def __getitem__(self, idx):
temp = random.randint(0, 1)
augment = Compose([
TimeStretch(min_rate=0.8, max_rate=1.25, p=0.5),
Shift(min_fraction=-0.5, max_fraction=0.5, p=0.5, rollover=False)
])
self.wavPath = str(self.data.iloc[idx, 0])
self.label = self.data.iloc[idx, 1]
self.signal, self.sr = torchaudio.load(self.wavPath)
if (temp == 1):
self.signal = torch.from_numpy(
augment(samples=self.signal.numpy(), sample_rate=self.sr))
self.spectogram = torchaudio.transforms.Spectrogram(n_fft=320,
win_length=320,
hop_length=160)(
self.signal)
self.logSpectogram = torchaudio.transforms.AmplitudeToDB()(
self.spectogram)
#self.tempImg=torchvision.transforms.ToPILImage()(self.logSpectogram)
#self.tempImg=self.tempImg.convert("RGB")
#self.spectogramImageTensor=self.vision_transform(self.tempImg)
return self.logSpectogram, self.label
def pitch_shift(data_path,
file_info,
n_repeats=3,
min_semitones=-4,
max_semitones=4):
# Create the augmenter
augmenter = Compose([
PitchShift(min_semitones=min_semitones,
max_semitones=max_semitones,
p=1.0)
])
# Iterate through the Gibbon audio files only
for j in file_info[file_info.label == 1].index:
for i in range(n_repeats):
# Read audio file
rate, samples = wavfile.read(data_path + 'Clean/' +
file_info.at[j, 'fname'])
# Set the output path
output_file_path = data_path + 'Augmented/PitchShift_{:03d}_'.format(
i) + file_info.at[j, 'fname']
# Perform time stretch
augmented_samples = augmenter(samples=samples, sample_rate=rate)
# Save the new audio
wavfile.write(filename=output_file_path,
rate=rate,
data=augmented_samples)
def augmented_feature_engineering(wavFile, settings):
fs, rawWav = scipy.io.wavfile.read(wavFile)
wavData = rawWav
if (settings['CHANNELS'] == 2):
wavData = rawWav[:, 0]
augmenter = Compose([
AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.015, p=0.5),
TimeStretch(min_rate=0.8, max_rate=1.25, p=0.5),
Shift(min_fraction=-0.5, max_fraction=0.5, p=0.5),
])
wavData = augmenter(samples=np.array(wavData, dtype="float32"),
sample_rate=fs)
data_row = []
input_type = settings['FEATURE_ENGINEERING_TYPE']
if (input_type == TYPE_FEATURE_ENGINEERING_NORM_MFCC):
mfcc_result1 = mfcc(wavData,
samplerate=fs,
nfft=1103,
numcep=30,
nfilt=40,
preemph=0.5,
winstep=0.005,
winlen=0.015,
appendEnergy=False)
data_row.extend(mfcc_result1.ravel())
elif (input_type == TYPE_FEATURE_ENGINEERING_RAW_WAVE):
data_row = wavData
else:
print("OLD MFCC TYPE IS NOT SUPPORTED FOR TRAINING PYTORCH")
return data_row
def add_gaussian_noise(data_path,
file_info,
n_repeats=3,
min_amp=0.001,
max_amp=0.015):
# Create the augmenter
augmenter = Compose([
AddGaussianNoise(min_amplitude=min_amp, max_amplitude=max_amp, p=1.0)
])
# Iterate through the Gibbon audio files only
for j in file_info[file_info.label == 1].index:
for i in range(n_repeats):
# Read audio file
rate, samples = wavfile.read(data_path + 'Clean/' +
file_info.at[j, 'fname'])
# Set the output path
output_file_path = data_path + 'Augmented/AddGaussianNoise_{:03d}_'.format(
i) + file_info.at[j, 'fname']
# Add gaussian noise
augmented_samples = augmenter(samples=samples, sample_rate=rate)
# Save the new audio
wavfile.write(filename=output_file_path,
rate=rate,
data=augmented_samples)
def __init__(
self,
directory_path: str,
dataset_folds: List[str],
train: bool,
arguments: Dict[str, Any],
augmentations: List[str],
train_cqts_path: str = None,
train_gfccs_path: str = None,
):
self.arguments = arguments
self.paths = []
self.train = train
for filename in tqdm(natsorted(os.listdir(directory_path))):
if filename[0] not in dataset_folds:
continue
path = os.path.join(directory_path, filename)
self.paths.append(path)
self.augmentations = Compose([
