def test_batch_spectrogram(self):
waveform, sample_rate = torchaudio.load(self.test_filepath)
# Single then transform then batch
expected = transforms.Spectrogram()(waveform).repeat(3, 1, 1, 1)
# Batch then transform
computed = transforms.Spectrogram()(waveform.repeat(3, 1, 1))
self.assertTrue(computed.shape == expected.shape, (computed.shape, expected.shape))
self.assertTrue(torch.allclose(computed, expected))
def __init__(self, d, src_path, batch_size, device):
super(VAE, self).__init__()
'''
=========== ARGUMENTS ===========
> d - dimensionality of latent space
> src_path - path to source samples
> batch_size - number of training examples in single batch
> device - CPU or GPU in use
=================================
'''
self.enc = ResNetBigger(d=d)
self.dec = nn.Sequential(nn.Linear(d, 40), nn.ReLU(),
nn.Linear(40, 50), nn.Sigmoid())
# For computing spectrogram then normalizing
self.spectrogram = T.Spectrogram(
n_fft=2048,
win_length=None,
hop_length=512,
power=2,
)
self.amp_to_db = T.AmplitudeToDB(stype='power')
self.src = torch.from_numpy(np.load(src_path))
self.src = self.src.unsqueeze(0).repeat(batch_size, 1, 1)
self.src = self.src.to(device)
self.d = d
self.device = device
def preprocess(file_path='../DATASETS/LJSpeech-1.1/metadata.csv',
root_dir='../DATASETS/LJSpeech-1.1'):
with open(file_path, encoding='utf8') as file:
data_ = [line.strip().split('|') for line in file]
root_dir = root_dir
sample_rate = 8000
resample = transforms.Resample(orig_freq=22050, new_freq=sample_rate)
spectogram = transforms.Spectrogram(n_fft=1024, hop_length=256)
to_mel = transforms.MelScale(n_mels=80,
sample_rate=sample_rate,
n_stft=1024 // 2 + 1)
mel_data = torch.zeros(len(data_), 316, 80)
mel_len = torch.empty(len(data_), dtype=torch.int)
for idx, data in enumerate(tqdm(data_)):
path, text = data[0], data[1]
path = f'{root_dir}/wavs/{path}.wav'
data, sample_rate = torchaudio.load(path)
data = resample(data)
data = spectogram(data)
data = to_mel(data)
data = data.transpose(1, 2).squeeze(0)
mel_data[idx, :data.size(0)] = data
mel_len[idx] = data.size(0)
torch.save(mel_data, f'{root_dir}/mel_data.pt')
torch.save(mel_len, f'{root_dir}/mel_len.pt')
def test_spectrogram(self):
specgram = transforms.Spectrogram(center=False,
pad_mode="reflect",
onesided=False)
self.assertEqual(specgram.center, False)
self.assertEqual(specgram.pad_mode, "reflect")
self.assertEqual(specgram.onesided, False)
def test_Spectrogram_complex(self):
n_fft = 400
hop_length = 200
sample_rate = 16000
waveform = get_whitenoise(
sample_rate=sample_rate,
n_channels=1,
).to(self.device, self.dtype)
expected = librosa.core.spectrum._spectrogram(
y=waveform[0].cpu().numpy(),
n_fft=n_fft,
hop_length=hop_length,
power=1)[0]
result = T.Spectrogram(
n_fft=n_fft,
hop_length=hop_length,
power=None,
return_complex=True,
).to(self.device, self.dtype)(waveform)[0]
self.assertEqual(result.abs(),
torch.from_numpy(expected),
atol=1e-5,
rtol=1e-5)
def __getitem__(self, index):
filename = self.data_path[index]
n_fft = 128
#fbins = n_fft//2 + 1
spec_transform = transforms.Spectrogram(n_fft = n_fft, normalized = False)
label = int(filename.split("/")[-1].split("_")[0])
