update

examples/nx_denoise/step_2_train_model.py

@@ -288,8 +288,9 @@ def main():
             metric_g = discriminator.forward(clean_audios, audio_g.detach())
             loss_metric = F.mse_loss(metric_g.flatten(), one_labels)
 
-            # loss_gen_all = loss_mag * 0.9 + loss_pha * 0.3  + loss_com * 0.1 + loss_metric * 0.05 + loss_time * 0.2
-            loss_gen_all = loss_mag * 0.1 + loss_pha * 0.1  + loss_com * 0.1 + loss_metric * 0.9 + loss_time * 0.9
+            loss_gen_all = loss_mag * 0.9 + loss_pha * 0.3  + loss_com * 0.1 + loss_metric * 0.05 + loss_time * 0.2
+            # loss_gen_all = loss_mag * 0.1 + loss_pha * 0.1  + loss_com * 0.1 + loss_metric * 0.9 + loss_time * 0.9
+            # 2.02
 
             loss_gen_all.backward()
             optim_g.step()

toolbox/torchaudio/models/nx_denoise/transformers/transformers.py

@@ -456,19 +456,19 @@ def main2():
 def main():
 
     encoder = TSTransformerEncoder(
-        input_size=16,
-        hidden_size=64,
-        attention_heads=4,
-        num_blocks=4,
+        input_size=8,
+        hidden_size=16,
+        attention_heads=2,
+        num_blocks=2,
         dropout_rate=0.1,
     )
     # print(encoder)
 
-    x = torch.ones([4, 16, 200, 32])
+    x = torch.ones([4, 8, 200, 8])
     y = encoder.forward(xs=x)
     print(y.shape)
 
-    x = torch.ones([4, 16, 200, 32])
+    x = torch.ones([4, 8, 200, 8])
     y = encoder.forward_chunk_by_chunk(xs=x)
     print(y.shape)