name2augmentation[name] for name in augmentations
if name in name2augmentation.keys()
])
self.train_cqts_path = train_cqts_path
self.train_gfccs_path = train_gfccs_path
if self.arguments["cqt"]:
assert (self.train_cqts_path is not None
), "cqt feature is True but there is no train_cqts_path"
if self.arguments["gfcc"]:
assert (self.train_gfccs_path is not None
), "gfcc feature is True but there is no train_gfccs_path"
def get_transforms(bckgrd_aug_dir=None, secondary_bckgrd_aug_dir=None):
list_of_aug = [
# AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.015, p=0.3),
AddGaussianNoise(p=0.2),
AddGaussianSNR(p=0.2),
Gain(min_gain_in_db=-15, max_gain_in_db=15, p=0.3)
]
if bckgrd_aug_dir is not None:
list_of_aug.append(AddBackgroundNoise(bckgrd_aug_dir, p=0.2))
if secondary_bckgrd_aug_dir is not None:
list_of_aug.append(
AddShortNoises(secondary_bckgrd_aug_dir,
min_time_between_sounds=0.0,
max_time_between_sounds=15.0,
burst_probability=0.5,
p=0.6))
list_of_aug += [
AddGaussianNoise(p=0.2),
AddGaussianSNR(p=0.2),
Gain(min_gain_in_db=-15, max_gain_in_db=15, p=0.3)
]
augmenter = Compose(list_of_aug)
transforms = {
"train": get_training_augmentation(augmenter),
"valid": get_validation_augmentation()
}
return transforms
def transform_speakers(self, speakers):
"""Used to transform the separate speakers, without the added noise"""
speaker_augment = Compose([
t for t in augment.transforms
if 'noise' not in t.__str__().lower()
])
return (speaker_augment(speakers, sample_rate=self.sample_rate))
def build_transforms(train=True):
return Compose([
AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.015, p=0.5),
TimeStretch(min_rate=0.8, max_rate=1.25, p=0.5),
PitchShift(min_semitones=-4, max_semitones=4, p=0.5),
Shift(min_fraction=-0.5, max_fraction=0.5, p=0.5),
])
def make_transform():
return Compose([
FrequencyMask(min_frequency_band=0.005,
max_frequency_band=0.10,
p=0.25),
TimeStretch(min_rate=0.15, max_rate=.25, p=0.25),
AddGaussianSNR(min_SNR=0.001, max_SNR=.25, p=0.25)
])
def get_stretched_audio(self, x, sr):
composition = []
composition.append(
AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.015, p=0.5))
augmenter = Compose(composition)
aug_chord = augmenter(samples=x, sample_rate=sr)
return aug_chord
def __init__(self, dataset):
self.dataset = dataset
self.sample_rate = TRAINING_CONFIG['audio_sample_rate']
self.augmenter = Compose([
AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.015, p=0.5),
TimeStretch(min_rate=0.8, max_rate=1.25, p=0.5),
PitchShift(min_semitones=-4, max_semitones=4, p=0.5),
Shift(min_fraction=-0.5, max_fraction=0.5, p=0.5),
])
def __init__(self):
super(Augment_Time, self).__init__()
self.p = 0.5
self.augmenter = Compose([
AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.01, p=0.3),
TimeStretch(min_rate=0.8, max_rate=1.25, p=0.5),
PitchShift(min_semitones=-4, max_semitones=4, p=0.5),
FrequencyMask(),
TimeMask()
#Shift(min_fraction=-0.5, max_fraction=0.5, p=0.5),
])
def __init__(self, desired_sr, mode, time_aug=False):
super(ReadAudio, self).__init__()
self.desired_sr = desired_sr
self.augmenter = Compose([
AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.01, p=0.3),
TimeStretch(min_rate=0.8, max_rate=1.25, p=0.3),
PitchShift(min_semitones=-4, max_semitones=4, p=0.5),
#Shift(min_fraction=-0.5, max_fraction=0.5, p=0.5),
])
self.mode = mode
self.time_aug = time_aug
def __getitem__(self,idx):
augment = Compose([
TimeStretch(min_rate=0.8, max_rate=1.25, p=0.5),
Shift(min_fraction=-0.5, max_fraction=0.5, p=0.5,rollover=False)
])
temp1=random.randint(0,1)
temp2=random.randint(0,1)
temp3=random.randint(0,1)
self.anchor=str(self.data.iloc[idx,0])