soundSource = filename.split("/")[-1].split("_")[1]
number = filename.split("/")[-1].split("_")[2]
wave, sample_rate = torchaudio.load_wav(filename)
spec = spec_transform(wave)
log_spec = (spec + 1e-9).log2()[0, :, :]
width = 65
height = log_spec.shape[0]
dim = (width, height)
log_spec = cv2.resize(log_spec.numpy(), dim, interpolation = cv2.INTER_AREA)
plt.figure()
plt.imshow(log_spec)
plt.show()
return log_spec, label, soundSource
def __init__(self, sample_rate=16000, n_fft=800, win_length=800, hop_length=200,
n_mels=80, rescale=True, rescaling_max=0.9, max_abs_value=4.,
preemphasis=0.97, preemphasize=True, fmin=55, fmax=7600,
min_level_db=-100, ref_level_db=20, symmetric_mels=True):
super(logFbankCal, self).__init__()
# these basic hyparams can be removed
self.sample_rate = sample_rate
self.n_fft = n_fft
self.win_length = win_length
self.hop_length = hop_length
self.n_mels = n_mels
self.fmin = fmin
self.fmax = fmax
self.rescale = rescale
self.rescaling_max = torch.tensor(rescaling_max, dtype=torch.float)
self.preemphasize = preemphasize
self.flipped_filter = torch.FloatTensor([-preemphasis, 1.]).unsqueeze(0).unsqueeze(0)
self.stftCal = transforms.Spectrogram(n_fft, win_length, hop_length, power=None)
mel_basis = librosa_mel_fn(sample_rate, n_fft, n_mels, fmin, fmax)
self.mel_basis = torch.from_numpy(mel_basis).float()
self.symmetric_mels = symmetric_mels
self.ref_level_db = torch.tensor(ref_level_db, dtype=torch.float)
self.min_level_db = torch.tensor(min_level_db, dtype=torch.float)
self.min_level = torch.tensor(np.exp(min_level_db / 20 * np.log(10)), dtype=torch.float)
self.max_abs_value = torch.tensor(max_abs_value, dtype=torch.float)
def test_spectrogram(self, kwargs):
# replication_pad1d_backward_cuda is not deteministic and
# gives very small (~2.7756e-17) difference.
#
# See https://github.com/pytorch/pytorch/issues/54093
transform = T.Spectrogram(**kwargs)
waveform = get_whitenoise(sample_rate=8000, duration=0.05, n_channels=2)
self.assert_grad(transform, [waveform], nondet_tol=1e-10)
def __init__(self, sample_rate, n_fft, win_length, hop_length, n_mels):
super(PhaseFbankCal, self).__init__()
self.complexSpec = transforms.Spectrogram(n_fft=n_fft,
win_length=win_length,
hop_length=hop_length,
power=None)
self.mel_scale = transforms.MelScale(n_mels=n_mels,
sample_rate=sample_rate,
n_stft=n_fft // 2 + 1)
def compute_stft(waveform: Tensor, n_fft: int, win_length: int,
hop_length: int) -> Tuple[Tensor, Tensor]:
device = waveform.device
spectrogram = _transf.Spectrogram(n_fft=n_fft,
win_length=win_length,
hop_length=hop_length,
power=None).to(device)
amplitude, phase = _func.magphase(spectrogram(waveform))
return amplitude, phase
def __init__(self, n_fft, hop_length):
"""Calculate spectrogram of a set of 1D signals.
The first dimension is batch/channel and the second should be time.
Arguments:
n_fft {int} -- Size of time window over which to calculate each FFT.
hop_length {int} -- The stride length between the start of each FFT window.
"""
self.spec_fn = transforms.Spectrogram(n_fft=n_fft,
hop_length=hop_length)
def test_roundtrip_spectrogram(self, **args):
"""Test the spectrogram + inverse spectrogram results in approximate identity."""
