def __getitem__(self, idx):
file = self.file_names[idx]
if self.working_dir is not None:
file = os.path.join(self.working_dir, file)
sample = T.load_audio_file(file, self.sr, self.mono)
if self.transform is not None:
sample = self.transform(sample)
return sample
def __init__(
self,
file_name,
sequence_len: int,
hop: int,
sr: int = 44100,
fft_size: int = 4096,
fft_hop: int = 441,
n_freq_bins: int = 256,
freq_compression: str = "linear",
f_min: int = 200,
f_max: int = 18000,
cache_dir=None #added
):
self.sequence_len = sequence_len
self.hop = hop
self.audio = T.load_audio_file(file_name, sr=sr, mono=True)
self.n_frames = self.audio.shape[1]
self.t = [
T.PreEmphasize(DefaultSpecDatasetOps["preemphases"]),
T.Spectrogram(fft_size, fft_hop, center=False),
]
if freq_compression == "linear":
self.t.append(T.Interpolate(n_freq_bins, sr, f_min, f_max))
elif freq_compression == "mel":
self.t.append(
T.F2M(sr=sr, n_mels=n_freq_bins, f_min=f_min, f_max=f_max))
elif freq_compression == "mfcc":
t_mel = T.F2M(sr=sr, n_mels=n_freq_bins, f_min=f_min, f_max=f_max)
self.t.append(T.Compose(t_mel, T.M2MFCC()))
else:
raise "Undefined frequency compression"
self.t.append(
T.Amp2Db(min_level_db=DefaultSpecDatasetOps["min_level_db"]))
self.t.append(
T.Normalize(
min_level_db=DefaultSpecDatasetOps["min_level_db"],
ref_level_db=DefaultSpecDatasetOps["ref_level_db"],
))
#self.file_reader = AsyncFileReader()
self.t = T.CachedSpectrogram(
cache_dir=cache_dir,
spec_transform=T.Compose(self.t),
n_fft=fft_size,
hop_length=fft_hop,
#file_reader=AsyncFileReader()
)
def __init__(
self,
file_name,
sequence_len: int,
hop: int,
sr: int = 44100,
fft_size: int = 4096,
fft_hop: int = 441,
n_freq_bins: int = 256,
freq_compression: str = "linear",
f_min: int = 200,
f_max: int = 18000,
center=True
):
self.sp = signal.signal_proc()
self.hop = hop
self.center = center
self.filename = file_name
self.sequence_len = sequence_len
self.audio = T.load_audio_file(file_name, sr=sr, mono=True)
self.n_frames = self.audio.shape[1]
spec_t = [
T.PreEmphasize(DefaultSpecDatasetOps["preemphases"]),
T.Spectrogram(fft_size, fft_hop, center=self.center),
]
self.spec_transforms = T.Compose(spec_t)
if freq_compression == "linear":
self.t_compr_f = (T.Interpolate(n_freq_bins, sr, f_min, f_max))
elif freq_compression == "mel":
self.t_compr_f = (T.F2M(sr=sr, n_mels=n_freq_bins, f_min=f_min, f_max=f_max))
elif freq_compression == "mfcc":
t_mel = T.F2M(sr=sr, n_mels=n_freq_bins, f_min=f_min, f_max=f_max)
self.t_compr_f = (T.Compose(t_mel, T.M2MFCC()))
else:
raise "Undefined frequency compression"
self.t_compr_a = T.Amp2Db(min_level_db=DefaultSpecDatasetOps["min_level_db"])
self.t_norm = T.Normalize(
min_level_db=DefaultSpecDatasetOps["min_level_db"],
ref_level_db=DefaultSpecDatasetOps["ref_level_db"]
)
def __init__(self,
file_name,
sequence_len: int,
hop: int,
sr: int = 44100,
fft_size: int = 4096,
fft_hop: int = 441,
n_freq_bins: int = 256,
freq_compression: str = "linear",
f_min: int = 200,
f_max: int = 18000):
self.sequence_len = sequence_len
self.hop = hop
self.audio = T.load_audio_file(file_name, sr=sr, mono=True)
self.n_frames = self.audio.shape[
1] # total num of samples in the audio (transposed mono)
self.t = [
T.PreEmphasize(DefaultSpecDatasetOps["preemphases"]),
T.Spectrogram(fft_size, fft_hop, center=False),
]
if freq_compression == "linear":
self.t.append(T.Interpolate(n_freq_bins, sr, f_min, f_max))
