:albemic: | Tests.

training.py

@@ -38,9 +38,9 @@ device = torch.device(args.device if torch.cuda.is_available() else "cpu")
 print(f"Using device: {device}")
 
 mfcc_transform = T.MFCC(
-    sample_rate=44100,  # Adjust to your sample rate
+    sample_rate=44100,
     n_mfcc=20,
-    melkwargs={'n_fft': 2048, 'hop_length': 512} # adjust n_fft and hop_length to your needs.
+    melkwargs={'n_fft': 2048, 'hop_length': 256}
 ).to(device)
 
 def gpu_mfcc_loss(y_true, y_pred):
@@ -49,7 +49,8 @@ def gpu_mfcc_loss(y_true, y_pred):
     min_len = min(mfccs_true.shape[2], mfccs_pred.shape[2])
     mfccs_true = mfccs_true[:, :, :min_len]
     mfccs_pred = mfccs_pred[:, :, :min_len]
-    return torch.mean((mfccs_true - mfccs_pred)**2)
+    loss = torch.mean((mfccs_true - mfccs_pred)**2)
+    return loss
 
 def discriminator_train(high_quality, low_quality, real_labels, fake_labels):
     optimizer_d.zero_grad()
@@ -99,7 +100,7 @@ dataset = AudioDataset(dataset_dir, device)
 
 # ========= SINGLE =========
 
-train_data_loader = DataLoader(dataset, batch_size=128, shuffle=True)
+train_data_loader = DataLoader(dataset, batch_size=16, shuffle=True)
 
 # Initialize models and move them to device
 generator = SISUGenerator()
@@ -161,7 +162,7 @@ def start_training():
             if debug:
                 print(d_loss, combined_loss, adversarial_loss, mfcc_l)
             scheduler_d.step(d_loss)
-            scheduler_g.step(combined_loss)
+            #scheduler_g.step(combined_loss)
 
             # ========= SAVE LATEST AUDIO =========
             high_quality_audio = (high_quality_clip[0][0], high_quality_clip[1][0])