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

Merge new-arch, because it has proven to give the best results #1

Merged
NikkeDoy merged 14 commits from new-arch into main 2025-04-30 23:47:41 +03:00
Conversation 0 Commits 14 Files Changed 8 +7 -7
1 changed files with 7 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files
Showing only changes of commit 3936b6c160 - Show all commits

14
training.py
View File

@ -101,7 +101,7 @@ dataset = AudioDataset(dataset_dir, device)
# ========= SINGLE =========
train_data_loader = DataLoader(dataset, batch_size=8, shuffle=True)
train_data_loader = DataLoader(dataset, batch_size=12, shuffle=True)
# Initialize models and move them to device
generator = SISUGenerator()
@ -118,7 +118,7 @@ if args.discriminator is not None:
discriminator.load_state_dict(torch.load(args.discriminator, map_location=device, weights_only=True))
# Loss
criterion_g = nn.MSELoss()
criterion_g = nn.BCEWithLogitsLoss()
criterion_d = nn.BCEWithLogitsLoss()
# Optimizers
@ -163,8 +163,8 @@ def start_training():
if debug:
print(d_loss, adversarial_loss)
scheduler_d.step(d_loss)
scheduler_g.step(adversarial_loss)
scheduler_d.step(d_loss.detach())
scheduler_g.step(adversarial_loss.detach())
# ========= SAVE LATEST AUDIO =========
high_quality_audio = (high_quality_clip[0][0], high_quality_clip[1][0])
@ -175,9 +175,9 @@ def start_training():
if generator_epoch % 10 == 0:
print(f"Saved epoch {new_epoch}!")
torchaudio.save(f"./output/epoch-{new_epoch}-audio-crap.wav", low_quality_audio[0].cpu(), high_quality_audio[1]) # <-- Because audio clip was resampled in data.py from original to crap and to original again.
torchaudio.save(f"./output/epoch-{new_epoch}-audio-ai.wav", ai_enhanced_audio[0].cpu(), ai_enhanced_audio[1])
torchaudio.save(f"./output/epoch-{new_epoch}-audio-orig.wav", high_quality_audio[0].cpu(), high_quality_audio[1])
torchaudio.save(f"./output/epoch-{new_epoch}-audio-crap.wav", low_quality_audio[0].cpu().detach(), high_quality_audio[1]) # <-- Because audio clip was resampled in data.py from original to crap and to original again.
torchaudio.save(f"./output/epoch-{new_epoch}-audio-ai.wav", ai_enhanced_audio[0].cpu().detach(), ai_enhanced_audio[1])
torchaudio.save(f"./output/epoch-{new_epoch}-audio-orig.wav", high_quality_audio[0].cpu().detach(), high_quality_audio[1])
if debug:
print(generator.state_dict().keys())