waveform = get_whitenoise(sample_rate=8000,
duration=0.5,
dtype=self.dtype)
s = T.Spectrogram(**args, power=None)
inv_s = T.InverseSpectrogram(**args)
transformed = s.forward(waveform)
restored = inv_s.forward(transformed, length=waveform.shape[-1])
self.assertEqual(waveform, restored, atol=1e-6, rtol=1e-6)
def __init__(self, d, src_path, batch_size, device, dropout_rate=0.5, encoder = 'ResNet'):
super(VAE, self).__init__()
'''
=========== ARGUMENTS ===========
> d - dimensionality of latent space
> src_path - path to source samples
> batch_size - number of training examples in single batch
> device - CPU or GPU in use
=================================
'''
if encoder == 'ResNet':
self.enc = ResNetBigger(d=d, dropout_rate=dropout_rate)
elif encoder == 'CNN':
self.enc = CNN(d=d, batch_size=batch_size, dropout_rate=dropout_rate)
elif encoder == 'linear':
self.enc == nn.Sequential(
nn.Linear(1025, 400),
nn.ReLU(),
nn.Linear(400, 100),
nn.ReLU(),
nn.Linear(100, 50),
nn.ReLU(),
nn.Linear(50, d)
)
self.dec = nn.Sequential(
nn.Linear(d, 40),
nn.ReLU(),
nn.Linear(40, 50),
nn.Softmax()
)
self.softmax = nn.Softmax(dim=0)
# For computing spectrogram then normalizing
self.spectrogram = T.Spectrogram(
n_fft=2048,
win_length=None,
hop_length=512,
power=2,
)
self.amp_to_db = T.AmplitudeToDB(stype='power')
self.src = torch.from_numpy(np.load(src_path))
self.src = self.src.unsqueeze(0).repeat(batch_size, 1, 1)
self.src = self.src.to(device)
self.d = d
self.global_step = 0
self.epoch = 0
def _wav_to_spec(path=None, wav=None, sr=sample_rate, engine='librosa'):
''' STFT Spectrogram with absolute values '''
if path is None and wav is None:
raise ValueError
if path is not None:
wav, _ = librosa.core.load(path, sr=None)
if engine == 'librosa':
return np.abs(librosa.stft(wav, **stft_params))
elif engine == 'torch':
return tf.Spectrogram(**stft_params,
power=power)(torch.from_numpy(wav))
raise ValueError(engine)
def spectro_gram(aud, spectro_type='mel', n_mels=64, n_fft=1024, hop_len=None):
sig,sr = aud
f_min, f_max, ws, top_db, pad = 0.0, None, None, 80, 0
# spec has shape [channel, n_mels, time], where channel is mono, stereo etc
if (spectro_type == 'mel'):
spec = transforms.MelSpectrogram(sr, n_fft, ws, hop_len, f_min, f_max, pad, n_mels)(sig)
elif (spectro_type == 'mfcc'):
pass
else:
spec = transforms.Spectrogram(n_fft, ws, hop_len, pad, normalize=False)(sig)
# Convert to decibels
spec = transforms.AmplitudeToDB(top_db=top_db)(spec)
return (spec)
def get_train_transforms(
config: object,
transforms_set: TformsSet = TformsSet.Audtorch) -> object:
if config.use_mels:
if transforms_set == TformsSet.TorchAudio:
trans = tforms_vision.Compose([
tforms_torch.Resample(orig_freq=44100,
new_freq=config.resampling_rate),
tforms_torch.MelSpectrogram(sample_rate=config.resampling_rate,
n_fft=config.n_fft,
win_length=config.hop_length,
hop_length=config.hop_length,
f_min=float(config.fmin),
f_max=float(config.fmax),
pad=0,
n_mels=config.n_mels),
tforms_torch.AmplitudeToDB(stype='power', top_db=80),
#tforms_aud.RandomCrop(config.max_length_frames), # Raises "Can't call numpy() on Variable that requires grad. Use var.detach().numpy() instead."
])
elif transforms_set == TformsSet.MySet: # this works
trans = tforms_aud.Compose([
tforms_torch.Resample(orig_freq=44100,
new_freq=config.resampling_rate),
tforms_mine.Spectrogram(config),
tforms_aud.RandomCrop(config.max_length_frames)
])
else:
if transforms_set == TformsSet.TorchAudio: # this works
trans = tforms_aud.Compose([
tforms_torch.Resample(orig_freq=44100,
new_freq=config.resampling_rate),
tforms_torch.Spectrogram(n_fft=config.n_fft,
win_length=config.hop_length,
hop_length=config.hop_length,
pad=0,
power=2,
normalized=True),
tforms_torch.AmplitudeToDB(stype='power', top_db=80),
tforms_aud.RandomCrop(config.max_length_frames)
])
elif transforms_set == TformsSet.MySet: # this works
trans = tforms_aud.Compose([
tforms_torch.Resample(orig_freq=44100,
new_freq=config.resampling_rate),
tforms_mine.Spectrogram(config),
tforms_aud.RandomCrop(config.max_length_frames)
])
return trans
def get_spectrogram(
n_fft = 400,
win_len = None,
hop_len = None,
power = 2.0,
):
waveform, _ = get_speech_sample()
spectrogram = T.Spectrogram(
n_fft=n_fft,
win_length=win_len,
hop_length=hop_len,
center=True,
pad_mode="reflect",
power=power,
)
return spectrogram(waveform)
def test_ecoacoustics_dm(root: Path):
dm = EcoacousticsDataModule(root=root,
segment_len=30.0,
target_attrs="habitat",
train_transforms=AT.Spectrogram())
dm.prepare_data()
dm.setup()