self.positive=self.data.iloc[idx,1]
self.negative=self.data.iloc[idx,2]
self.signalAnchor,self.srAnchor=torchaudio.load(self.anchor)
self.signalPositive,self.srPositive=torchaudio.load(self.positive)
self.signalNegative,self.srNegative=torchaudio.load(self.negative)
if (temp1==1):
self.signalAnchor=torch.from_numpy(augment(samples=self.signalAnchor.numpy(),sample_rate=self.srAnchor))
if (temp2==1):
self.signalPositive=torch.from_numpy(augment(samples=self.signalPositive.numpy(),sample_rate=self.srPositive))
if (temp3==1):
self.signalNegative=torch.from_numpy(augment(samples=self.signalNegative.numpy(),sample_rate=self.srNegative))
self.spectogramAnchor=torchaudio.transforms.Spectrogram(n_fft=320,hop_length=160,win_length=320)(self.signalAnchor)
self.logSpectogramAnchor=torchaudio.transforms.AmplitudeToDB()(self.spectogramAnchor)
self.spectogramPositive=torchaudio.transforms.Spectrogram(n_fft=320,hop_length=160,win_length=320)(self.signalPositive)
self.logSpectogramPositive=torchaudio.transforms.AmplitudeToDB()(self.spectogramPositive)
self.spectogramNegative=torchaudio.transforms.Spectrogram(n_fft=320,hop_length=160,win_length=320)(self.signalNegative)
self.logSpectogramNegative=torchaudio.transforms.AmplitudeToDB()(self.spectogramNegative)
#self.tempImgAnchor=torchvision.transforms.ToPILImage()(self.logSpectogramAnchor)
#self.tempImgAnchor=self.tempImgAnchor.convert("RGB")
#self.spectogramAnchorImageTensor=self.vision_transform(self.tempImgAnchor)
#self.tempImgPositive=torchvision.transforms.ToPILImage()(self.logSpectogramPositive)
#self.tempImgPositive=self.tempImgPositive.convert("RGB")
#self.spectogramPositiveImageTensor=self.vision_transform(self.tempImgPositive)
#self.tempImgNegative=torchvision.transforms.ToPILImage()(self.logSpectogramNegative)
#self.tempImgNegative=self.tempImgNegative.convert("RGB")
#self.spectogramNegativeImageTensor=self.vision_transform(self.tempImgNegative)
return self.logSpectogramAnchor,self.logSpectogramPositive,self.logSpectogramNegative
def get_transforms(bckgrd_aug_dir=None):
list_of_aug = [
AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.015, p=0.3),
AddGaussianSNR(p=0.3)
]
if bckgrd_aug_dir is not None:
list_of_aug.append(AddBackgroundNoise(bckgrd_aug_dir,p=0.5))
augmenter = Compose(list_of_aug)
transforms = {
"train": get_training_augmentation(augmenter),
"valid": get_validation_augmentation()
}
return transforms
def compose_augmentations(rir_path):
impulse_path = os.path.join(rir_path, 'simulated_rirs')
noise_path = os.path.join(rir_path, 'pointsource_noises')
if not (os.path.exists(impulse_path) and os.path.exists(noise_path)):
raise ValueError(
'Unable to augment signal, rir_path "{}" does not exist.'.format(
rir_path))
return Compose([
AddGaussianSNR(min_SNR=0.2, max_SNR=0.5, p=0.5),
AddImpulseResponse(impulse_path, leave_length_unchanged=True, p=0.3),
AddBackgroundNoise(noise_path, p=0.3),
AddShortNoises(noise_path, max_snr_in_db=80, p=0.3)
])
def augment_stretched_noise(data, sr, label, noise=True, stretch=True):
composition = []
if noise:
composition.append(
AddGaussianNoise(min_amplitude=0.001,
max_amplitude=0.015,
p=0.5))
if stretch:
composition.append(TimeStretch(min_rate=0.8, max_rate=1.25, p=0.5))
augmenter = Compose(composition)
aug_chord = augmenter(samples=data, sample_rate=sr)
mfccs = librosa.feature.mfcc(y=aug_chord, sr=sr, n_mfcc=40)
mfccs_processed = np.mean(mfccs.T, axis=0)
return mfccs_processed
def __getitem__(self, idx: int):
wav_path, ebird_code = self.file_list[idx]
y, sr = sf.read(wav_path, dtype='float32')
signal_aug = Compose([
AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.015, p=0.5),
TimeStretch(min_rate=0.9, max_rate=1.15, p=0.5)
