update

examples/dtln/run.sh

@@ -6,7 +6,13 @@ sh run.sh --stage 2 --stop_stage 2 --system_version windows --file_folder_name f
 --noise_dir "E:/Users/tianx/HuggingDatasets/nx_noise/data/noise" \
 --speech_dir "E:/Users/tianx/HuggingDatasets/nx_noise/data/speech"
 
-sh run.sh --stage 2 --stop_stage 2 --system_version centos --file_folder_name file_dir --final_model_name dtln-nx-dns3 \
+sh run.sh --stage 2 --stop_stage 2 --system_version centos --file_folder_name file_dir-256 --final_model_name dtln-256-nx-dns3 \
+--config_file "yaml/config-256.yaml" \
+--noise_dir "/data/tianxing/HuggingDatasets/nx_noise/data/noise/dns3-noise" \
+--speech_dir "/data/tianxing/HuggingDatasets/nx_noise/data/speech/dns3-speech"
+
+sh run.sh --stage 2 --stop_stage 2 --system_version centos --file_folder_name file_dir-512 --final_model_name dtln-512-nx-dns3 \
+--config_file "yaml/config-512.yaml" \
 --noise_dir "/data/tianxing/HuggingDatasets/nx_noise/data/noise/dns3-noise" \
 --speech_dir "/data/tianxing/HuggingDatasets/nx_noise/data/speech/dns3-speech"
 

examples/dtln/yaml/{config.yaml → config-256.yaml}

@@ -24,6 +24,6 @@ max_epochs: 100
 clip_grad_norm: 10.0
 seed: 1234
 
-batch_size: 64
 num_workers: 4
+batch_size: 64
 eval_steps: 15000

examples/dtln/yaml/config-512.yaml

@@ -0,0 +1,29 @@
+model_name: "DTLN"
+
+# spec
+sample_rate: 8000
+fft_size: 512
+hop_size: 128
+win_type: hann
+
+# data
+max_snr_db: 20
+min_snr_db: -10
+
+# model
+encoder_size: 512
+
+# train
+lr: 0.001
+lr_scheduler: "CosineAnnealingLR"
+lr_scheduler_kwargs:
+  T_max: 250000
+  eta_min: 0.0001
+
+max_epochs: 100
+clip_grad_norm: 10.0
+seed: 1234
+
+num_workers: 4
+batch_size: 64
+eval_steps: 15000

toolbox/torchaudio/models/dtln/modeling_dtln.py

@@ -133,7 +133,8 @@ class DTLNModel(nn.Module):
         )
 
         self.sep1 = SeperationBlock(input_size=(fft_size // 2 + 1),
-                                    hidden_size=128,
+                                    # hidden_size=128,
+                                    hidden_size=self.encoder_size // 2,
                                     dropout=0.25,
                                     )
 
@@ -148,7 +149,8 @@ class DTLNModel(nn.Module):
         self.encoder_norm1 = InstantLayerNormalization(channels=self.encoder_size)
 
         self.sep2 = SeperationBlock(input_size=self.encoder_size,
-                                    hidden_size=128,
+                                    # hidden_size=128,
+                                    hidden_size=self.encoder_size // 2,
                                     dropout=0.25,
                                     )
 

toolbox/torchaudio/models/dtln/yaml/config-160.yaml

@@ -11,7 +11,7 @@ min_snr_db: -10
 encoder_size: 256
 
 max_epochs: 100
-batch_size: 4
+batch_size: 64
 num_workers: 4
 seed: 1234
 eval_steps: 25000

toolbox/torchaudio/models/dtln/yaml/config-256.yaml

@@ -11,7 +11,7 @@ min_snr_db: -10
 encoder_size: 256
 
 max_epochs: 100
-batch_size: 4
+batch_size: 64
 num_workers: 4
 seed: 1234
 eval_steps: 25000

toolbox/torchaudio/models/dtln/yaml/config-512.yaml

@@ -0,0 +1,29 @@
+model_name: "DTLN"
+
+# spec
+sample_rate: 8000
+fft_size: 512
+hop_size: 128
+win_type: hann
+
+# data
+max_snr_db: 20
+min_snr_db: -10
+
+# model
+encoder_size: 512
+
+# train
+lr: 0.001
+lr_scheduler: "CosineAnnealingLR"
+lr_scheduler_kwargs:
+  T_max: 250000
+  eta_min: 0.0001
+
+max_epochs: 100
+clip_grad_norm: 10.0
+seed: 1234
+
+num_workers: 4
+batch_size: 64
+eval_steps: 15000