# Test loading a sample.
train_dl = dm.train_dataloader()
test_sample = next(iter(train_dl))
# Test size().
assert test_sample.x.size()[1] == dm.dims[0]
assert test_sample.x.size()[2] == dm.dims[1]
assert test_sample.x.size()[3] == dm.dims[2]
def __call__(self, data):
signal = data["signal"]
sr = data['sample_rate']
self.n_fft = int(np.ceil(0.025 * sr))
self.win_length = int(np.ceil(0.025 * sr))
self.hop_length = int(np.ceil(0.01 * sr))
spec = nn.Sequential(
T.Spectrogram(n_fft=self.n_fft,
win_length=self.win_length,
hop_length=self.hop_length), T.AmplitudeToDB())
data['Spectrogram'] = spec(signal)
data['input'] = spec(signal)
return data
def get_time_frequency_transform(config):
"""
Returns a nn.Sequential block to do a time-frequency transform, and crop to the desired size.
The spectrogram has shape: [batch, channels, freq_bins, frames]
:param config:
:return:
"""
if config.use_mels:
transformer = nn.Sequential(
tforms_torch.MelSpectrogram(sample_rate=config.new_fs,
n_fft=config.n_fft,
win_length=config.win_length,
hop_length=config.hop_length,
f_min=float(config.fmin),
f_max=float(config.fmax),
pad=0,
n_mels=config.n_mels),
#utils.make_module(tforms_mine.RandomCrop)(config.max_length_frames),
tforms_mine.RandomCrop(
(1, config.n_mels if config.use_mels else config.n_fft // 2 +
1, config.max_length_frames),
value=0),
tforms_torch.AmplitudeToDB(stype='power', top_db=80),
tforms_mine.ReScaleSpec([-1, 1]),
)
else:
transformer = nn.Sequential(
tforms_torch.Spectrogram(n_fft=config.n_fft,
win_length=config.win_length,
hop_length=config.hop_length,
pad=0,
power=2,
normalized=True),
tforms_mine.RandomCrop(
(1, config.n_mels if config.use_mels else config.n_fft // 2 +
1, config.max_length_frames),
value=0),
tforms_mine.AmplitudeToDB(stype='power', top_db=80),
#utils.make_module(tforms_mine.RandomCrop)(config.max_length_frames),
tforms_mine.ReScaleSpec([-1, 1]),
)
return transformer
def tfm_spectro(ad, sr):
# We must reshape signal for torchaudio to generate the spectrogram.
ws=512
hop=256
to_db_scale=False
n_fft=1024
f_min=0.0
f_max=-80
pad=0
n_mels=128
#mel = transforms.MelSpectrogram(sr, n_mels=n_mels, n_fft=n_fft, hop=hop, f_min=f_min, f_max=f_max, pad=pad)(ad)
sp = transforms.Spectrogram()(ad)
mel = transforms.MelScale()(sp)
#mel = mel.permute(0,2,1) # swap dimension, mostly to look sane to a human.
#if to_db_scale: mel = transforms.SpectrogramToDB(stype='magnitude', top_db=f_max)(mel)
#mel = mel.detach().numpy()
if to_db_scale:
mel = 20*torch.log10(mel)
return mel
def __init__(self,
table_path,
alphabet,
dataset_path='',
max_len=0,
preprocess='raw',
normalize=False,
pkwargs=None):
global transform, do_normalize
super(SpeechDataset, self).__init__()
self.table = pd.read_csv(table_path)
self.dataset_path = dataset_path
self.intencode = IntegerEncode(alphabet)
self.max_len = max_len
if preprocess == "mfcc":
transform = transforms.MFCC(sample_rate=pkwargs['sr'],
n_mfcc=pkwargs['num_features'])
elif preprocess == "spectrogram":
transform = transforms.Spectrogram(
n_fft=pkwargs['n_fft'], normalized=pkwargs['normalized'])
def test_Spectrogram(self, n_fft, hop_length, power):
sample_rate = 16000
waveform = get_whitenoise(
sample_rate=sample_rate,
n_channels=1,
).to(self.device, self.dtype)
expected = librosa.core.spectrum._spectrogram(
y=waveform[0].cpu().numpy(),
n_fft=n_fft,
hop_length=hop_length,
power=power)[0]
result = T.Spectrogram(
n_fft=n_fft,
hop_length=hop_length,
power=power,
).to(self.device, self.dtype)(waveform)[0]
self.assertEqual(result,
torch.from_numpy(expected),
atol=1e-5,
rtol=1e-5)
def compareTforms(config):
'''
Here I compare different transfromations sets for spectrograms, using (torchaudio, audtorch, and my own custom
spectrogram using librosa. This codes is applied to a sample audio file from the librispeech dataset.