])
len_y = len(y)
effective_length = sr * PERIOD
if len_y < effective_length:
new_y = np.zeros(effective_length, dtype=np.float32)
start = np.random.randint(effective_length - len_y)
new_y[start:start + len_y] = y
y = new_y.astype(np.float32)
elif len_y > effective_length:
start = np.random.randint(len_y - effective_length)
y = y[start:start + effective_length].astype(np.float32)
else:
y = y.astype(np.float32)
if self.waveform_transforms:
y = signal_aug(samples=y, sample_rate=sr)
melspec = librosa.feature.melspectrogram(
y, sr=sr, **self.melspectrogram_parameters)
melspec = librosa.power_to_db(melspec).astype(np.float32)
if self.spectrogram_transforms:
melspec = self.spectrogram_transforms(melspec)
else:
pass
#img_aug = transforms.Compose([transforms.RandomHorizontalFlip(p=0.5)])
image = mono_to_color(melspec)
height, width, _ = image.shape
image = cv2.resize(
image, (int(width * self.img_size / height), self.img_size))
#image = img_aug(image)
image = np.moveaxis(image, 2, 0)
image = (image / 255.0).astype(np.float32)
labels = np.zeros(len(BIRD_CODE), dtype="f")
labels[BIRD_CODE[ebird_code]] = 1
return image, labels
def compose(sounds_path):
_p = 0.2
transforms = [
MyGain(p=_p),
AddGaussianNoise(p=_p),
Shift(p=_p, min_fraction=-0.25, max_fraction=0.25),
FrequencyMask(p=_p),
TimeMask(p=_p, max_band_part=0.25),
AddGaussianSNR(p=_p),
ClippingDistortion(p=_p, max_percentile_threshold=20),
AddBackgroundNoise(sounds_path=sounds_path, p=_p),
TimeStretch(p=_p/10),
PitchShift(p=_p/30),
]
return Compose(transforms, p=0.4, shuffle=True)
def __init__(
self,
dataset_path: str,
dataset_folds: List[int],
train: bool,
augmentations: List[str],
):
self.train = train
self.paths = []
for filename in tqdm(natsorted(os.listdir(dataset_path))):
if filename[0] not in dataset_folds:
continue
path = os.path.join(dataset_path, filename)
self.paths.append(path)
self.sampling_rate = int(dataset_path.split("_")[-1])
self.augmentations = Compose([
name2augmentation[name] for name in augmentations
if name in name2augmentation.keys()
])
def raw_audio_process(transform_fn):
augment_fn = Compose([
AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.015, p=0.5),
TimeStretch(min_rate=0.8, max_rate=1.25, p=0.5),
PitchShift(min_semitones=-4, max_semitones=4, p=0.5),
Shift(min_fraction=-0.5, max_fraction=0.5, p=0.5)
])
@wraps(transform_fn)
def augment_audio(audio, **kwargs):
sr = kwargs.setdefault('sr', 22050)
n_win = kwargs.setdefault('n_win', 20)
win_length = int(n_win * sr / 1000)
audio = augment_fn(audio)
return transform_fn(audio,
win_length=win_length,
hop_length=win_length // 4)
return augment_audio
def process_fn(output='stft', spec_aug=False, p=0.5, sr=22050):
augment_fn = Compose([
AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.015, p=p),
TimeStretch(min_rate=0.8, max_rate=1.25, p=p),
PitchShift(min_semitones=-4, max_semitones=4, p=p),
Shift(min_fraction=-0.5, max_fraction=0.5, p=p)
])
win_length = int(20 * sr / 1000)
if output == 'stft':
def stft_transform(feats):
if feats.ndim == 1:
feats = augment_fn(samples=feats, sample_rate=sr)
feats = np.log(
np.abs(librosa.stft(feats, 1023, win_length=win_length)).T
+ 1e-12)
if spec_aug:
feats = spec_augment(feats)
return feats
return stft_transform
if output == 'lms':
def lms_transform(feats):
if feats.ndim == 1:
feats = augment_fn(samples=feats, sample_rate=sr)
hop_length = win_length // 4
feats = np.log(
np.abs(
librosa.feature.melspectrogram(
feats,
n_fft=win_length,
hop_length=hop_length,
win_length=win_length)).T + 1e-12)
if spec_aug:
feats = spec_augment(feats)
return feats
return lms_transform
