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 test_compose(input, idx, axis):
t = transforms.Compose(
[transforms.Crop(idx, axis=axis),
transforms.Normalize(axis=axis)])
expected_output = F.crop(input, idx, axis=axis)
expected_output = F.normalize(expected_output, axis=axis)
assert np.array_equal(t(input), expected_output)
root = '/m/cs/work/falconr1/datasets/mtg-jamendo-dataset-master'
split = 0
start_id = config.start_id * config.chunk_size
stop_id = start_id + config.chunk_size
full_output_path = os.path.join(root, config.output_path)
dataset = JamendoAudioFolder_audtorch(
root,
config.subset,
split,
config.mode,
return_fname=True,
transform=tforms2.Compose([
tforms2.RandomCrop(config.max_length),
tforms2.Downmix(1),
tforms2.Normalize()
]),
)
## TODO use the audio dataset, not the npy
# dataset = JamendoAudioFolder_npy(root,
# config.subset,
# split,
# config.mode,
# trim_to_size=config.max_length,
# return_fname=True,
# transform=tforms2.Compose([
# tforms2.Downmix(1),
# tforms2.Normalize()]
sampling_rate=sampling_rate,
mean=mean,
stdev=stdev)
noise, label = next(iter(dataset))
samples = int(np.ceil(duration * sampling_rate))
assert noise.shape == (1, samples)
assert label == 'white noise'
assert -1 <= np.max(np.abs(noise)) <= 1
assert len(dataset) == 1
# --- datasets/utils.py ---
crop = transforms.RandomCrop(8192)
resamp1 = transforms.Resample(48000, 44100)
resamp2 = transforms.Resample(44100, 16000)
t1 = transforms.Compose([crop, resamp1])
t2 = transforms.Compose([crop, resamp1, resamp2])
t3 = transforms.Compose([resamp1, crop, resamp2])
d0 = datasets.WhiteNoise(duration=0.5, sampling_rate=48000, transform=crop)
d1 = datasets.WhiteNoise(duration=0.5, sampling_rate=48000, transform=t1)
d2 = datasets.WhiteNoise(duration=0.5, sampling_rate=48000, transform=t2)
d3 = datasets.WhiteNoise(duration=0.5, sampling_rate=48000, transform=t3)
df_empty = pd.DataFrame()
df_a = pd.DataFrame(data=[0], columns=['a'])
df_ab = pd.DataFrame(data=[('0', 1)], columns=['a', 'b'])
@pytest.mark.parametrize('list_of_datasets', [
(d2, d3), pytest.param([d0, d1], marks=xfail(raises=ValueError))
])
def test_audioconcatdataset(list_of_datasets):
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
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()