⚗️ | Added some stupid ways for training + some makeup

utils/MultiResolutionSTFTLoss.py

@@ -8,8 +8,9 @@ import torchaudio.transforms as T
 
 class MultiResolutionSTFTLoss(nn.Module):
     """
-    Computes a loss based on multiple STFT resolutions, including both
-    spectral convergence and log STFT magnitude components.
+    Multi-resolution STFT loss.
+    Combines spectral convergence loss and log-magnitude loss
+    across multiple STFT resolutions.
     """
 
     def __init__(
@@ -20,43 +21,67 @@ class MultiResolutionSTFTLoss(nn.Module):
         eps: float = 1e-7,
     ):
         super().__init__()
-        self.stft_transforms = nn.ModuleList(
-            [
-                T.Spectrogram(
-                    n_fft=n_fft, win_length=win_len, hop_length=hop_len, power=None
-                )
-                for n_fft, hop_len, win_len in zip(fft_sizes, hop_sizes, win_lengths)
-            ]
-        )
+
         self.eps = eps
+        self.n_resolutions = len(fft_sizes)
+
+        self.stft_transforms = nn.ModuleList()
+        for n_fft, hop_len, win_len in zip(fft_sizes, hop_sizes, win_lengths):
+            window = torch.hann_window(win_len)
+            stft = T.Spectrogram(
+                n_fft=n_fft,
+                hop_length=hop_len,
+                win_length=win_len,
+                window_fn=lambda _: window,
+                power=None,  # Keep complex output
+                center=True,
+                pad_mode="reflect",
+                normalized=False,
+            )
+            self.stft_transforms.append(stft)
 
     def forward(
         self, y_true: torch.Tensor, y_pred: torch.Tensor
     ) -> Dict[str, torch.Tensor]:
-        sc_loss = 0.0  # Spectral Convergence Loss
-        mag_loss = 0.0  # Log STFT Magnitude Loss
+        """
+        Args:
+            y_true: (B, T) or (B, 1, T) waveform
+            y_pred: (B, T) or (B, 1, T) waveform
+        """
+        # Ensure correct shape (B, T)
+        if y_true.dim() == 3 and y_true.size(1) == 1:
+            y_true = y_true.squeeze(1)
+        if y_pred.dim() == 3 and y_pred.size(1) == 1:
+            y_pred = y_pred.squeeze(1)
+
+        sc_loss = 0.0
+        mag_loss = 0.0
 
         for stft in self.stft_transforms:
-            stft.to(y_pred.device)  # Ensure transform is on the correct device
+            stft = stft.to(y_pred.device)
 
-            # Get complex STFTs
+            # Complex STFTs: (B, F, T, 2)
             stft_true = stft(y_true)
             stft_pred = stft(y_pred)
 
-            # Get magnitudes
+            # Magnitudes
             stft_mag_true = torch.abs(stft_true)
             stft_mag_pred = torch.abs(stft_pred)
 
             # --- Spectral Convergence Loss ---
-            # || |S_true| - |S_pred| ||_F  /  || |S_true| ||_F
             norm_true = torch.linalg.norm(stft_mag_true, dim=(-2, -1))
             norm_diff = torch.linalg.norm(stft_mag_true - stft_mag_pred, dim=(-2, -1))
             sc_loss += torch.mean(norm_diff / (norm_true + self.eps))
 
             # --- Log STFT Magnitude Loss ---
             mag_loss += F.l1_loss(
-                torch.log(stft_mag_pred + self.eps), torch.log(stft_mag_true + self.eps)
+                torch.log(stft_mag_pred + self.eps),
+                torch.log(stft_mag_true + self.eps),
             )
 
+        # Average across resolutions
+        sc_loss /= self.n_resolutions
+        mag_loss /= self.n_resolutions
         total_loss = sc_loss + mag_loss
+
         return {"total": total_loss, "sc": sc_loss, "mag": mag_loss}