elif freq_compression == "mel":
self.t.append(
T.F2M(sr=sr, n_mels=n_freq_bins, f_min=f_min, f_max=f_max))
elif freq_compression == "mfcc":
t_mel = T.F2M(sr=sr, n_mels=n_freq_bins, f_min=f_min, f_max=f_max)
self.t.append(T.Compose(t_mel, T.M2MFCC()))
else:
raise "Undefined frequency compression"
self.t.append(
T.Amp2Db(min_level_db=DefaultSpecDatasetOps["min_level_db"]))
self.t.append(
T.Normalize(
min_level_db=DefaultSpecDatasetOps["min_level_db"],
ref_level_db=DefaultSpecDatasetOps["ref_level_db"],
))
self.t = T.Compose(self.t)
def __init__(self,
file_names: Iterable[str],
working_dir=None,
cache_dir=None,
sr=44100,
n_fft=4096,
hop_length=441,
freq_compression="linear",
n_freq_bins=256,
f_min=0,
f_max=18000,
seq_len=128,
augmentation=False,
noise_files=[],
min_max_normalize=False,
*args,
**kwargs):
super().__init__(file_names, working_dir, sr, *args, **kwargs)
if self.dataset_name is not None:
self._logger.info("Init dataset {}...".format(self.dataset_name))
self.n_fft = n_fft
self.hop_length = hop_length
self.f_min = f_min
self.f_max = f_max
valid_freq_compressions = ["linear", "mel", "mfcc"]
if freq_compression not in valid_freq_compressions:
raise ValueError(
"{} is not a valid freq_compression. Must be one of {}",
format(freq_compression, valid_freq_compressions),
)
self.freq_compression = freq_compression
self.possible_call_labels = re.compile("|".join(["call"]))
self.possible_nocall_labels = re.compile("|".join(["noise"]))
self._logger.debug("Number of files : {}".format(len(self.file_names)))
_n_calls = 0
for f in self.file_names:
if self.is_call(f):
_n_calls += 1
self._logger.debug("Number of calls: {}".format(_n_calls))
self._logger.debug(
"Number of noise: {}".format(len(self.file_names) - _n_calls))
self.augmentation = augmentation
spec_transforms = [
lambda fn: T.load_audio_file(fn, sr=sr),
T.PreEmphasize(DefaultSpecDatasetOps["preemphases"]),
T.Spectrogram(n_fft, hop_length, center=False),
]
self.file_reader = AsyncFileReader()
if cache_dir is None:
self.t_spectrogram = T.Compose(spec_transforms)
else:
self.t_spectrogram = T.CachedSpectrogram(
cache_dir=cache_dir,
spec_transform=T.Compose(spec_transforms),
n_fft=n_fft,
hop_length=hop_length,
file_reader=AsyncFileReader(),
)
if augmentation:
self._logger.debug(
"Init augmentation transforms for time and pitch shift")
self.t_amplitude = T.RandomAmplitude(3, -6)
self.t_timestretch = T.RandomTimeStretch()
self.t_pitchshift = T.RandomPitchSift()
else:
self._logger.debug("Running without augmentation")
if self.freq_compression == "linear":
self.t_compr_f = T.Interpolate(n_freq_bins, sr, f_min, f_max)
elif self.freq_compression == "mel":
self.t_compr_f = T.F2M(sr=sr,
n_mels=n_freq_bins,
f_min=f_min,
f_max=f_max)
elif self.freq_compression == "mfcc":
self.t_compr_f = T.Compose(
T.F2M(sr=sr, n_mels=n_freq_bins, f_min=f_min, f_max=f_max))
self.t_compr_mfcc = T.M2MFCC(n_mfcc=32)
else:
raise "Undefined frequency compression"
if augmentation:
if noise_files:
self._logger.debug(
"Init augmentation transform for random noise addition")
self.t_addnoise = T.RandomAddNoise(
noise_files,
self.t_spectrogram,
T.Compose(self.t_timestretch, self.t_pitchshift,
self.t_compr_f),
min_length=seq_len,
return_original=True)
else:
self.t_addnoise = None
self.t_compr_a = T.Amp2Db(
min_level_db=DefaultSpecDatasetOps["min_level_db"])
if min_max_normalize:
self.t_norm = T.MinMaxNormalize()