def __init__(self, path_audio, y, resample_freq = 32000, max_length=3, augmentation=[], validation=False, num_class=264, pseudo_labels=None):
self.labels2idx = {'Pump': 0, 'Spinach': 1, 'abalimi': 2, 'afukirira': 3, 'agriculture': 4, 'akammwanyi': 5, 'akamonde': 6, 'akasaanyi': 7, 'akatunda': 8, 'akatungulu': 9,
'akawuka': 10, 'amakoola': 11, 'amakungula': 12, 'amalagala': 13, 'amappapaali': 14, 'amatooke': 15, 'banana': 16, 'beans': 17, 'bibala': 18, 'bulimi': 19, 'butterfly': 20, 'cabbages': 21,
'cassava': 22, 'caterpillar': 23, 'caterpillars': 24, 'coffee': 25, 'crop': 26, 'ddagala': 27, 'dig': 28, 'disease': 29, 'doodo': 30, 'drought': 31, 'ebbugga': 32, 'ebibala': 33, 'ebigimusa': 34,
'ebijanjaalo': 35, 'ebijjanjalo': 36, 'ebikajjo': 37, 'ebikolo': 38, 'ebikongoliro': 39, 'ebikoola': 40, 'ebimera': 41, 'ebinyebwa': 42, 'ebirime': 43, 'ebisaanyi': 44, 'ebisooli': 45,
'ebisoolisooli': 46, 'ebitooke': 47, 'ebiwojjolo': 48, 'ebiwuka': 49, 'ebyobulimi': 50, 'eddagala': 51, 'eggobe': 52, 'ejjobyo': 53, 'ekibala': 54, 'ekigimusa': 55, 'ekijanjaalo': 56,
'ekikajjo': 57, 'ekikolo': 58, 'ekikoola': 59, 'ekimera': 60, 'ekirime': 61, 'ekirwadde': 62, 'ekisaanyi': 63, 'ekitooke': 64, 'ekiwojjolo': 65, 'ekyeya': 66, 'emboga': 67, 'emicungwa': 68,
'emisiri': 69, 'emiyembe': 70, 'emmwanyi': 71, 'endagala': 72, 'endokwa': 73, 'endwadde': 74, 'enkota': 75, 'ennima': 76, 'ennimiro': 77, 'ennyaanya': 78, 'ensigo': 79, 'ensiringanyi': 80, 'ensujju': 81,
'ensuku': 82, 'ensukusa': 83, 'enva endiirwa': 84, 'eppapaali': 85, 'faamu': 86, 'farm': 87, 'farmer': 88, 'farming instructor': 89, 'fertilizer': 90, 'fruit': 91, 'fruit picking': 92,
'garden': 93, 'greens': 94, 'ground nuts': 95, 'harvest': 96, 'harvesting': 97, 'insect': 98, 'insects': 99, 'irish potatoes': 100, 'irrigate': 101, 'kaamulali': 102, 'kasaanyi': 103, 'kassooli': 104,
'kikajjo': 105, 'kikolo': 106, 'kisaanyi': 107, 'kukungula': 108, 'leaf': 109, 'leaves': 110, 'lumonde': 111, 'lusuku': 112, 'maize': 113, 'maize stalk borer': 114, 'maize streak virus': 115, 'mango': 116, 'mangoes': 117, 'matooke': 118,
'matooke seedlings': 119, 'medicine': 120, 'miceere': 121, 'micungwa': 122, 'mpeke': 123, 'muceere': 124, 'mucungwa': 125, 'mulimi': 126, 'munyeera': 127, 'muwogo': 128,
'nakavundira': 129, 'nambaale': 130, 'namuginga': 131, 'ndwadde': 132, 'nfukirira': 133, 'nnakati': 134, 'nnasale beedi': 135, 'nnimiro': 136, 'nnyaanya': 137, 'npk': 138, 'nursery bed': 139,
'obulimi': 140, 'obulwadde': 141, 'obumonde': 142, 'obusaanyi': 143, 'obutunda': 144, 'obutungulu': 145, 'obuwuka': 146, 'okufukirira': 147, 'okufuuyira': 148, 'okugimusa': 149, 'okukkoola': 150,
'okukungula': 151, 'okulima': 152, 'okulimibwa': 153, 'okunnoga': 154, 'okusaasaana': 155, 'okusaasaanya': 156, 'okusiga': 157,
'okusimba': 158, 'okuzifuuyira': 159, 'olusuku': 160, 'omuceere': 161, 'omucungwa': 162, 'omulimi': 163, 'omulimisa': 164, 'omusiri': 165, 'omuyembe': 166,
'onion': 167, 'orange': 168, 'pampu': 169, 'passion fruit': 170, 'pawpaw': 171, 'pepper': 172, 'plant': 173, 'plantation': 174, 'ppaapaali': 175, 'pumpkin': 176, 'rice': 177, 'seed': 178,
'sikungula': 179, 'sow': 180, 'spray': 181, 'spread': 182, 'suckers': 183, 'sugarcane': 184, 'sukumawiki': 185, 'super grow': 186, 'sweet potatoes': 187, 'tomatoes': 188, 'vegetables': 189,
'watermelon': 190, 'weeding': 191, 'worm': 192}
self.idx2labels = {k:v for v,k in self.labels2idx.items()}
identity = np.eye(num_class)
self.augmentation = set(augmentation)
self.samples = path_audio #+ path_augment
self.max_length = max_length # 99% are shorter than 3 sec
self.resample_freq=resample_freq
self.validation = validation
self.y = np.array([identity[self.labels2idx[t]] for t in y]).astype(np.float32) #+ [self.labels2idx[t] for t in y_aug]
