update

examples/rnnoise/yaml/config.yaml

@@ -3,29 +3,33 @@ model_name: "rnnoise"
 # spec
 sample_rate: 8000
 segment_size: 32000
-nfft: 512
-win_size: 512
-hop_size: 256
+nfft: 160
+win_size: 160
+hop_size: 80
 win_type: hann
 
-# data
-max_snr_db: 20
-min_snr_db: -10
+erb_bins: 32
+min_freq_bins_for_erb: 2
 
 # model
 conv_size: 256
 gru_size: 256
 
-# train
-max_epochs: 100
-batch_size: 32
-num_workers: 4
-seed: 1234
+# data
+max_snr_db: 20
+min_snr_db: -10
 
+# train
 lr: 0.001
-lr_scheduler: CosineAnnealingLR
-lr_scheduler_kwargs: {}
+lr_scheduler: "CosineAnnealingLR"
+lr_scheduler_kwargs:
+  T_max: 250000
+  eta_min: 0.0001
 
-weight_decay: 0.00001
+max_epochs: 100
 clip_grad_norm: 10.0
-eval_steps: 20000
+seed: 1234
+
+batch_size: 64
+num_workers: 4
+eval_steps: 15000

toolbox/torchaudio/models/rnnoise/configuration_rnnoise.py

@@ -21,17 +21,16 @@ class RNNoiseConfig(PretrainedConfig):
                  min_snr_db: float = -10,
                  max_snr_db: float = 20,
 
-                 max_epochs: int = 100,
-                 batch_size: int = 4,
-                 num_workers: int = 4,
-                 seed: int = 1234,
-
                  lr: float = 0.001,
                  lr_scheduler: str = "CosineAnnealingLR",
                  lr_scheduler_kwargs: dict = None,
 
-                 weight_decay: float = 0.00001,
+                 max_epochs: int = 100,
                  clip_grad_norm: float = 10.,
+                 seed: int = 1234,
+
+                 batch_size: int = 64,
+                 num_workers: int = 4,
                  eval_steps: int = 25000,
 
                  **kwargs
@@ -53,17 +52,16 @@ class RNNoiseConfig(PretrainedConfig):
         self.min_snr_db = min_snr_db
         self.max_snr_db = max_snr_db
 
-        self.max_epochs = max_epochs
-        self.batch_size = batch_size
-        self.num_workers = num_workers
-        self.seed = seed
-
         self.lr = lr
         self.lr_scheduler = lr_scheduler
         self.lr_scheduler_kwargs = lr_scheduler_kwargs or dict()
 
-        self.weight_decay = weight_decay
+        self.max_epochs = max_epochs
         self.clip_grad_norm = clip_grad_norm
+        self.seed = seed
+
+        self.batch_size = batch_size
+        self.num_workers = num_workers
         self.eval_steps = eval_steps
 
 

toolbox/torchaudio/models/rnnoise/yaml/config.yaml

@@ -3,32 +3,33 @@ model_name: "rnnoise"
 # spec
 sample_rate: 8000
 segment_size: 32000
-nfft: 512
-win_size: 512
-hop_size: 256
+nfft: 160
+win_size: 160
+hop_size: 80
 win_type: hann
 
 erb_bins: 32
 min_freq_bins_for_erb: 2
 
-# data
-max_snr_db: 20
-min_snr_db: -10
-
 # model
 conv_size: 256
 gru_size: 256
 
-# train
-max_epochs: 100
-batch_size: 32
-num_workers: 4
-seed: 1234
+# data
+max_snr_db: 20
+min_snr_db: -10
 
+# train
 lr: 0.001
-lr_scheduler: CosineAnnealingLR
-lr_scheduler_kwargs: {}
+lr_scheduler: "CosineAnnealingLR"
+lr_scheduler_kwargs:
+  T_max: 250000
+  eta_min: 0.0001
 
-weight_decay: 0.00001
+max_epochs: 100
 clip_grad_norm: 10.0
-eval_steps: 20000
+seed: 1234
+
+batch_size: 64
+num_workers: 4
+eval_steps: 15000