update

examples/cnn_vad_by_webrtcvad/step_4_train_model.py

@@ -272,7 +272,7 @@ def main():
             dice_loss = dice_loss_fn.forward(probs, targets)
             lsnr_loss = model.lsnr_loss_fn(lsnr, clean_audios, noisy_audios)
 
-            loss = 1.0 * bce_loss + 1.0 * dice_loss + 0.3 * lsnr_loss
+            loss = 1.0 * bce_loss + 1.0 * dice_loss + 0.03 * lsnr_loss
             if torch.any(torch.isnan(loss)) or torch.any(torch.isinf(loss)):
                 logger.info(f"find nan or inf in loss. continue.")
                 continue
@@ -352,7 +352,7 @@ def main():
                         dice_loss = dice_loss_fn.forward(probs, targets)
                         lsnr_loss = model.lsnr_loss_fn(lsnr, clean_audios, noisy_audios)
 
-                        loss = 1.0 * bce_loss + 1.0 * dice_loss + 0.3 * lsnr_loss
+                        loss = 1.0 * bce_loss + 1.0 * dice_loss + 0.03 * lsnr_loss
                         if torch.any(torch.isnan(loss)) or torch.any(torch.isinf(loss)):
                             logger.info(f"find nan or inf in loss. continue.")
                             continue

toolbox/torchaudio/models/vad/cnn_vad/modeling_cnn_vad.py

@@ -201,10 +201,6 @@ class CNNVadModel(nn.Module):
             raise AssertionError("Input signals must have the same shape")
         noise = noisy - clean
 
-        print(f"lsnr: {lsnr.shape}")
-        print(f"clean: {clean.shape}")
-        print(f"noisy: {noisy.shape}")
-
         if clean.dim() == 2:
             clean = torch.unsqueeze(clean, dim=1)
         if noise.dim() == 2:
@@ -227,9 +223,6 @@ class CNNVadModel(nn.Module):
         lsnr_gth = self.lsnr_fn.forward(stft_clean, stft_noise)
         # lsnr_gth shape: [b, t]
 
-        print(f"lsnr: {lsnr.shape}")
-        print(f"lsnr_gth: {lsnr_gth.shape}")
-
         loss = F.mse_loss(lsnr, lsnr_gth)
         return loss