self.num_class = num_class
self.noise = Compose([AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.015, p=0.6),
TimeStretch(min_rate=0.8, max_rate=1.25, p=0.6),
PitchShift(min_semitones=-4, max_semitones=4, p=0.5),
Shift(min_fraction=-0.5, max_fraction=0.5, p=0.5),
Gain(min_gain_in_db=-12, max_gain_in_db=12, p=0.6),
])
if pseudo_labels is not None:
self.add_pl(pseudo_labels[0], pseudo_labels[1])
def __init__(
self,
manifest_path,
sample_rate,
max_sample_size=None,
min_sample_size=None,
shuffle=True,
min_length=0,
pad=False,
normalize=False,
):
super(AugmentedFileAudioDataset, self).__init__(
manifest_path=manifest_path,
sample_rate=sample_rate,
max_sample_size=max_sample_size,
min_sample_size=min_sample_size,
shuffle=shuffle,
min_length=min_length,
pad=pad,
normalize=normalize,
)
self.pre_transform = Compose([
#AddGaussianNoise(min_amplitude=1e-3, max_amplitude=5e-2, p=0.8),
#PitchShift(min_semitones=-4, max_semitones=4, p=0.8),
FrequencyMask(min_frequency_band=0.0,
max_frequency_band=0.05,
p=0.5),
TimeMask(min_band_part=0.0, max_band_part=0.05, p=0.5)
#ClippingDistortion(min_percentile_threshold=10, max_percentile_threshold=40, p=0.2),
])
random_reverb = RandomReverb()
random_clip = RandomClip()
random_time_dropout = RandomTimeDropout()
self.post_transform = augment.EffectChain().reverb(
random_reverb).channels(1).clip(random_clip) #.time_dropout(200)
def __init__(
self,
sound_file_paths,
batch_size=8,
augment=True,
save_augmented_sounds_to_path=None,
fixed_sound_length=FIXED_SOUND_LENGTH,
num_mels=NUM_MELS,
preprocessing_fn=None,
):
self.sound_file_paths = sound_file_paths
self.batch_size = batch_size
self.augment = augment
self.save_augmented_sounds_to_path = save_augmented_sounds_to_path
self.fixed_sound_length = fixed_sound_length
self.min_num_samples = (fixed_sound_length + 3) * HOP_LENGTH
self.num_mels = num_mels
self.preprocessing_fn = preprocessing_fn
self.laughter_paths = self.sound_file_paths["laughter"]
self.non_laughter_paths = []
for category in self.sound_file_paths:
if not is_laughter_category(category):
self.non_laughter_paths += self.sound_file_paths[category]
if save_augmented_sounds_to_path:
os.makedirs(save_augmented_sounds_to_path, exist_ok=True)
self.augmenter = Compose(
[
AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.002, p=0.1),
TimeStretch(min_rate=0.8, max_rate=1.25, p=0.02),
PitchShift(min_semitones=-3, max_semitones=3, p=0.02),
Shift(min_fraction=-0.5, max_fraction=0.5, p=0.5),
]
)
def load_wav_file(sound_file_path):
sample_rate, sound_np = wavfile.read(sound_file_path)
if sample_rate != SAMPLE_RATE:
raise Exception("Unexpected sample rate {} (expected {})".format(
sample_rate, SAMPLE_RATE))
if sound_np.dtype != np.float32:
assert sound_np.dtype == np.int16
sound_np = sound_np / 32767 # ends up roughly between -1 and 1
return sound_np
augmenter = Compose([
AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.015, p=0.5),
TimeStretch(min_rate=0.8, max_rate=1.25, p=0.5),
PitchShift(min_semitones=-4, max_semitones=4, p=0.5),
Shift(min_fraction=-0.5, max_fraction=0.5, p=0.5),
])
current_dir = os.path.dirname(__file__)
output_dir = os.path.join(current_dir, "output")
os.makedirs(output_dir, exist_ok=True)
samples = load_wav_file(os.path.join(current_dir, "acoustic_guitar_0.wav"))
for i in tqdm(range(20)):
output_file_path = os.path.join(output_dir, "{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path, rate=SAMPLE_RATE, data=augmented_samples)
def generate(self, wave_file, output_dir):
"""
For each transformation, apply it to an example sound and write the transformed sounds to
an output folder.
