| """ |
| Configuration for CAL-MOS Hugging Face models. |
| |
| CAL-MOS: Bridging Layers with Adapters for Robust MOS Prediction Across |
| Speech Foundation Models, Interspeech 2026. |
| """ |
|
|
| from transformers import PretrainedConfig |
|
|
|
|
| class CalMOSConfig(PretrainedConfig): |
| """Configuration for the CAL-MOS Adapters + Mean (A+M) model.""" |
|
|
| model_type = "calmos" |
|
|
| def __init__( |
| self, |
| backbone_name=None, |
| num_feature_layers=None, |
| adapter_input_dim=None, |
| adapter_dim=None, |
| mlp_input_dim=None, |
| mlp_hidden_dim=None, |
| mlp_num_layers=1, |
| mlp_output_size=1, |
| mlp_dropout=0.1, |
| target_sr=16000, |
| mos_min=1.0, |
| mos_max=5.0, |
| layer_weight_strategy="layer_adapter", |
| pooling_strategy="mean", |
| freeze_backbone=True, |
| **kwargs, |
| ): |
| super().__init__(**kwargs) |
|
|
| self.backbone_name = backbone_name |
| self.num_feature_layers = num_feature_layers |
| self.adapter_input_dim = adapter_input_dim |
| self.adapter_dim = adapter_dim |
| self.mlp_input_dim = mlp_input_dim |
| self.mlp_hidden_dim = mlp_hidden_dim |
| self.mlp_num_layers = mlp_num_layers |
| self.mlp_output_size = mlp_output_size |
| self.mlp_dropout = mlp_dropout |
|
|
| self.target_sr = target_sr |
| self.mos_min = mos_min |
| self.mos_max = mos_max |
|
|
| self.layer_weight_strategy = layer_weight_strategy |
| self.pooling_strategy = pooling_strategy |
| self.freeze_backbone = freeze_backbone |
|
|
