NikkeDoy/SISU
1
0
Fork 0
Code Issues Pull Requests Actions Packages Projects Releases Wiki Activity

💩 | Hopefully this does something else.

Browse Source
This commit is contained in:
NikkeDoy
2024-12-18 02:55:57 +02:00
parent eea4e565bc
commit dcde20387a
1 changed files with 10 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

17
training.py
View File

@ -45,20 +45,21 @@ scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer_d, mode='min',
# Training loop # Training loop
num_epochs = 500 num_epochs = 500
lambda_gp = 10
for epoch in range(num_epochs): for epoch in range(num_epochs):
original, crap_audio = torch.empty((1,2,3)), torch.empty((1,2,3)) original = torch.empty((2))
crap_audio = torch.empty((2))
for low_quality, high_quality in tqdm.tqdm(train_data_loader): for low_quality, high_quality in tqdm.tqdm(train_data_loader):
low_quality = low_quality.to(device)
high_quality = high_quality.to(device) high_quality = high_quality.to(device)
batch_size = low_quality.size(0) low_quality = low_quality.to(device)
batch_size = high_quality.size(0)
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
optimizer_d.zero_grad()
real_outputs = discriminator(high_quality) real_outputs = discriminator(high_quality)
fake_audio = generator(low_quality) fake_outputs = discriminator(generator(low_quality))
fake_outputs = discriminator(fake_audio.detach())
d_loss_real = criterion(real_outputs, real_labels) d_loss_real = criterion(real_outputs, real_labels)
d_loss_fake = criterion(fake_outputs, fake_labels) d_loss_fake = criterion(fake_outputs, fake_labels)
d_loss = (d_loss_real + d_loss_fake) * 0.5 d_loss = (d_loss_real + d_loss_fake) * 0.5
@ -67,10 +68,12 @@ for epoch in range(num_epochs):
# Train Generator # Train Generator
optimizer_g.zero_grad() optimizer_g.zero_grad()
fake_audio = generator(low_quality)
fake_outputs = discriminator(fake_audio) fake_outputs = discriminator(fake_audio)
g_loss = criterion(fake_outputs, real_labels) g_loss = criterion(fake_outputs, real_labels)
g_loss.backward() g_loss.backward()
optimizer_g.step() optimizer_g.step()
original = high_quality original = high_quality
crap_audio = fake_audio crap_audio = fake_audio