This code was done mostly to post as an issue in github. As a minimal working example.
'''
config.use_mels = False
config.win_length = 400
config.hop_length = 400
config.n_fft = 2048
config.resampling_rate = 16000
augment1 = tforms2.Compose([
myTforms.ToTensor(),
tforms.Spectrogram(
n_fft=2048,
win_length=400, # 400 samples @ 16k = 25 ms,
hop_length=400,
pad=0,
power=2,
normalized=False),
tforms.AmplitudeToDB(stype='power', top_db=80)
])
augment2 = tforms2.Compose([
tforms2.Spectrogram(
window_size=400, # 400 samples @ 16k = 25 ms
hop_size=400,
fft_size=2048),
myTforms.ToTensor(),
tforms.AmplitudeToDB(stype='magnitude', top_db=80)
])
augment3 = tforms2.Compose([myTforms.Spectrogram(config)])
data1 = dsets.LibriSpeech(
root='/m/cs/work/falconr1/datasets/librespeech/LibriSpeech',
sets='dev-clean',
download=False,
transform=augment1)
data2 = dsets.LibriSpeech(
root='/m/cs/work/falconr1/datasets/librespeech/LibriSpeech',
sets='dev-clean',
download=False,
transform=augment2)
data3 = dsets.LibriSpeech(
root='/m/cs/work/falconr1/datasets/librespeech/LibriSpeech',
sets='dev-clean',
download=False,
transform=augment3)
plt.figure(figsize=(16, 8))
titles = ['torchaudio', 'audtorch', 'myset']
for i, data in enumerate([data1, data2, data3]):
spec, label = data[0]
if isinstance(spec, torch.Tensor):
spec = spec.numpy()
plt.subplot(1, 3, i + 1)
plt.imshow(spec.squeeze(),
interpolation='nearest',
cmap='inferno',
origin='lower',
aspect='auto')
plt.colorbar()
plt.title(titles[i])
plt.savefig(os.path.join('./results', 'Test_Output_compare_specs.png'))
plt.show()
def test_Spectrogram_return_complex(self):
tensor = torch.rand((1, 1000))
self._assert_consistency(
T.Spectrogram(power=None, return_complex=True), tensor)
def test_Spectrogram(self):
tensor = torch.rand((1, 1000))
self._assert_consistency(T.Spectrogram(), tensor)
# samples = [0, 1, 2, 2547]
# plot_samples(samples)
df_train, df_val = train_test_split(df,
train_size=SPLIT_RATIO,
test_size=1 - SPLIT_RATIO,
random_state=RANDOM_STATE)
print(len(df_train))
print(len(df_val))
print(f'-----------------\n' \
f'{df_train.loc[0]}')
spectrogram = T.Spectrogram(n_fft=128, hop_length=64)
classes = torch.Tensor(df.target.unique())
model = resnet18(pretrained=True)
model.conv1 = nn.Conv2d(1,
model.conv1.out_channels,
kernel_size=model.conv1.kernel_size[0],
stride=model.conv1.stride[0],
padding=model.conv1.padding[0])
num_ftrs = model.fc.in_features
model.fc = nn.Linear(num_ftrs, len(classes))
train = df.sample(frac=1, random_state=RANDOM_STATE)
X = train.file_path
y = train.target
model.fit(torch.Tensor(np.load(X[0])), target[0])
