| """Configuration class for the Vesuvius ink-detection model. |
| |
| See modeling_inkdetection.py for the model itself. Designed to be |
| loaded with `AutoConfig.from_pretrained(..., trust_remote_code=True)`. |
| """ |
| from transformers import PretrainedConfig |
|
|
|
|
| class InkDetectionConfig(PretrainedConfig): |
| """Configuration for an InkDetectionModel. |
| |
| Architecture: ResNet3D-50 backbone (Hara et al., 2018) initialised |
| from Kinetics-700; per-stage max-pool over the z (depth) axis to |
| collapse 5-D features to 4-D; small 2-D U-Net decoder; 1x1 conv |
| output head producing one sigmoid logit channel. |
| """ |
| model_type = "inkdetection_resnet3d" |
|
|
| def __init__( |
| self, |
| in_channels: int = 1, |
| input_depth: int = 62, |
| input_size: int = 256, |
| backbone_depth: int = 50, |
| backbone_channels=(256, 512, 1024, 2048), |
| num_classes: int = 1, |
| decoder_upscale: int = 1, |
| |
| train_segment: str = "l_2", |
| train_inklabels: str = "l_2_inklabels.png", |
| train_steps: int = 12396, |
| train_tiles: int = 0, |
| train_loss_final: float = 0.0, |
| **kwargs, |
| ): |
| super().__init__(**kwargs) |
| self.in_channels = in_channels |
| self.input_depth = input_depth |
| self.input_size = input_size |
| self.backbone_depth = backbone_depth |
| self.backbone_channels = list(backbone_channels) |
| self.num_classes = num_classes |
| self.decoder_upscale = decoder_upscale |
| self.train_segment = train_segment |
| self.train_inklabels = train_inklabels |
| self.train_steps = train_steps |
| self.train_tiles = train_tiles |
| self.train_loss_final = train_loss_final |
|
|
