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

⚗️ | Experiment with other layer layouts.

Browse Source
This commit is contained in:
NikkeDoy
2024-12-21 23:27:38 +02:00
parent b7d7e95c89
commit 70e20f53d4
6 changed files with 142 additions and 91 deletions
Download Patch File Download Diff File Expand all files Collapse all files

168
training.py
View File

@ -13,6 +13,60 @@ from data import AudioDataset
from generator import SISUGenerator
from discriminator import SISUDiscriminator
# Mel Spectrogram Loss
class MelSpectrogramLoss(nn.Module):
def __init__(self, sample_rate=44100, n_fft=2048, hop_length=512, n_mels=128):
super(MelSpectrogramLoss, self).__init__()
self.mel_transform = torchaudio.transforms.MelSpectrogram(
sample_rate=sample_rate,
n_fft=n_fft,
hop_length=hop_length,
n_mels=n_mels
).to(device) # Move to device
def forward(self, y_pred, y_true):
mel_pred = self.mel_transform(y_pred)
mel_true = self.mel_transform(y_true)
return F.l1_loss(mel_pred, mel_true)
def snr(y_true, y_pred):
noise = y_true - y_pred
signal_power = torch.mean(y_true ** 2)
noise_power = torch.mean(noise ** 2)
snr_db = 10 * torch.log10(signal_power / noise_power)
return snr_db
def discriminator_train(high_quality, low_quality, scale, real_labels, fake_labels):
optimizer_d.zero_grad()
discriminator_decision_from_real = discriminator(high_quality)
# TODO: Experiment with criterions HERE!
d_loss_real = criterion_d(discriminator_decision_from_real, real_labels)
generator_output = generator(low_quality, scale)
discriminator_decision_from_fake = discriminator(generator_output.detach())
# TODO: Experiment with criterions HERE!
d_loss_fake = criterion_d(discriminator_decision_from_fake, fake_labels)
d_loss = (d_loss_real + d_loss_fake) / 2.0
d_loss.backward()
nn.utils.clip_grad_norm_(discriminator.parameters(), max_norm=1.0) #Gradient Clipping
optimizer_d.step()
return d_loss
def generator_train(low_quality, scale, real_labels):
optimizer_g.zero_grad()
generator_output = generator(low_quality, scale)
discriminator_decision = discriminator(generator_output)
# TODO: Fix this shit
g_loss = criterion_g(discriminator_decision, real_labels)
g_loss.backward()
optimizer_g.step()
return generator_output
# Check for CUDA availability
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
print(f"Using device: {device}")
@ -27,8 +81,8 @@ val_size = int(dataset_size-train_size)
train_dataset, val_dataset = random_split(dataset, [train_size, val_size])
train_data_loader = DataLoader(train_dataset, batch_size=8, shuffle=True)
val_data_loader = DataLoader(val_dataset, batch_size=8, shuffle=True)
train_data_loader = DataLoader(train_dataset, batch_size=1, shuffle=True)
val_data_loader = DataLoader(val_dataset, batch_size=1, shuffle=True)
# Initialize models and move them to device
generator = SISUGenerator()
@ -39,6 +93,7 @@ discriminator = discriminator.to(device)
# Loss
criterion_g = nn.L1Loss()
criterion_g_mel = MelSpectrogramLoss().to(device)
criterion_d = nn.BCEWithLogitsLoss()
# Optimizers
@ -49,87 +104,80 @@ optimizer_d = optim.Adam(discriminator.parameters(), lr=0.0001, betas=(0.5, 0.99
scheduler_g = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer_g, mode='min', factor=0.5, patience=5)
scheduler_d = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer_d, mode='min', factor=0.5, patience=5)
def snr(y_true, y_pred):
noise = y_true - y_pred
signal_power = torch.mean(y_true ** 2)
noise_power = torch.mean(noise ** 2)
snr_db = 10 * torch.log10(signal_power / noise_power)
return snr_db
def discriminator_train(discriminator, optimizer, criterion, generator, real_labels, fake_labels, high_quality, low_quality):
optimizer.zero_grad()
discriminator_decision_from_real = discriminator(high_quality)
d_loss_real = criterion(discriminator_decision_from_real, real_labels)
generator_output = generator(low_quality)
discriminator_decision_from_fake = discriminator(generator_output.detach())
d_loss_fake = criterion(discriminator_decision_from_fake, fake_labels)
d_loss = (d_loss_real + d_loss_fake) / 2.0
d_loss.backward()
nn.utils.clip_grad_norm_(discriminator.parameters(), max_norm=1.0) #Gradient Clipping
optimizer.step()
# print(f"Discriminator Loss: {d_loss.item():.4f}, Mean Real Logit: {discriminator_decision_from_real.mean().item():.2f}, Mean Fake Logit: {discriminator_decision_from_fake.mean().item():.2f}")
def start_training():
# Training loop
# discriminator_epochs = 1000
generator_epochs = 500
for generator_epoch in range(generator_epochs):
low_quality_audio = torch.empty((1))
