update

examples/vm_sound_classification/conv2d_classifier.yaml

@@ -13,21 +13,21 @@ mel_spectrogram_param:
 conv2d_block_param_list:
   - batch_norm: true
     in_channels: 1
-    out_channels: 4
+    out_channels: 32
     kernel_size: 3
     stride: 1
     dilation: 3
     activation: relu
     dropout: 0.1
-  - in_channels: 4
-    out_channels: 4
+  - in_channels: 32
+    out_channels: 32
     kernel_size: 5
     stride: 2
     dilation: 3
     activation: relu
     dropout: 0.1
-  - in_channels: 4
-    out_channels: 4
+  - in_channels: 32
+    out_channels: 32
     kernel_size: 3
     stride: 1
     dilation: 2
@@ -35,7 +35,7 @@ conv2d_block_param_list:
     dropout: 0.1
 
 cls_head_param:
-  input_dim: 108
+  input_dim: 864
   num_layers: 2
   hidden_dims:
     - 128

examples/vm_sound_classification/run.sh

@@ -13,7 +13,7 @@ E:/programmer/asr_datasets/voicemail/wav_finished/id-ID/wav_finished/*/*.wav" \
 sh run.sh --stage 2 --stop_stage 2 --system_version windows --file_folder_name file_dir --final_model_name vm_sound_classification3
 sh run.sh --stage 3 --stop_stage 3 --system_version windows --file_folder_name file_dir --final_model_name vm_sound_classification3
 
-sh run.sh --stage 0 --stop_stage 5 --system_version centos --file_folder_name file_dir --final_model_name vm_sound_classification8-ch4 \
+sh run.sh --stage 0 --stop_stage 5 --system_version centos --file_folder_name file_dir --final_model_name vm_sound_classification8-ch32 \
 --filename_patterns "/data/tianxing/PycharmProjects/datasets/voicemail/*/wav_finished/*/*.wav"
 
 

examples/vm_sound_classification/step_3_train_model.py

@@ -181,7 +181,7 @@ def main():
     # )
     lr_scheduler = torch.optim.lr_scheduler.MultiStepLR(
         optimizer,
-        milestones=[10000, 20000, 30000], gamma=0.5
+        milestones=[10000, 20000, 30000, 40000, 50000], gamma=0.5
     )
     focal_loss = FocalLoss(
         num_classes=vocabulary.get_vocab_size(namespace="labels"),