aam params

nnet/xsets.py

@@ -166,7 +166,7 @@ class SideSet(Dataset):
             self.transformation = dataset["train"]["transformation"]
         else:
             self.duration = dataset["eval"]["duration"]
-        self.transformation = dataset["eval"]["transformation"]
+            self.transformation = dataset["eval"]["transformation"]
 
         self.sample_number = int(self.duration * self.sample_rate)
 

nnet/xvector.py

@@ -408,6 +408,7 @@ class Xtractor(torch.nn.Module):
 
             self.preprocessor = MelSpecFrontEnd(n_mels=80)
             self.sequence_network = PreResNet34()
+            self.embedding_size = 256
 
             self.before_speaker_embedding = torch.nn.Linear(in_features=5120,
                                                             out_features=self.embedding_size)
@@ -444,8 +445,8 @@ class Xtractor(torch.nn.Module):
             if self.loss == "aam":
                 self.after_speaker_embedding = ArcMarginProduct(self.embedding_size,
                                                                 int(self.speaker_number),
-                                                                s = 30,
-                                                                m = 0.2,
+                                                                s = 20,
+                                                                m = 0.3,
                                                                 easy_margin = False)
 
             elif self.loss == 'aps':
@@ -1000,9 +1001,9 @@ def xtrain(speaker_number,
         param_list.append({'params': model.module.after_speaker_embedding.parameters(), 'weight_decay': model.module.after_speaker_embedding_weight_decay})
 
     optimizer = _optimizer(param_list, **_options)
-    scheduler = torch.optim.lr_scheduler.StepLR(optimizer,
-                                                step_size=20 * training_loader.__len__(),
-                                                gamma=0.5)
+    scheduler = scheduler = torch.optim.lr_scheduler.MultiStepLR(optimizer, 
+                                                                 milestones=[10000,50000,100000], 
+                                                                 gamma=0.5)
 
     if mixed_precision:
         scaler = torch.cuda.amp.GradScaler()