self._logger.debug("Init min-max-normalization activated")
else:
self.t_norm = T.Normalize(
min_level_db=DefaultSpecDatasetOps["min_level_db"],
ref_level_db=DefaultSpecDatasetOps["ref_level_db"],
)
self._logger.debug("Init 0/1-dB-normalization activated")
self.t_subseq = T.PaddedSubsequenceSampler(seq_len,
dim=1,
random=augmentation)
def __init__(
self,
file_names: Iterable[str],
working_dir=None,
cache_dir=None,
sr=44100,
n_fft=2048, #4096
hop_length=220, #441
freq_compression="linear",
n_freq_bins=256, # determines the width of the image
f_min=0,
f_max=18000,
seq_len=128, # shd be adjusted together with sequence_len in class StridedAudioDataset (called by predict.py)
augmentation=False,
noise_files=[],
*args,
**kwargs):
super().__init__(file_names, working_dir, sr, *args, **kwargs)
if self.dataset_name is not None:
self._logger.info("Init dataset {}...".format(self.dataset_name))
self.n_fft = n_fft
self.hop_length = hop_length
self.f_min = f_min
self.f_max = f_max
# mel: log transformation of freq (Hz scale to Mel scale)
# attention: Mel-spectrograms as a network input led to an excessive loss of resolution in higher frequency bands, which was
# a big problem considering the high-frequency pulsed calls and whistles.
valid_freq_compressions = ["linear", "mel", "mfcc"]
if freq_compression not in valid_freq_compressions:
raise ValueError(
"{} is not a valid freq_compression. Must be one of {}",
format(freq_compression, valid_freq_compressions),
)
self.freq_compression = freq_compression
# combine a RegExp pattern into pattern objects for pattern matching
self.possible_call_labels = re.compile("|".join(["call"]))
self.possible_nocall_labels = re.compile("|".join(["noise"]))
self._logger.debug("Number of files : {}".format(len(self.file_names)))
_n_calls = 0
for f in self.file_names:
if self.is_call(f):
_n_calls += 1
self._logger.debug("Number of calls: {}".format(_n_calls))
self._logger.debug(
"Number of noise: {}".format(len(self.file_names) - _n_calls))
self.augmentation = augmentation
spec_transforms = [
lambda fn: T.load_audio_file(fn, sr=sr), # return: a vector tensor
T.PreEmphasize(DefaultSpecDatasetOps["preemphases"]),
T.Spectrogram(n_fft, hop_length, center=False),
]
self.file_reader = AsyncFileReader()
# if user chooses to not cache .spec by omitting the directory
if cache_dir is None:
self.t_spectrogram = T.Compose(spec_transforms)
else:
# where .spec is created and stored
# n_fft, hop_length: meta in spec_dict
self.t_spectrogram = T.CachedSpectrogram(
cache_dir=cache_dir,
spec_transform=T.Compose(spec_transforms),
n_fft=n_fft,
hop_length=hop_length,
file_reader=AsyncFileReader(),
)
if augmentation:
self._logger.debug(
"Init augmentation transforms for time and pitch shift")
self.t_amplitude = T.RandomAmplitude(3, -6)
self.t_timestretch = T.RandomTimeStretch()
self.t_pitchshift = T.RandomPitchSift()
else:
self._logger.debug("Running without augmentation")
if self.freq_compression == "linear":
self.t_compr_f = T.Interpolate(n_freq_bins, sr, f_min, f_max)
elif self.freq_compression == "mel":
self.t_compr_f = T.F2M(sr=sr,
n_mels=n_freq_bins,
f_min=f_min,
f_max=f_max)
elif self.freq_compression == "mfcc":
self.t_compr_f = T.Compose(
T.F2M(sr=sr, n_mels=n_freq_bins, f_min=f_min, f_max=f_max),
T.M2MFCC())
else:
raise "Undefined frequency compression"
if augmentation:
if noise_files:
self._logger.debug(
"Init augmentation transform for random noise addition")
