def test_convert_to_sequences():
feature_extractor = FeatureExtractor(pad_mode="constant")
audio_rep = np.zeros((feature_extractor.sequence_frames * 2, 2)) # (64, 2)
frames = feature_extractor.convert_to_sequences(audio_rep) # (3, 32, 2)
assert frames.shape == (3, 32, 2)
# Check that ignore last samples
audio_rep = np.zeros(
(feature_extractor.sequence_frames * 2 + 10, 2)) # (74, 2)
frames = feature_extractor.convert_to_sequences(audio_rep) # (3, 32, 2)
assert frames.shape == (3, 32, 2)
feature_extractor = FeatureExtractor(sequence_hop_time=-1,
pad_mode="constant")
frames = feature_extractor.convert_to_sequences(audio_rep) # (1, 74, 2)
assert frames.shape == (1, 32, 2)
feature_extractor = FeatureExtractor(sequence_hop_time=-1,
sequence_time=-1,
pad_mode="constant")
frames = feature_extractor.convert_to_sequences(audio_rep) # (1, 74, 2)
assert frames.shape == (1, 74, 2)
# Test it together with pad_audio
feature_extractor = FeatureExtractor(pad_mode="constant")
n_frames = feature_extractor.sequence_frames
n_samples = n_frames * feature_extractor.audio_hop + feature_extractor.audio_win
audio = np.zeros(n_samples)
audio_pad = feature_extractor.pad_audio(audio)
stft = librosa.core.stft(
audio_pad,
n_fft=1024,
hop_length=feature_extractor.audio_hop,
win_length=feature_extractor.audio_win,
center=False,
)
spectrogram = np.abs(stft)**2
spectrogram = spectrogram.T
frames = feature_extractor.convert_to_sequences(spectrogram)
assert frames.shape == (1, feature_extractor.sequence_frames, 513)
n_frames = 2 * feature_extractor.sequence_frames
n_samples = n_frames * feature_extractor.audio_hop + feature_extractor.audio_win
audio = np.zeros(n_samples)
audio_pad = feature_extractor.pad_audio(audio)
stft = librosa.core.stft(
audio_pad,
n_fft=1024,
hop_length=feature_extractor.audio_hop,
win_length=feature_extractor.audio_win,
center=False,
)
spectrogram = np.abs(stft)**2
spectrogram = spectrogram.T
frames = feature_extractor.convert_to_sequences(spectrogram)
assert frames.shape == (3, feature_extractor.sequence_frames, 513)
def test_pad_audio():
# No sequence slicing
# sequence_frames = 32
feature_extractor = FeatureExtractor(sequence_hop_time=-1,
pad_mode='constant')
n_frames = feature_extractor.sequence_frames - 1
n_samples = n_frames * feature_extractor.audio_hop + feature_extractor.audio_win
audio = np.ones(n_samples)
audio_pad = feature_extractor.pad_audio(audio)
assert len(
audio_pad
) == feature_extractor.sequence_frames * feature_extractor.audio_hop + feature_extractor.audio_win
# No sequence slicing, audio larger than sequence_time
n_frames = feature_extractor.sequence_frames + 1
n_samples = n_frames * feature_extractor.audio_hop + feature_extractor.audio_win
audio = np.ones(n_samples)
audio_pad = feature_extractor.pad_audio(audio)
assert len(
audio_pad
) == feature_extractor.sequence_frames * feature_extractor.audio_hop + feature_extractor.audio_win
# Sequence slicing, audio shorter than one sequence
# sequence_frames = 32
# sequence_hop = 16
feature_extractor = FeatureExtractor(pad_mode='constant')
n_frames = feature_extractor.sequence_frames - 1
n_samples = n_frames * feature_extractor.audio_hop + feature_extractor.audio_win
audio = np.ones(n_samples)
audio_pad = feature_extractor.pad_audio(audio)
assert len(
audio_pad
) == feature_extractor.sequence_frames * feature_extractor.audio_hop + feature_extractor.audio_win
# Sequence slicing, audio larger than one sequence
# sequence_frames = 32
# sequence_hop = 16
feature_extractor = FeatureExtractor(pad_mode='constant')
n_frames = feature_extractor.sequence_frames + 1
n_samples = n_frames * feature_extractor.audio_hop + feature_extractor.audio_win
audio = np.ones(n_samples)
audio_pad = feature_extractor.pad_audio(audio)
assert len(audio_pad) == (
feature_extractor.sequence_frames + feature_extractor.sequence_hop
) * feature_extractor.audio_hop + feature_extractor.audio_win
# Sequence slicing, audio length equal to two sequences
feature_extractor = FeatureExtractor(pad_mode='constant')
n_frames = 2 * feature_extractor.sequence_frames
n_samples = n_frames * feature_extractor.audio_hop + feature_extractor.audio_win
audio = np.ones(n_samples)
audio_pad = feature_extractor.pad_audio(audio)
assert len(audio_pad) == (
2 * feature_extractor.sequence_frames
) * feature_extractor.audio_hop + feature_extractor.audio_win