BN bias

nnet/xvector.py

@@ -522,7 +522,7 @@ class Xtractor(torch.nn.Module):
             self.sequence_network = PreHalfResNet34()
             self.embedding_size = 256
             self.before_speaker_embedding = torch.nn.Sequential(OrderedDict([
-                                        ("emb1", torch.nn.Linear(in_features = 5120, out_features = self.embedding_size)),
+                                        ("emb1", torch.nn.Linear(in_features = 5120, out_features = self.embedding_size, bias=False)),
                                         ("bn1", torch.nn.BatchNorm1d(self.embedding_size))
                                     ]))
 
@@ -541,7 +541,7 @@ class Xtractor(torch.nn.Module):
             self.sequence_network_weight_decay = 0.00002
             self.stat_pooling_weight_decay = 0.00002
             self.before_speaker_embedding_weight_decay = 0.00002
-            self.after_speaker_embedding_weight_decay = 0.0002
+            self.after_speaker_embedding_weight_decay = 0 #0.0002
 
         elif model_archi == "rawnet2":
 
@@ -1222,7 +1222,7 @@ def get_optimizer(model, model_opts, train_opts, training_loader):
         scheduler = torch.optim.lr_scheduler.CyclicLR(optimizer=optimizer,
                                                       base_lr=1e-8,
                                                       max_lr=train_opts["lr"],
-                                                      step_size_up=training_loader.__len__() * 12,
+                                                      step_size_up=training_loader.__len__() * 16,
                                                       step_size_down=None,
                                                       cycle_momentum=cycle_momentum,
                                                       mode="triangular2")