Update configuration_wavtokenizer.py

configuration_wavtokenizer.py

@@ -85,20 +85,20 @@ class WavTokenizerConfig(PretrainedConfig):
         
         # Feature dimensions
         feature_dim: int = 512,
-        encoder_dim: int = 64,
+        encoder_dim: int = 32,
         encoder_rates: list = None,
         latent_dim: int = None,
         
         # Quantizer parameters
         codebook_size: int = 4096,
-        codebook_dim: int = 8,
+        codebook_dim: int = 512,
         num_quantizers: int = 1,
         
         # Backbone parameters
         backbone_type: str = "vocos",
-        backbone_dim: int = 512,
-        backbone_num_blocks: int = 8,
-        backbone_intermediate_dim: int = 1536,
+        backbone_dim: int = 768,
+        backbone_num_blocks: int = 12,
+        backbone_intermediate_dim: int = 2304,
         backbone_kernel_size: int = 7,
         backbone_layer_scale_init_value: float = 1e-6,
         
@@ -126,7 +126,7 @@ class WavTokenizerConfig(PretrainedConfig):
         # Feature dimensions
         self.feature_dim = feature_dim
         self.encoder_dim = encoder_dim
-        self.encoder_rates = encoder_rates if encoder_rates is not None else [8, 5, 4, 2]
+        self.encoder_rates = encoder_rates if encoder_rates is not None else [2, 4, 5, 8]
         self.latent_dim = latent_dim if latent_dim is not None else feature_dim
         
         # Quantizer
@@ -160,4 +160,4 @@ class WavTokenizerConfig(PretrainedConfig):
     @property
     def frame_rate(self) -> float:
         """Returns the frame rate (tokens per second)."""
-        return self.sample_rate / self.hop_length
+        return self.sample_rate / self.hop_length