| """TIPSv2 DPT head configuration.""" |
|
|
| from transformers import PretrainedConfig |
|
|
|
|
| _BACKBONE_REPO_BY_HIDDEN_SIZE = { |
| 768: "google/tipsv2-b14", |
| 1024: "google/tipsv2-l14", |
| 1152: "google/tipsv2-so400m14", |
| 1536: "google/tipsv2-g14", |
| } |
|
|
|
|
| class TIPSv2DPTConfig(PretrainedConfig): |
| """Configuration for TIPSv2 DPT dense prediction heads.""" |
|
|
| model_type = "tipsv2_dpt" |
|
|
| def __init__( |
| self, |
| backbone_config=None, |
| neck_hidden_sizes=(96, 192, 384, 768), |
| fusion_hidden_size=256, |
| reassemble_factors=(4, 2, 1, 0.5), |
| readout_type="project", |
| num_depth_bins=256, |
| min_depth=1e-3, |
| max_depth=10.0, |
| **kwargs, |
| ): |
| super().__init__(**kwargs) |
| backbone_config = backbone_config or {} |
| hidden_size = backbone_config.get("hidden_size", 768) |
| out_indices = backbone_config.get("out_indices", (3, 6, 9, 12)) |
| self.backbone_repo = _BACKBONE_REPO_BY_HIDDEN_SIZE[hidden_size] |
| self.embed_dim = hidden_size |
| self.channels = fusion_hidden_size |
| self.post_process_channels = list(neck_hidden_sizes) |
| self.block_indices = [out_index - 1 for out_index in out_indices] |
| self.readout_type = readout_type |
| self.num_depth_bins = num_depth_bins |
| self.min_depth = min_depth |
| self.max_depth = max_depth |
| self.num_seg_classes = self.num_labels |
|
|