"""
samples = load_wav_file(wave_file)
_filename = os.path.basename(wave_file).split('.')[0]
# AddImpulseResponse
if self.AddImpulseResponse[0]:
augmenter = Compose([
AddImpulseResponse(p=1.0, ir_path=os.path.join(DEMO_DIR, "ir"))
])
output_file_path = os.path.join(
output_dir,
_filename + "_AddImpulseResponse{:03d}.wav".format(0))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# FrequencyMask
if self.FrequencyMask[0]:
augmenter = Compose([FrequencyMask(p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir,
_filename + "_FrequencyMask{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# TimeMask
if self.TimeMask[0]:
augmenter = Compose([TimeMask(p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, _filename + "_TimeMask{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# AddGaussianSNR
if self.AddGaussianSNR[0]:
augmenter = Compose([AddGaussianSNR(p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir,
_filename + "_AddGaussianSNR{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# AddGaussianNoise
if self.AddGaussianNoise[0]:
augmenter = Compose([
AddGaussianNoise(min_amplitude=0.001,
max_amplitude=0.015,
p=1.0)
])
for i in range(5):
output_file_path = os.path.join(
output_dir,
_filename + "_AddGaussianNoise{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# TimeStretch
if self.TimeStretch[0]:
augmenter = Compose(
[TimeStretch(min_rate=0.5, max_rate=1.5, p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, _filename + "_TimeStretch{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# PitchShift
if self.PitchShift[0]:
augmenter = Compose(
[PitchShift(min_semitones=-6, max_semitones=12, p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, _filename + "_PitchShift{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# Shift
if self.Shift[0]:
augmenter = Compose(
[Shift(min_fraction=-0.5, max_fraction=0.5, p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, _filename + "_Shift{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# Shift without rollover
if self.ShiftWithoutRoll[0]:
augmenter = Compose([
Shift(min_fraction=-0.2,
max_fraction=0.2,
rollover=False,
p=1.0)
])
for i in range(5):
output_file_path = os.path.join(
output_dir,
_filename + "_ShiftWithoutRollover{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# Normalize
if self.Normalize[0]:
augmenter = Compose([Normalize(p=1.0)])
output_file_path = os.path.join(
output_dir, _filename + "_Normalize{:03d}.wav".format(0))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# Resample
if self.Resample[0]:
augmenter = Compose([
Resample(min_sample_rate=12000, max_sample_rate=44100, p=1.0)
])
for i in range(5):
output_file_path = os.path.join(
output_dir, _filename + "_Resample{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# ClippingDistortion
if self.ClippingDistortion[0]:
augmenter = Compose(
[ClippingDistortion(max_percentile_threshold=10, p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir,
_filename + "_ClippingDistortion{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# AddBackgroundNoise
if self.AddBackgroundNoise[0]:
augmenter = Compose([
AddBackgroundNoise(sounds_path=os.path.join(
DEMO_DIR, "background_noises"),
p=1.0)
])
for i in range(5):
output_file_path = os.path.join(
output_dir,
_filename + "_AddBackgroundNoise{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# AddWhiteNoise
if self.AddWhiteNoise[0]:
augmenter = Compose([
AddBackgroundNoise(sounds_path=os.path.join(
DEMO_DIR, "white_noises"),
p=1.0)
])
for i in range(self.AddWhiteNoise[1]):
output_file_path = os.path.join(
output_dir,
_filename + "_AddWhiteNoise{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# AddPinkNoise
if self.AddPinkNoise[0]:
augmenter = Compose([
AddBackgroundNoise(sounds_path=os.path.join(
DEMO_DIR, "pink_noises"),
p=1.0)
])
for i in range(self.AddPinkNoise[1]):
output_file_path = os.path.join(
output_dir,
_filename + "_AddPinkNoise{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
# AddShortNoises
if self.AddShortNoises[0]:
augmenter = Compose([
AddShortNoises(
sounds_path=os.path.join(DEMO_DIR, "short_noises"),
min_snr_in_db=0,
max_snr_in_db=8,
min_time_between_sounds=2.0,
max_time_between_sounds=4.0,
burst_probability=0.4,
min_pause_factor_during_burst=0.01,
max_pause_factor_during_burst=0.95,
min_fade_in_time=0.005,
max_fade_in_time=0.08,
min_fade_out_time=0.01,
max_fade_out_time=0.1,
p=1.0,
)
])
for i in range(5):
output_file_path = os.path.join(
output_dir,
_filename + "_AddShortNoises{:03d}.wav".format(i))
augmented_samples = augmenter(samples=samples,
sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path,
rate=SAMPLE_RATE,
data=augmented_samples)
DEMO_DIR = os.path.dirname(__file__)
if __name__ == "__main__":
"""
For each transformation, apply it to an example sound and write the transformed sounds to
an output folder.
