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
elif config.platform == 3:
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),
@pytest.mark.parametrize('stdev', [1, 0.5])
def test_whitenoise(duration, sampling_rate, mean, stdev):
dataset = datasets.WhiteNoise(duration=duration,
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', [
def test_randomcrop(input, size, axis):
t = transforms.RandomCrop(size, axis=axis)
t.fix_randomization = True
assert np.array_equal(t(input), t(input))
assert np.array_equal(t(input), F.crop(input, t.idx, axis=t.axis))
def get_DataLoader(config):
train_transforms = get_Transforms()
if config.platform == 0:
root = '/Volumes/scratch/work/falconr1/datasets/mtg-jamendo-dataset-master'
elif config.platform == 2:
root = '/scratch/work/falconr1/datasets/mtg-jamendo-dataset-master'
elif config.platform == 3:
root = '/m/cs/work/falconr1/datasets/mtg-jamendo-dataset-master'
subset = config.subset
split = 0
mode = 'train'
if config.dataset == 'JamendoSpecFolder':
dataset = JamendoSpecFolder(root,
subset,
split,
mode,
spec_folder='data/processed/spec_npy',
transform=train_transforms)
elif config.dataset == 'JamendoSpecHDF5':
dataset = JamendoSpecHDF5(root,
subset,
split,
mode,
train_transforms,
hdf5_filename='data/processed/jamendo.hdf5')
elif config.dataset == 'JamendoSpecLMDB':
dataset = JamendoSpecLMDB(root,
subset,
split,
mode,
train_transforms,
lmdb_path='data/processed/triton')
elif config.dataset == 'JamendoSpecLMDBsubdir':
dataset = JamendoSpecLMDBsubdir(root,
subset,
split,
mode,
train_transforms,
lmdb_path='data/processed/chunks')
elif config.dataset == 'fake':
dataset = dset.FakeData(image_size=(1, 96, 1366),
transform=transforms.Compose([
transforms.RandomCrop((96, 256), pad_if_needed=True, padding_mode='reflect'),
transforms.ToTensor()
]))
elif config.dataset == 'SVHN':
dataset = dset.SVHN(root='/m/cs/work/falconr1/datasets/SVHN',
transform=transforms.Compose([
transforms.RandomCrop((96, 256), pad_if_needed=True, padding_mode='reflect'),
transforms.ToTensor()
]),
download=True)
elif config.dataset == 'JamendoAudioFolder_torchaudio':
dataset = JamendoAudioFolder_torchaudio(root,
subset,
split,
mode,
transform=transforms.Compose([
tforms.MelSpectrogram(sr=44100,
n_fft=512,
ws=256,
hop=256,
f_min=20.0,
f_max=8000,
pad=0,
n_mels=96),
transforms.ToPILImage(),
transforms.RandomCrop((96, 256), pad_if_needed=True, padding_mode='reflect'),
transforms.ToTensor(),
])
)
elif config.dataset == 'JamendoAudioFolder_audtorch':
dataset = JamendoAudioFolder_audtorch(root,
subset,
split,
mode,
## transform=tforms2.RandomCrop(size=256*44100),
# transform=tforms2.Compose([
# tforms2.Downmix(1),
# tforms2.Normalize(),
# tforms2.Spectrogram(window_size=256,
# hop_size=256,
# fft_size=512),
# tforms2.Log(),
# # tforms2.LogSpectrogram(window_size=256,
# # hop_size=256,
# # normalize=True),
# myTforms.Debugger(),
# myTforms.CFL2FLC(),
# transforms.ToPILImage(),
# transforms.RandomCrop((96, 256), pad_if_needed=True, padding_mode='reflect'),
# transforms.ToTensor(),
# ])
)
elif config.dataset == 'JamendoAudioFolder_npy':
dataset = JamendoAudioFolder_npy(root,
subset,
split,
mode,
trim_to_size=config.trim_size,
###transform=tforms2.Downmix(1),
#transform=tforms2.RandomCrop(size=30*44100),
# transform=tforms2.Compose([
# tforms2.Downmix(1),
# tforms2.Normalize(),
# tforms2.Spectrogram(window_size=256,
# hop_size=256,
# fft_size=512),
# tforms2.Log(),
# # tforms2.LogSpectrogram(window_size=256,
# # hop_size=256,
# # normalize=True),
# myTforms.Debugger(),
# myTforms.CFL2FLC(),
# transforms.ToPILImage(),
# transforms.RandomCrop((96, 256), pad_if_needed=True, padding_mode='reflect'),
# transforms.ToTensor(),
# ])
)
elif config.dataset == 'JamendoAudioFolder_torch':
dataset = JamendoAudioFolder_torch(root,
subset,
split,
mode,
###transform=tforms2.Downmix(1),
transform=tforms2.RandomCrop(size=30*44100),
# transform=tforms2.Compose([
# tforms2.Downmix(1),
# tforms2.Normalize(),
# tforms2.Spectrogram(window_size=256,
# hop_size=256,
# fft_size=512),
# tforms2.Log(),
# # tforms2.LogSpectrogram(window_size=256,
# # hop_size=256,
# # normalize=True),
# myTforms.Debugger(),
# myTforms.CFL2FLC(),
# transforms.ToPILImage(),
# transforms.RandomCrop((96, 256), pad_if_needed=True, padding_mode='reflect'),
# transforms.ToTensor(),
# ])
)
subset_indices = np.random.choice(range(len(dataset)), config.data_limit, replace=False)
print('------ Dataset length = {}, using {} samples.'.format(len(dataset), len(subset_indices)))
if config.collate_fn == 'seq2seq':
collate = Seq2Seq([-1,-1], batch_first=None, sort_sequences=False)
#collate = Seq2Seq_short([-1, -1], batch_first=None, sort_sequences=False)
else:
collate = torch.utils.data.dataloader.default_collate
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=config.batch_size,
# shuffle=True,
num_workers=config.num_workers,
pin_memory=True,
sampler=torch.utils.data.sampler.SubsetRandomSampler(subset_indices),
collate_fn=collate,
drop_last=True,
)
return dataloader