#
# To get the frequency make-up of an audio signal as it varies with time,
# you can use ``Spectrogram``.
#
waveform, sample_rate = get_speech_sample()
n_fft = 1024
win_length = None
hop_length = 512
# define transformation
spectrogram = T.Spectrogram(
n_fft=n_fft,
win_length=win_length,
hop_length=hop_length,
center=True,
pad_mode="reflect",
power=2.0,
)
# Perform transformation
spec = spectrogram(waveform)
print_stats(spec)
plot_spectrogram(spec[0], title='torchaudio')
######################################################################
# GriffinLim
# ----------
#
# To recover a waveform from a spectrogram, you can use ``GriffinLim``.
#
val_path = r'X:\DS Training Data\samples\val.npy'
val_data = Dataset(val_path)
val_loader = torch.utils.data.DataLoader(val_data,
drop_last=True,
**params)
test_path = r'X:\DS Training Data\samples\test.npy'
test_data = Dataset(test_path)
test_loader = torch.utils.data.DataLoader(test_data,
drop_last=True,
**params)
spectrogram = T.Spectrogram(
n_fft=2048,
win_length=None,
hop_length=512,
power=2,
)
amp_to_db = T.AmplitudeToDB(stype='power')
src_path = test_path = r'X:\DS Training Data\samples\src.npy'
model = VAE(d=32,
src_path=src_path,
batch_size=params['batch_size'],
device=device,
dropout_rate=0.25,
encoder='CNN').cuda()
print(model)
criterion = loss_function
def get_train_transforms(config: object,
set: TformsSet = TformsSet.Audtorch) -> object:
if config.use_mels:
if set == TformsSet.TorchAudio:
trans = transforms.Compose([
tforms2.Crop((441000, 441000 + 441000)),
tforms.MelSpectrogram(sample_rate=config.resampling_rate,
n_fft=config.n_fft,
win_length=config.hop_length,
hop_length=config.hop_length,
f_min=float(config.fmin),
f_max=float(config.fmax),
pad=0,
n_mels=config.n_mels),
tforms.AmplitudeToDB(stype='power', top_db=80),
# transforms.ToPILImage(),
# transforms.RandomCrop((96, 256), pad_if_needed=True,
# padding_mode='reflect'),
# transforms.ToTensor(),
])
elif set == TformsSet.Audtorch: ## no real mel spectrogram in audtorch
trans = tforms2.Compose([
myTforms.ToNumpy(),
tforms2.Crop((441000, 441000 + 441000)),
# tforms2.Normalize(),
tforms2.Spectrogram(
window_size=config.hop_length,
hop_size=config.hop_length,
fft_size=config.n_fft,
),
tforms2.Log(),
myTforms.ToTensor(),
tforms.AmplitudeToDB(stype='magnitude', top_db=80)
])
elif set == TformsSet.MySet:
trans = tforms2.Compose([
tforms2.Crop((441000, 441000 + 441000)),
myTforms.Spectrogram(config)
])
else:
if set == TformsSet.TorchAudio:
trans = transforms.Compose([
tforms2.Crop((441000, 441000 + 441000)),
tforms.Spectrogram(n_fft=config.n_fft,
win_length=config.hop_length,
hop_length=config.hop_length,
pad=0,
power=2,
normalized=True),
tforms.AmplitudeToDB(stype='power', top_db=80),
# tforms.MelSpectrogram(sample_rate=config.resampling_rate,
# n_fft=config.n_fft,
# win_length=config.hop_length,
# hop_length=config.hop_length,
# f_min=float(config.fmin),
# f_max=float(config.fmax),
# pad=0,
# n_mels=config.n_mels),
#transforms.ToPILImage(),
#transforms.RandomCrop((96, 256), pad_if_needed=True,
# padding_mode='reflect'),
#transforms.ToTensor(),
])
elif set == TformsSet.Audtorch:
trans = tforms2.Compose([
myTforms.ToNumpy(),
tforms2.Crop((441000, 441000 + 441000)),
#tforms2.Normalize(),
tforms2.Spectrogram(
window_size=config.hop_length,
hop_size=config.hop_length,
fft_size=config.n_fft,
),
myTforms.ToTensor(),
tforms.AmplitudeToDB(stype='magnitude', top_db=80)
])
elif set == TformsSet.MySet:
trans = tforms2.Compose([
tforms2.Crop((441000, 441000 + 441000)),
myTforms.Spectrogram(config)
])
return trans