update

examples/vm_sound_classification/conv2d_classifier.yaml

@@ -18,21 +18,21 @@ spec_augment_param:
 conv2d_block_param_list:
   - batch_norm: true
     in_channels: 1
-    out_channels: 16
+    out_channels: 32
     kernel_size: 3
     stride: 1
     dilation: 3
     activation: relu
     dropout: 0.1
-  - in_channels: 16
-    out_channels: 16
+  - in_channels: 32
+    out_channels: 32
     kernel_size: 5
     stride: 2
     dilation: 3
     activation: relu
     dropout: 0.1
-  - in_channels: 16
-    out_channels: 16
+  - in_channels: 32
+    out_channels: 32
     kernel_size: 3
     stride: 1
     dilation: 2
@@ -40,11 +40,11 @@ conv2d_block_param_list:
     dropout: 0.1
 
 cls_head_param:
-  input_dim: 432
+  input_dim: 864
   num_layers: 2
   hidden_dims:
     - 128
     - 32
   activations: relu
   dropout: 0.1
-  num_labels: 4
+  num_labels: 2

examples/vm_sound_classification/run.sh

@@ -2,22 +2,16 @@
 
 : <<'END'
 
-sh run.sh --stage 0 --stop_stage 1 --system_version windows --file_folder_name file_dir --final_model_name vm_sound_classification4-ch16 \
+sh run.sh --stage 0 --stop_stage 1 --system_version windows --file_folder_name file_dir --final_model_name vm_sound_classification4-ch32 \
 --filename_patterns "E:/Users/tianx/HuggingDatasets/vm_sound_classification/data/wav_finished/wav_finished/en-US/wav_finished/*/*.wav \
 E:/Users/tianx/HuggingDatasets/vm_sound_classification/data/wav_finished/id-ID/wav_finished/*/*.wav" \
 --label_plan 4
 
-sh run.sh --stage 2 --stop_stage 2 --system_version windows --file_folder_name file_dir --final_model_name vm_sound_classification4-ch16 \
+sh run.sh --stage 2 --stop_stage 2 --system_version windows --file_folder_name file_dir --final_model_name vm_sound_classification2-ch32 \
 --filename_patterns "E:/Users/tianx/HuggingDatasets/vm_sound_classification/data/wav_finished/wav_finished/en-US/wav_finished/*/*.wav \
 E:/Users/tianx/HuggingDatasets/vm_sound_classification/data/wav_finished/id-ID/wav_finished/*/*.wav" \
 --label_plan 4
 
-sh run.sh --stage 2 --stop_stage 2 --system_version windows --file_folder_name file_dir --final_model_name vm_sound_classification3 --label_plan 4
-sh run.sh --stage 3 --stop_stage 3 --system_version windows --file_folder_name file_dir --final_model_name vm_sound_classification3 --label_plan 4
-
-sh run.sh --stage 0 --stop_stage 5 --system_version centos --file_folder_name file_dir --final_model_name vm_sound_classification4-ch16 \
---filename_patterns "/data/tianxing/PycharmProjects/datasets/voicemail/*/wav_finished/*/*.wav" --label_plan 4
-
 sh run.sh --stage 0 --stop_stage 5 --system_version centos --file_folder_name file_dir --final_model_name vm_sound_classification2-ch32 \
 --filename_patterns "/data/tianxing/PycharmProjects/datasets/voicemail/*/wav_finished/*/*.wav" --label_plan 2
 

toolbox/torchaudio/models/cnn_audio_classifier/modeling_cnn_audio_classifier.py

@@ -273,6 +273,12 @@ class WaveEncoder(nn.Module):
             conv2d_block_param_list=conv2d_block_param_list,
         )
 
+    @torch.jit.ignore
+    def do_spec_augment(self, spec: torch.Tensor) -> torch.Tensor:
+        if self.training:
+            spec = self.spec_augment.forward(spec)
+        return spec
+
     def forward(self, inputs: torch.Tensor):
         # x: [batch_size, spec_dim, seq_length]
         x = inputs
@@ -283,8 +289,7 @@ class WaveEncoder(nn.Module):
             x = x.log()
             x = x - torch.mean(x, dim=-1, keepdim=True)
 
-            if self.training:
-                x = self.spec_augment.forward(x)
+            x = self.do_spec_augment(x)
 
         x = x.transpose(1, 2)