fix: resolve bugs in WarmupLR_withStepDecay and train config (#446)

* fix: resolve bugs in WarmupLR_withStepDecay and train config

* fix: keep cmvn default as True, disable via config

Author: shangguanqituan

examples/voxceleb/v2/conf/w2vbert_s1_lora.yaml

@@ -48,6 +48,14 @@ dataset_args:
       lora_dropout: 0.0
       bias: "none"
   # Note: w2vbert frontend handles feature extraction, no fbank_args needed
+  cmvn: False
+  spec_aug: False
+  spec_aug_args:
+    num_t_mask: 1
+    num_f_mask: 1
+    max_t: 10
+    max_f: 8
+    prob: 0.6
 
 # Model definition
 model: W2VBert_Adapter_MFA  # The new speaker model class

examples/voxceleb/v2/conf/w2vbert_s2_ft.yaml

@@ -42,6 +42,14 @@ dataset_args:
     frozen: False  # Unfreeze the base W2V-BERT model
     use_lora: False  # No LoRA in this stage
     lora_config_args: null
+  cmvn: False
+  spec_aug: False
+  spec_aug_args:
+    num_t_mask: 1
+    num_f_mask: 1
+    max_t: 10
+    max_f: 8
+    prob: 0.6
 
 # Model definition (load S1 model, unfreeze frontend, remove LoRA)
 model: W2VBert_Adapter_MFA

examples/voxceleb/v2/conf/w2vbert_s3_lmft.yaml

@@ -43,6 +43,14 @@ dataset_args:
     frozen: False  # Keep unfrozen
     use_lora: False
     lora_config_args: null
+  cmvn: False
+  spec_aug: False
+  spec_aug_args:
+    num_t_mask: 1
+    num_f_mask: 1
+    max_t: 10
+    max_f: 8
+    prob: 0.6
 
 # Model definition (load S2 model)
 model: W2VBert_Adapter_MFA

wespeaker/bin/train.py

@@ -176,7 +176,9 @@ def train(config='conf/config.yaml', **kwargs):
         logger.info("<== Loss ==>")
         logger.info("loss criterion is: " + configs['loss'])
 
-    configs['optimizer_args']['lr'] = configs['scheduler_args']['initial_lr']
+    if 'initial_lr' in configs['scheduler_args']:
+        configs['optimizer_args']['lr'] = configs['scheduler_args'].pop(
+            'initial_lr')
     optimizer = getattr(torch.optim,
                         configs['optimizer'])(ddp_model.parameters(),
                                               **configs['optimizer_args'])

wespeaker/utils/schedulers.py

@@ -111,7 +111,7 @@ def __init__(
         gamma: float = 0.1,
     ):
         self.optimizer = optimizer
-        self.warmup_step = warmup_step
+        self.warmup_step = warmup_step * epoch_iter
         self.num_epochs = num_epochs
         self.gamma = gamma
         self.epoch_iter = epoch_iter  # number of batches in each epoch