🐛 | Changed training loop.
This commit is contained in:
1
.gitignore
vendored
1
.gitignore
vendored
@ -164,3 +164,4 @@ cython_debug/
|
|||||||
backup/
|
backup/
|
||||||
dataset/
|
dataset/
|
||||||
old-output/
|
old-output/
|
||||||
|
*.wav
|
||||||
|
BIN
output.wav
BIN
output.wav
Binary file not shown.
61
training.py
61
training.py
@ -46,64 +46,39 @@ scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer_d, mode='min',
|
|||||||
# Training loop
|
# Training loop
|
||||||
num_epochs = 500
|
num_epochs = 500
|
||||||
for epoch in range(num_epochs):
|
for epoch in range(num_epochs):
|
||||||
latest_crap_audio = torch.empty((2,3), dtype=torch.int64)
|
original, crap_audio = torch.empty((1,2,3)), torch.empty((1,2,3))
|
||||||
for high_quality, low_quality in tqdm.tqdm(train_data_loader):
|
for low_quality, high_quality in tqdm.tqdm(train_data_loader):
|
||||||
# Check for NaN values in input tensors
|
|
||||||
if torch.isnan(low_quality).any() or torch.isnan(high_quality).any():
|
|
||||||
continue
|
|
||||||
|
|
||||||
high_quality = high_quality.to(device)
|
|
||||||
low_quality = low_quality.to(device)
|
low_quality = low_quality.to(device)
|
||||||
|
high_quality = high_quality.to(device)
|
||||||
batch_size = low_quality.size(0)
|
batch_size = low_quality.size(0)
|
||||||
|
|
||||||
# Labels
|
|
||||||
real_labels = torch.ones(batch_size, 1).to(device)
|
real_labels = torch.ones(batch_size, 1).to(device)
|
||||||
fake_labels = torch.zeros(batch_size, 1).to(device)
|
fake_labels = torch.zeros(batch_size, 1).to(device)
|
||||||
|
|
||||||
# Train Discriminator
|
# Train Discriminator
|
||||||
optimizer_d.zero_grad()
|
optimizer_d.zero_grad()
|
||||||
outputs = discriminator(high_quality)
|
real_outputs = discriminator(high_quality)
|
||||||
d_loss_real = criterion(outputs, real_labels)
|
fake_audio = generator(low_quality)
|
||||||
d_loss_real.backward()
|
fake_outputs = discriminator(fake_audio.detach())
|
||||||
|
d_loss_real = criterion(real_outputs, real_labels)
|
||||||
resampled_audio = generator(low_quality)
|
d_loss_fake = criterion(fake_outputs, fake_labels)
|
||||||
|
d_loss = (d_loss_real + d_loss_fake) * 0.5
|
||||||
outputs = discriminator(resampled_audio.detach())
|
d_loss.backward()
|
||||||
d_loss_fake = criterion(outputs, fake_labels)
|
|
||||||
d_loss_fake.backward()
|
|
||||||
|
|
||||||
|
|
||||||
# Gradient clipping for discriminator
|
|
||||||
clip_value = 2.0
|
|
||||||
for param in discriminator.parameters():
|
|
||||||
if param.grad is not None:
|
|
||||||
param.grad.clamp_(-clip_value, clip_value)
|
|
||||||
|
|
||||||
optimizer_d.step()
|
optimizer_d.step()
|
||||||
|
|
||||||
d_loss = d_loss_real + d_loss_fake
|
|
||||||
|
|
||||||
# Train Generator
|
# Train Generator
|
||||||
optimizer_g.zero_grad()
|
optimizer_g.zero_grad()
|
||||||
outputs = discriminator(resampled_audio)
|
fake_outputs = discriminator(fake_audio)
|
||||||
g_loss = criterion(outputs, real_labels)
|
g_loss = criterion(fake_outputs, real_labels)
|
||||||
g_loss.backward()
|
g_loss.backward()
|
||||||
|
|
||||||
# Gradient clipping for generator
|
|
||||||
clip_value = 1.0
|
|
||||||
for param in generator.parameters():
|
|
||||||
if param.grad is not None:
|
|
||||||
param.grad.clamp_(-clip_value, clip_value)
|
|
||||||
|
|
||||||
optimizer_g.step()
|
optimizer_g.step()
|
||||||
|
original = high_quality
|
||||||
scheduler.step(d_loss + g_loss)
|
crap_audio = fake_audio
|
||||||
latest_crap_audio = resampled_audio
|
|
||||||
|
|
||||||
if epoch % 10 == 0:
|
if epoch % 10 == 0:
|
||||||
print(latest_crap_audio.size())
|
print(crap_audio.size())
|
||||||
torchaudio.save(f"./epoch-{epoch}-audio.wav", latest_crap_audio[0].cpu(), 44100)
|
torchaudio.save(f"./epoch-{epoch}-audio.wav", crap_audio[0].cpu(), 44100)
|
||||||
|
torchaudio.save(f"./epoch-{epoch}-audio-orig.wav", original[0].cpu(), 44100)
|
||||||
|
|
||||||
print(f'Epoch [{epoch+1}/{num_epochs}]')
|
print(f'Epoch [{epoch+1}/{num_epochs}]')
|
||||||
|
|
||||||
torch.save(generator.state_dict(), "generator.pt")
|
torch.save(generator.state_dict(), "generator.pt")
|
||||||
|
Reference in New Issue
Block a user