple(
energy_threshold=energy_threshold,
spectral_flatness_threshold=spectral_flatness_threshold)
sf_pro = SegmentFeatureProcessor([act_det, ffc_ext],
ff_pro=ff_pro,
audio_root=DATA_PATH)
parser = CSVFileLabelParser(TEST_FILE2LABEL_PATH,
label_file=TEST_LABEL_PATH)
sc_gen = SegmentContainerGenerator(DATA_PATH,
sf_pro,
label_parser=parser,
seg_duration=seg_duration,
seg_overlap=seg_overlap)
mb_gen_1 = MiniBatchGen(
sc_gen, batch_size, {
"mel_spectrum": {
"feature_size": num_features,
"n_time_bins": n_time_bins
}
}, 0)
# parse json file
with open(CONFIG_PATH) as config_file:
config = json.loads(config_file.read())
config["features"] = [{
'name': 'audio_chunk',
'config': {}
}, {
'name': 'mel_spectrum',
'config': {
'n_mels': 64,
'min_freq': 0,
'max_freq': 11025,
'log_amp': 1
}
}]
# construct minibatch generator from config
mb_gen_dict = MiniBatchGen.from_config(config)
mb_gen_2 = mb_gen_dict["default"]
# execute and compare
try:
mb_gen_1_e = mb_gen_1.execute(active_segments_only=True,
with_targets=True,
with_filenames=False)
mb_gen_2_e = mb_gen_2.execute(active_segments_only=True,
with_targets=True,
with_filenames=False)
if not os.path.exists(FEATURE_ROOT):
os.makedirs(FEATURE_ROOT)
for mb1, mb2 in zip(mb_gen_1_e, mb_gen_2_e):
assert np.all(mb1[0]["mel_spectrum"] == mb2[0]["mel_spectrum"])
assert np.all(mb1[1] == mb2[1])
except Exception as e:
pytest.fail("Unexpected Error: {}".format(e))
finally:
shutil.rmtree(FEATURE_ROOT)
def test_gen_minibatches_2d_w_pca_scaler(self):
sample_rate = 22050
win_size = 256
hop_size = 128
energy_threshold = 0.2
spectral_flatness_threshold = 0.3
seg_duration = 0.1
seg_overlap = 0.5
batch_size = 10
num_features = 16
n_time_bins = 17
af_gen = AudioFrameGen(sample_rate=sample_rate,
win_size=win_size,
hop_size=hop_size)
en_ext = EnergyExtractor()
sf_ext = SpectralFlatnessExtractor()
mel_ext = MelSpectrumExtractor(sample_rate=sample_rate,
fft_size=win_size,
n_mels=64,
min_freq=0,
max_freq=sample_rate / 2)
ff_pro = FrameFeatureProcessor(af_gen, [en_ext, sf_ext, mel_ext],
feature_container_root=FEATURE_ROOT)
pca = joblib.load(
os.path.join(DATA_PATH, "transform/mel64_pca16_norm/pca.jl"))
scaler = joblib.load(
os.path.join(DATA_PATH, "transform/mel64_pca16_norm/scaler.jl"))
ffc_ext = FrameFeatureChunkExtractor("mel_spectrum", pca, scaler)
act_det = Simple(
energy_threshold=energy_threshold,
spectral_flatness_threshold=spectral_flatness_threshold)
sf_pro = SegmentFeatureProcessor([act_det, ffc_ext],
ff_pro=ff_pro,
audio_root=DATA_PATH)
parser = CSVFileLabelParser(TEST_FILE2LABEL_PATH,
label_file=TEST_LABEL_PATH)
sc_gen = SegmentContainerGenerator(DATA_PATH,
sf_pro,
label_parser=parser,
seg_duration=seg_duration,
seg_overlap=seg_overlap)
sc_gen_e = sc_gen.execute()
active_segments = []
# compare data in segment and corresponding data in feature container
for sc in sc_gen_e:
fc_path = os.path.join(FEATURE_ROOT,
sc.audio_path.replace(".wav", ".fc.jl"))
fc = feature_container.FeatureContainer.load(fc_path)
for s in sc.segments:
if hasattr(s, 'activity') and s.activity:
start_ind = fc.time_to_frame_ind(s.start_time)
end_ind = start_ind + n_time_bins
data = scaler.transform(
pca.transform(fc.features["mel_spectrum"]["data"]
[start_ind:end_ind]))
assert np.all(data == s.features["mel_spectrum"])
active_segments.append(s)
# compare data in segment and corresponding data in minibatches
mb_gen = MiniBatchGen(
sc_gen, batch_size, {
"mel_spectrum": {
"feature_size": num_features,
"n_time_bins": n_time_bins
}
}, 0)
mb_gen_e = mb_gen.execute(active_segments_only=True,
with_targets=False,
with_filenames=False)
count = 0
for mb, in mb_gen_e:
for data in mb["mel_spectrum"]:
assert np.all(data[0].T ==
active_segments[count].features["mel_spectrum"])
count += 1
def test_init(self):
try:
Simple(energy_threshold=0.3, spectral_flatness_threshold=0.2)
except Exception as e:
pytest.fail("Unexpected Error: {}".format(e))