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 __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])
5,
},
{
"instance": ClippingDistortion(p=1.0),
"num_runs": 5
},
{
"instance":
FrequencyMask(min_frequency_band=0.5,
max_frequency_band=0.6,
p=1.0),
"num_runs":
5,
},
{
"instance": Gain(min_gain_in_db=-6, max_gain_in_db=6, p=1.0),
"num_runs": 5
},
{
"instance": PitchShift(min_semitones=-4, max_semitones=4, p=1.0),
"num_runs": 5,
},
{
"instance": LoudnessNormalization(p=1.0),
"num_runs": 5
},
{
"instance": Mp3Compression(backend="lameenc", p=1.0),
"num_runs": 5,
"name": "Mp3CompressionLameenc",
},
min_snr=5.0,
max_snr=20.0,
**kwargs):
super().__init__(always_apply, p)
self.min_snr = min_snr
self.max_snr = max_snr
def apply(self, y: np.ndarray, **params):
snr = np.random.uniform(self.min_snr, self.max_snr)
a_signal = np.sqrt(y**2).max()
a_noise = a_signal / (10**(snr / 20))
pink_noise = cn.powerlaw_psd_gaussian(1, len(y))
a_pink = np.sqrt(pink_noise**2).max()
augmented = (y + pink_noise * 1 / a_pink * a_noise).astype(y.dtype)
return augmented
AUGMENT = Compose([
PitchShift(min_semitones=-4, max_semitones=4, p=0.2),
Gain(min_gain_in_db=-12, max_gain_in_db=12, p=0.2),
AddGaussianNoise(min_amplitude=0.001, max_amplitude=0.015, p=0.2),
AddGaussianSNR(min_SNR=0.1, max_SNR=1, p=0.2),
PinkNoiseSNR(min_snr=5.0, max_snr=10, p=0.2)
])
def do_aug(samples, sample_rate):
return AUGMENT(samples, sample_rate)
output_file_path = os.path.join(
output_dir, "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)
# PolarityInversion
augmenter = Compose([PolarityInversion(p=1.0)])
output_file_path = os.path.join(
output_dir, "PolarityInversion_{:03d}.wav".format(0)
)
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path, rate=SAMPLE_RATE, data=augmented_samples)
# Gain
augmenter = Compose(
[
Gain(
min_gain_in_db=-6,
max_gain_in_db=6,
p=1.0,
)
]
)
for i in range(5):
output_file_path = os.path.join(
output_dir, "Gain_{:03d}.wav".format(i)
)
augmented_samples = augmenter(samples=samples, sample_rate=SAMPLE_RATE)
wavfile.write(output_file_path, rate=SAMPLE_RATE, data=augmented_samples)