def do_noise_test(model, device):
"""
Test on the noisy data.
:param model: pytorch model to be tested
:type model: torch.nn.Module
:param device:
:type device: torch.device
"""
for noise in [0.0, 0.05, 0.1, 0.15, 0.2, 0.25, 0.3, 0.35, 0.4, 0.45, 0.5]:
noise_wavdata_to_tensor = [
LoadAudio(),
FixAudioLength(),
AddNoise(noise),
ToMelSpectrogram(n_mels=32),
ToTensor("mel_spectrogram", "input"),
Unsqueeze("input")
]
cachefile = "gsc_test_noise{}.npz".format("{:.2f}".format(noise)[2:])
test_dataset = dataset_from_wavfiles(EXTRACTPATH / "test",
noise_wavdata_to_tensor,
cachefilepath=DATAPATH /
cachefile)
test_loader = torch.utils.data.DataLoader(test_dataset,
batch_size=TEST_BATCH_SIZE)
results = test(model=model,
loader=test_loader,
criterion=F.nll_loss,
device=device)
print("Noise level: {}, Results: {}".format(noise, results))
desc="Processing audio")):
for xform in wavdata_to_tensor:
d = xform(d)
x[i] = d
y = np.array(ids, dtype=np.int)
print("Saving preprocessed data to {}".format(cachefilepath))
np.savez(cachefilepath, x, y)
if __name__ == "__main__":
random.seed(SEED)
np.random.seed(SEED)
test_wavdata_to_tensor = [
LoadAudio(),
FixAudioLength(),
ToMelSpectrogram(n_mels=32),
ToTensor("mel_spectrogram", "input"),
Unsqueeze("input"),
]
train_wavdata_to_tensor = [
LoadAudio(),
ChangeAmplitude(),
ChangeSpeedAndPitchAudio(),
FixAudioLength(),
ToSTFT(),
StretchAudioOnSTFT(),
TimeshiftAudioOnSTFT(),
FixSTFTDimension(),
def do_training(model, device):
"""
Train the model.
:param model: pytorch model to be trained
:type model: torch.nn.Module
:param device:
:type device: torch.device
"""
test_wavdata_to_tensor = [
LoadAudio(),
FixAudioLength(),
ToMelSpectrogram(n_mels=32),
ToTensor("mel_spectrogram", "input"),
Unsqueeze("input"),
]
valid_dataset = dataset_from_wavfiles(
EXTRACTPATH / "valid",
test_wavdata_to_tensor,
cachefilepath=DATAPATH / "gsc_valid.npz",
)
valid_loader = torch.utils.data.DataLoader(valid_dataset,
batch_size=VALID_BATCH_SIZE)
train_wavdata_to_tensor = [
LoadAudio(),
ChangeAmplitude(),
ChangeSpeedAndPitchAudio(),
FixAudioLength(),
ToSTFT(),
StretchAudioOnSTFT(),
TimeshiftAudioOnSTFT(),
FixSTFTDimension(),
ToMelSpectrogramFromSTFT(n_mels=32),
DeleteSTFT(),
ToTensor("mel_spectrogram", "input"),
Unsqueeze("input"),
]
sgd = optim.SGD(model.parameters(),
lr=LEARNING_RATE,
momentum=MOMENTUM,
weight_decay=WEIGHT_DECAY)
lr_scheduler = optim.lr_scheduler.StepLR(sgd,
step_size=1,
gamma=LEARNING_RATE_GAMMA)
for epoch in range(EPOCHS):
train_dataset = dataset_from_wavfiles(
EXTRACTPATH / "train",
train_wavdata_to_tensor,
cachefilepath=DATAPATH / "gsc_train{}.npz".format(epoch),
silence_percentage=0.1,
)
train_loader = torch.utils.data.DataLoader(
train_dataset,
batch_size=(FIRST_EPOCH_BATCH_SIZE
if epoch == 0 else TRAIN_BATCH_SIZE),
shuffle=True,
)
train(model=model,
loader=train_loader,
optimizer=sgd,
criterion=F.nll_loss,
device=device)
lr_scheduler.step()
model.apply(rezero_weights)
model.apply(update_boost_strength)
results = test(model=model,
loader=valid_loader,
criterion=F.nll_loss,
device=device)
print("Epoch {}: {}".format(epoch, results))