self.t_addnoise = T.RandomAddNoise(
noise_files,
self.t_spectrogram,
T.Compose(self.t_timestretch, self.t_pitchshift,
self.t_compr_f),
min_length=seq_len,
return_original=True
) # if return_original = True, both augmented and original specs are returned
else:
self.t_addnoise = None
self.t_compr_a = T.Amp2Db(
min_level_db=DefaultSpecDatasetOps["min_level_db"])
self.t_norm = T.Normalize(
min_level_db=DefaultSpecDatasetOps["min_level_db"],
ref_level_db=DefaultSpecDatasetOps["ref_level_db"],
)
self.t_subseq = T.PaddedSubsequenceSampler(seq_len,
dim=1,
random=augmentation)
def __init__(
self,
file_names: Iterable[str],
working_dir=None,
cache_dir=None,
sr=44100,
n_fft=1024,
hop_length=512,
freq_compression="linear",
n_freq_bins=256,
f_min=None,
f_max=18000,
*args,
**kwargs
):
super().__init__(file_names, working_dir, sr, *args, **kwargs)
if self.dataset_name is not None:
self._logger.info("Init dataset {}...".format(self.dataset_name))
self.sp = signal.signal_proc()
self.sr = sr
self.f_min = f_min
self.f_max = f_max
self.n_fft = n_fft
self.hop_length = hop_length
self.sp = signal.signal_proc()
self.freq_compression = freq_compression
valid_freq_compressions = ["linear", "mel", "mfcc"]
if self.freq_compression not in valid_freq_compressions:
raise ValueError(
"{} is not a valid freq_compression. Must be one of {}",
format(self.freq_compression, valid_freq_compressions),
)
self._logger.debug(
"Number of test files: {}".format(len(self.file_names))
)
spec_transforms = [
lambda fn: T.load_audio_file(fn, sr=sr),
T.PreEmphasize(DefaultSpecDatasetOps["preemphases"]),
T.Spectrogram(n_fft, hop_length, center=False)
]
self.file_reader = AsyncFileReader()
if cache_dir is None:
self.t_spectrogram = T.Compose(spec_transforms)
else:
self.t_spectrogram = T.CachedSpectrogram(
cache_dir=cache_dir,
spec_transform=T.Compose(spec_transforms),
n_fft=n_fft,
hop_length=hop_length,
file_reader=AsyncFileReader(),
)
if self.freq_compression == "linear":
self.t_compr_f = T.Interpolate(
n_freq_bins, sr, f_min, f_max
)
elif self.freq_compression == "mel":
self.t_compr_f = T.F2M(sr=sr, n_mels=n_freq_bins, f_min=f_min, f_max=f_max)
elif self.freq_compression == "mfcc":
self.t_compr_f = T.Compose(
T.F2M(sr=sr, n_mels=n_freq_bins, f_min=f_min, f_max=f_max), T.M2MFCC()
)
else:
raise "Undefined frequency compression"
self.t_compr_a = T.Amp2Db(min_level_db=DefaultSpecDatasetOps["min_level_db"])
self.t_norm = T.Normalize(
min_level_db=DefaultSpecDatasetOps["min_level_db"],
ref_level_db=DefaultSpecDatasetOps["ref_level_db"],
)
def __init__(
self,
file_names: Iterable[str],
working_dir=None,
cache_dir=None,
sr=44100,
n_fft=4096,
hop_length=441,
freq_compression="linear",
n_freq_bins=256,
f_min=0,
f_max=18000,
seq_len=128,
augmentation=False,
noise_files_train=[],
noise_files_val=[],
noise_files_test=[],
random=False,
*args,
**kwargs
):
super().__init__(file_names, working_dir, sr, *args, **kwargs)
if self.dataset_name is not None:
self._logger.info("Init dataset {}...".format(self.dataset_name))
self.sp = signal.signal_proc()
self.df = 15.0
self.exp_e = 0.1
self.bin_pow = 2.0
self.gaus_mean = 0.0
self.gaus_stdv = 12.5
self.poisson_lambda = 15.0
self.orig_noise_value = -5
self.f_min = f_min
self.f_max = f_max
self.n_fft = n_fft
self.random = random
self.hop_length = hop_length
self.augmentation = augmentation
self.file_reader = AsyncFileReader()
