🐛 | Changed training loop.

.gitignore

@@ -164,3 +164,4 @@ cython_debug/
 backup/
 dataset/
 old-output/
+*.wav

training.py

@@ -46,64 +46,39 @@ scheduler = torch.optim.lr_scheduler.ReduceLROnPlateau(optimizer_d, mode='min',
 # Training loop
 num_epochs = 500
 for epoch in range(num_epochs):
-    latest_crap_audio = torch.empty((2,3), dtype=torch.int64)
-    for high_quality, low_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)
+    original, crap_audio = torch.empty((1,2,3)), torch.empty((1,2,3))
+    for low_quality, high_quality in tqdm.tqdm(train_data_loader):
         low_quality = low_quality.to(device)
-
+        high_quality = high_quality.to(device)
         batch_size = low_quality.size(0)
-
-        # Labels
         real_labels = torch.ones(batch_size, 1).to(device)
         fake_labels = torch.zeros(batch_size, 1).to(device)
 
         # Train Discriminator
         optimizer_d.zero_grad()
-        outputs = discriminator(high_quality)
-        d_loss_real = criterion(outputs, real_labels)
-        d_loss_real.backward()
-
-        resampled_audio = generator(low_quality)
-
-        outputs = discriminator(resampled_audio.detach())
-        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)
-
+        real_outputs = discriminator(high_quality)
+        fake_audio = generator(low_quality)
+        fake_outputs = discriminator(fake_audio.detach())
+        d_loss_real = criterion(real_outputs, real_labels)
+        d_loss_fake = criterion(fake_outputs, fake_labels)
+        d_loss = (d_loss_real + d_loss_fake) * 0.5
+        d_loss.backward()
         optimizer_d.step()
 
-        d_loss = d_loss_real + d_loss_fake
-
         # Train Generator
         optimizer_g.zero_grad()
-        outputs = discriminator(resampled_audio)
-        g_loss = criterion(outputs, real_labels)
+        fake_outputs = discriminator(fake_audio)
+        g_loss = criterion(fake_outputs, real_labels)
         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()
-
-        scheduler.step(d_loss + g_loss)
-        latest_crap_audio = resampled_audio
+        original = high_quality
+        crap_audio = fake_audio
 
     if epoch % 10 == 0:
-        print(latest_crap_audio.size())
-        torchaudio.save(f"./epoch-{epoch}-audio.wav", latest_crap_audio[0].cpu(), 44100)
+        print(crap_audio.size())
+        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}]')
 
 torch.save(generator.state_dict(), "generator.pt")