| 12 | |
| 13 | @dataclass |
| 14 | class Options(BaseOptions): |
| 15 | model_type = "Decoder" |
| 16 | mixed_precision: str = 'bf16' |
| 17 | gradient_accumulation_steps: int = 1 |
| 18 | |
| 19 | # Dataset paths |
| 20 | root_path: str = "" |
| 21 | re10k_path: str = "" |
| 22 | co3d_path: str = "" |
| 23 | davis_path: str = "" |
| 24 | vos_path: str = "" |
| 25 | batch_size: int = 16 |
| 26 | num_workers: int = 8 |
| 27 | resume: Optional[str] = None |
| 28 | encoder_path: str = "PATH_TO_ENCODER_CHECKPOINT" |
| 29 | |
| 30 | lr: float = 5e-4 |
| 31 | enable_depth: bool = True |
| 32 | num_epochs: int = 200 |
| 33 | warmup_iters: int = 20000 |
| 34 | lr_decay_epochs: int = 200 |
| 35 | gradient_clip: float = 1.0 |
| 36 | forder: int = 1 |
| 37 | output_frames: int = 6 |
| 38 | lambda_lpips: float = 0.05 |
| 39 | lambda_depth: float = 0.01 |
| 40 | lambda_reg: float = 0.0 |
| 41 | lambda_mask: float = 3.0 |
| 42 | ignore_large_loss: float = 0.3 |
| 43 | lpips_start_epoch : int = 50 |
| 44 | depth_start_epoch : int = 0 |
| 45 | # workspace |
| 46 | workspace: str = './workspace' |
| 47 | |
| 48 | # Model architecture |
| 49 | patch_size: int = 8 |
| 50 | decoder_num_layers: int = 10 |
| 51 | decoder_hidden_dim: int = 768 |
| 52 | decoder_ratio: float = 2.0 |
| 53 | opacity_activation: str = "sigmoid" |
| 54 | use_augmentation: bool = True |
| 55 | |
| 56 | # Probabilistic sampling |
| 57 | use_pm: bool = True |
| 58 | fix_keys: List[str] = field(default_factory=lambda: ["rot_static", "rot_dynamic"]) |
| 59 | sample_keys: List[str] = field(default_factory=lambda: ["xyz_dynamic"]) |
| 60 | pred_keys: List[str] = field(default_factory=lambda: ["rgb", "opacity", "scale", "xyz_static"]) |
| 61 | |
| 62 | # Decoder-specific training |
| 63 | encoder_path: str = "" |
| 64 | use_dino: bool = True |
| 65 | drop_path_rate: float = 0.0 |
| 66 | pm_dynamic: bool = True |
| 67 | skip: bool = False |
| 68 | fix_opacity: bool = False |
| 69 | |
| 70 | # Training options |
| 71 | compile: bool = False |