self.noise_files_val = noise_files_val
self.noise_files_test = noise_files_test
self.freq_compression = freq_compression
self.noise_files_train = noise_files_train
valid_freq_compressions = ["linear", "mel", "mfcc"]
if self.freq_compression not in valid_freq_compressions:
raise ValueError(
"{} is not a valid freq_compression. Must be one of {}",
format(self.freq_compressio, valid_freq_compressions),
)
self._logger.debug(
"Number of files to denoise : {}".format(len(self.file_names))
)
spec_transforms = [
lambda fn: T.load_audio_file(fn, sr=sr),
T.PreEmphasize(DefaultSpecDatasetOps["preemphases"]),
T.Spectrogram(n_fft, hop_length, center=False),
]
if cache_dir is None:
self.t_spectrogram = T.Compose(spec_transforms)
else:
self.t_spectrogram = T.CachedSpectrogram(
cache_dir=cache_dir,
spec_transform=T.Compose(spec_transforms),
n_fft=n_fft,
hop_length=hop_length,
file_reader=AsyncFileReader())
if self.augmentation:
self._logger.debug("Init augmentation transforms for intensity, time, and pitch shift")
self.t_amplitude = T.RandomAmplitude(3, -6)
self.t_timestretch = T.RandomTimeStretch()
self.t_pitchshift = T.RandomPitchSift()
else:
#only for noise augmentation during validation phase - intensity, time and pitch augmentation is not used during validation/test
self.t_timestretch = T.RandomTimeStretch()
self.t_pitchshift = T.RandomPitchSift()
self._logger.debug("Running without intensity, time, and pitch augmentation")
if self.freq_compression == "linear":
self.t_compr_f = T.Interpolate(n_freq_bins, sr, f_min, f_max)
elif self.freq_compression == "mel":
self.t_compr_f = T.F2M(sr=sr, n_mels=n_freq_bins, f_min=f_min, f_max=f_max)
elif self.freq_compression == "mfcc":
self.t_compr_f = T.Compose(T.F2M(sr=sr, n_mels=n_freq_bins, f_min=f_min, f_max=f_max))
self.t_compr_mfcc = T.M2MFCC(n_mfcc=32)
else:
raise "Undefined frequency compression"
if self.augmentation and self.noise_files_train and self.dataset_name == "train":
self._logger.debug("Init training real-world noise files for noise2noise adding")
self.t_addnoise = T.RandomAddNoise(
self.noise_files_train,
self.t_spectrogram,
T.Compose(self.t_timestretch, self.t_pitchshift, self.t_compr_f),
min_length=seq_len,
min_snr=-2,
max_snr=-8,
return_original=True
)
elif not self.augmentation and self.noise_files_val and self.dataset_name == "val":
self._logger.debug("Init validation real-world noise files for noise2noise adding")
self.t_addnoise = T.RandomAddNoise(
self.noise_files_val,
self.t_spectrogram,
T.Compose(self.t_timestretch, self.t_pitchshift, self.t_compr_f),
min_length=seq_len,
min_snr=-2,
max_snr=-8,
return_original=True
)
elif not self.augmentation and self.noise_files_test and self.dataset_name == "test":
self._logger.debug("Init test real-world noise files for noise2noise adding")
self.t_addnoise = T.RandomAddNoise(
self.noise_files_test,
self.t_spectrogram,
T.Compose(self.t_timestretch, self.t_pitchshift, self.t_compr_f),
min_length=seq_len,
min_snr=-2,
max_snr=-8,
return_original=True
)
else:
self.t_addnoise = None
raise "ERROR: Init noise files for noise adding does not have a proper setup per split!"
self.t_compr_a = T.Amp2Db(min_level_db=DefaultSpecDatasetOps["min_level_db"])
self.t_norm = T.Normalize(
min_level_db=DefaultSpecDatasetOps["min_level_db"],
ref_level_db=DefaultSpecDatasetOps["ref_level_db"],
)
self.t_subseq = T.PaddedSubsequenceSampler(seq_len, dim=1, random=augmentation)