high_quality_audio = torch.empty((1))
ai_enhanced_audio = torch.empty((1))
# Training
for low_quality, high_quality in tqdm.tqdm(train_data_loader, desc=f"Epoch {generator_epoch+1}/{generator_epochs}"):
high_quality = high_quality.to(device)
low_quality = low_quality.to(device)
# ========= DISCRIMINATOR PRE-TRAINING =========
discriminator_epochs = 1
for discriminator_epoch in range(discriminator_epochs):
batch_size = high_quality.size(0)
# ========= TRAINING =========
for high_quality_clip, low_quality_clip in tqdm.tqdm(train_data_loader, desc=f"Epoch {discriminator_epoch+1}/{discriminator_epochs}"):
high_quality_sample = high_quality_clip[0].to(device)
low_quality_sample = low_quality_clip[0].to(device)
scale = high_quality_clip[0].shape[2]/low_quality_clip[0].shape[2]
# ========= LABELS =========
batch_size = high_quality_sample.size(0)
real_labels = torch.ones(batch_size, 1).to(device)
fake_labels = torch.zeros(batch_size, 1).to(device)
# Train Discriminator
# ========= DISCRIMINATOR =========
discriminator.train()
discriminator_train(high_quality_sample, low_quality_sample, scale, real_labels, fake_labels)
torch.save(discriminator.state_dict(), "models/discriminator-single-shot-pre-train.pt")
generator_epochs = 500
for generator_epoch in range(generator_epochs):
low_quality_audio = (torch.empty((1)), 1)
high_quality_audio = (torch.empty((1)), 1)
ai_enhanced_audio = (torch.empty((1)), 1)
# ========= TRAINING =========
for high_quality_clip, low_quality_clip in tqdm.tqdm(train_data_loader, desc=f"Epoch {generator_epoch+1}/{generator_epochs}"):
high_quality_sample = high_quality_clip[0].to(device)
low_quality_sample = low_quality_clip[0].to(device)
scale = high_quality_clip[0].shape[2]/low_quality_clip[0].shape[2]
# ========= LABELS =========
batch_size = high_quality_clip[0].size(0)
real_labels = torch.ones(batch_size, 1).to(device)
fake_labels = torch.zeros(batch_size, 1).to(device)
# ========= DISCRIMINATOR =========
discriminator.train()
for _ in range(3):
discriminator_train(discriminator, optimizer_d, criterion_d, generator, real_labels, fake_labels, high_quality, low_quality)
discriminator_train(high_quality_sample, low_quality_sample, scale, real_labels, fake_labels)
# Train Generator
# ========= GENERATOR =========
generator.train()
optimizer_g.zero_grad()
generator_output = generator_train(low_quality_sample, scale, real_labels)
# Generator loss: how well fake data fools the discriminator
generator_output = generator(low_quality)
discriminator_decision = discriminator(generator_output) # No detach here
g_loss = criterion_g(discriminator_decision, real_labels) # Train generator to produce real-like outputs
# ========= SAVE LATEST AUDIO =========
high_quality_audio = high_quality_clip
low_quality_audio = low_quality_clip
ai_enhanced_audio = (generator_output, high_quality_clip[1])
g_loss.backward()
optimizer_g.step()
low_quality_audio = low_quality
high_quality_audio = high_quality
ai_enhanced_audio = generator_output
metric = snr(high_quality_audio, ai_enhanced_audio)
metric = snr(high_quality_audio[0].to(device), ai_enhanced_audio[0])
print(f"Generator metric {metric}!")
scheduler_g.step(metric)
if generator_epoch % 10 == 0:
print(f"Saved epoch {generator_epoch}!")
torchaudio.save(f"./output/epoch-{generator_epoch}-audio-crap.wav", low_quality_audio[0].cpu(), 44100)
torchaudio.save(f"./output/epoch-{generator_epoch}-audio-ai.wav", ai_enhanced_audio[0].cpu(), 44100)
torchaudio.save(f"./output/epoch-{generator_epoch}-audio-orig.wav", high_quality_audio[0].cpu(), 44100)
torchaudio.save(f"./output/epoch-{generator_epoch}-audio-crap.wav", low_quality_audio[0][0].cpu(), low_quality_audio[1])
torchaudio.save(f"./output/epoch-{generator_epoch}-audio-ai.wav", ai_enhanced_audio[0][0].cpu(), ai_enhanced_audio[1])
torchaudio.save(f"./output/epoch-{generator_epoch}-audio-orig.wav", high_quality_audio[0][0].cpu(), high_quality_audio[1])
if generator_epoch % 50 == 0:
torch.save(discriminator.state_dict(), "discriminator.pt")
torch.save(generator.state_dict(), "generator.pt")
torch.save(discriminator.state_dict(), f"models/epoch-{generator_epoch}-discriminator.pt")
torch.save(generator.state_dict(), f"models/epoch-{generator_epoch}-generator.pt")
torch.save(discriminator.state_dict(), "discriminator.pt")
torch.save(generator.state_dict(), "generator.pt")
torch.save(discriminator.state_dict(), "models/epoch-500-discriminator.pt")
torch.save(generator.state_dict(), "models/epoch-500-generator.pt")
print("Training complete!")
start_training()