update

examples/conv_tasnet/step_2_train_model.py

@@ -207,7 +207,7 @@ def main():
     neg_si_snr_loss_fn = NegativeSISNRLoss(reduction="mean").to(device)
     neg_stoi_loss_fn = NegSTOILoss(sample_rate=8000, reduction="mean").to(device)
     mr_stft_loss_fn = MultiResolutionSTFTLoss(
-        # fft_size_list=[256, 512, 1024],
+        fft_size_list=[256, 512, 1024],
         win_size_list=[120, 240, 480],
         hop_size_list=[25, 50, 100],
         reduction="mean"

toolbox/torchaudio/losses/spectral.py

@@ -288,9 +288,9 @@ class MultiResolutionSTFTLoss(torch.nn.Module):
                  reduction: str = "mean",
                  ):
         super(MultiResolutionSTFTLoss, self).__init__()
-        fft_size_list = fft_size_list or [1024, 2048, 512]
-        win_size_list = win_size_list or [600, 1200, 240]
-        hop_size_list = hop_size_list or [120, 240, 50]
+        fft_size_list = fft_size_list or [512, 1024, 2048]
+        win_size_list = win_size_list or [240, 600, 1200]
+        hop_size_list = hop_size_list or [50, 120, 240]
 
         if not len(fft_size_list) == len(win_size_list) == len(hop_size_list):
             raise AssertionError