bugfixes

- disabling mp3 and vorbis codecs due to errors

nnet/augmentation.py

@@ -265,8 +265,8 @@ def data_augmentation(speech,
          configs = [
              ({"format": "wav", "encoding": 'ULAW', "bits_per_sample": 8}, "8 bit mu-law"),
              ({"format": "wav", "encoding": 'ALAW', "bits_per_sample": 8}, "8 bit a-law"),
-             ({"format": "mp3", "compression": -9}, "MP3"),
-             ({"format": "vorbis", "compression": -1}, "Vorbis")
+             #({"format": "mp3", "compression": -9}, "MP3"),
+             #({"format": "vorbis", "compression": -1}, "Vorbis")
          ]
          param, title = random.choice(configs)
          speech = torchaudio.functional.apply_codec(speech, sample_rate, **param)

nnet/xvector.py

@@ -634,10 +634,10 @@ class Xtractor(torch.nn.Module):
             elif self.loss == 'aps':
                 self.after_speaker_embedding = SoftmaxAngularProto(int(self.speaker_number))
 
-            self.preprocessor_weight_decay = 0.00002
-            self.sequence_network_weight_decay = 0.00002
-            self.stat_pooling_weight_decay = 0.00002
-            self.before_speaker_embedding_weight_decay = 0.00002
+            self.preprocessor_weight_decay = 0.0002
+            self.sequence_network_weight_decay = 0.0002
+            self.stat_pooling_weight_decay = 0.0002
+            self.before_speaker_embedding_weight_decay = 0.0002
             self.after_speaker_embedding_weight_decay = 0.0002
 
         elif model_archi == "rawnet2":
@@ -1067,6 +1067,8 @@ def update_training_dictionary(dataset_description,
     training_opts["scheduler"] = dict()
     training_opts["scheduler"]["type"] = "ReduceLROnPlateau"
     training_opts["scheduler"]["step_size_up"] = 10
+    training_opts["scheduler"]["epochs"] = 1
+    training_opts["scheduler"]["steps_per_epoch"] = 5994
     training_opts["scheduler"]["base_lr"] = 1e-8
     training_opts["scheduler"]["mode"] = "triangular2"
 
@@ -1619,7 +1621,7 @@ def xtrain(dataset_description,
                             scheduler,
                             device,
                             scaler=scaler,
-                            training_opts["mixed_precision"])
+                            mixed_precision=training_opts["mixed_precision"])
                             #aam_scheduler=aam_scheduler)
 
         # Cross validation