update

examples/conv_tasnet/step_1_prepare_data.py

@@ -10,8 +10,9 @@ import sys
 pwd = os.path.abspath(os.path.dirname(__file__))
 sys.path.append(os.path.join(pwd, "../../"))
 
-from tqdm import tqdm
 import librosa
+import numpy as np
+from tqdm import tqdm
 
 
 def get_args():
@@ -66,6 +67,8 @@ def target_second_signal_generator(data_dir: str, duration: int = 2, sample_rate
             signal_length = len(signal)
             win_size = int(duration * sample_rate)
             for begin in range(0, signal_length - win_size, win_size):
+                if np.sum(signal[begin: begin+win_size]) == 0:
+                    continue
                 row = {
                     "epoch_idx": epoch_idx,
                     "filename": filename.as_posix(),

toolbox/torchaudio/losses/perceptual.py

@@ -85,15 +85,13 @@ class PesqLoss(nn.Module):
         )
 
     def forward(self, denoise: torch.Tensor, clean: torch.Tensor):
-        max_val = torch.max(clean.abs())
-        if max_val == 0:
-            raise AssertionError
+
         batch_loss = self.loss_fn.forward(clean, denoise)
 
-        mask = ~(torch.isnan(batch_loss) | torch.isinf(batch_loss))
-        batch_loss = batch_loss[mask]
-        if len(batch_loss) == 0:
-            raise AssertionError
+        # mask = ~(torch.isnan(batch_loss) | torch.isinf(batch_loss))
+        # batch_loss = batch_loss[mask]
+        # if len(batch_loss) == 0:
+        #     raise AssertionError
 
         if self.reduction == "mean":
             loss = torch.mean(batch_loss)