"""
output_dir = os.path.join(DEMO_DIR, "output")
os.makedirs(output_dir, exist_ok=True)
samples = load_wav_file(os.path.join(DEMO_DIR, "acoustic_guitar_0.wav"))
# AddImpulseResponse
augmenter = Compose(
[AddImpulseResponse(p=1.0, ir_path=os.path.join(DEMO_DIR, "ir"))]
)
output_file_path = os.path.join(
output_dir, "AddImpulseResponse_{:03d}.wav".format(0)
)
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path, rate=SAMPLE_RATE, data=augmented_samples)
# FrequencyMask
augmenter = Compose([FrequencyMask(p=1.0)])
for i in range(5):
output_file_path = os.path.join(
output_dir, "FrequencyMask_{:03d}.wav".format(i)
)
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path, rate=SAMPLE_RATE, data=augmented_samples)
def __init__(self,
root_dir,
csv_dir,
conf,
bird_code,
inv_ebird_label,
num_test_samples=10,
bckgrd_aug_dir=None,
background_audio_dir=None,
file_type="mp3",
isTraining=True,
transform=None,
apply_mixer=False):
self.root_dir = root_dir
self.conf = conf
self.isTraining = isTraining
self.bird_code = bird_code
self.inv_ebird_label = inv_ebird_label
self.transform = transform
self.file_type = file_type
self.apply_mixer = apply_mixer
self.additional_loader_params = {
"worker_init_fn": self.init_workers_fn,
"collate_fn": self.collate_fn
}
self.sampler = ImbalancedDatasetSampler
df = pd.read_csv(csv_dir)
df.secondary_labels = df.secondary_labels.apply(eval)
self.data = list(df[["filename", "ebird_code",
"secondary_labels"]].to_dict('index').values())
self.background_audio_dir = background_audio_dir
if self.background_audio_dir is not None:
for bk in background_audio_dir.glob('**/*.wav'):
self.data.append({"filename": bk})
self.num_test_samples = num_test_samples
self.length = len(self.data)
if self.apply_mixer:
self.dict_grp = {}
for grp, d in df.groupby("ebird_code"):
self.dict_grp[grp] = d.index.values
self.possible_mixer_keys = list(self.dict_grp.keys())
if bckgrd_aug_dir is not None:
self.augmenter = Compose([
AddGaussianNoise(min_amplitude=0.001,
max_amplitude=0.015,
p=0.3),
AddGaussianSNR(p=0.3),
PitchShift(min_semitones=-4, max_semitones=4, p=0.3),
AddBackgroundNoise(bckgrd_aug_dir, p=0.5),
])
else:
self.augmenter = Compose([
AddGaussianNoise(min_amplitude=0.001,
max_amplitude=0.015,
p=0.3),
AddGaussianSNR(p=0.3),
PitchShift(min_semitones=-4, max_semitones=4, p=0.3)
])
del df
sound_np = np.divide(
sound_np, 32768, dtype=np.float32
)
number = os.path.split(audio_file)[-1][:-4]
transforms = [
{"instance": AddGaussianSNR(p=1.0), "num_runs": 3},
{"instance": TimeStretch(min_rate=0.4, max_rate=1.25, p=1.0), "num_runs": 5},
{
"instance": PitchShift(min_semitones=-5, max_semitones=5, p=1.0),
"num_runs": 6,
},
{"instance": Shift(min_fraction=-0.85, max_fraction=0.85, p=1.0), "num_runs": 4},
{"instance": Resample(p=1.0), "num_runs": 5},
{"instance": ClippingDistortion(p=1.0), "num_runs": 3},
]
for transform in transforms:
augmenter = Compose([transform["instance"]])
run_name = (
transform.get("name")
if transform.get("name")
else transform["instance"].__class__.__name__
)
for i in range(transform["num_runs"]):
output_file_path = os.path.join(
'augmented', "{}_{}_{:03d}.wav".format(number, run_name, i)
)
augmented_samples = augmenter(samples=sound_np, sample_rate=sample_rate)
wavfile.write(output_file_path, rate=sample_rate, data=augmented_samples)
