| 11 | |
| 12 | @dataclass |
| 13 | class Options: |
| 14 | model_type = "Encoder" |
| 15 | mixed_precision: str = 'bf16' |
| 16 | gradient_accumulation_steps: int = 1 |
| 17 | |
| 18 | # dataset |
| 19 | root_path: str = "" |
| 20 | re10k_path: str = "" |
| 21 | co3d_path: str = "" |
| 22 | davis_path: str = "" |
| 23 | vos_path: str = "" |
| 24 | |
| 25 | batch_size: int = 32 |
| 26 | num_workers: int = 8 |
| 27 | resume: Optional[str] = None |
| 28 | |
| 29 | # Data augmentation control |
| 30 | use_augmentation: bool = True |
| 31 | |
| 32 | lr: float = 5e-4 |
| 33 | enable_depth: bool = True |
| 34 | depth_loss_type: str = "ssitrim" |
| 35 | depth_model_name: str = 'vitl' |
| 36 | num_epochs: int = 250 |
| 37 | warmup_iters: int = 50000 |
| 38 | lr_decay_epochs: int = 250 |
| 39 | gradient_clip: float = 1.0 |
| 40 | orth_proj: bool = True |
| 41 | lambda_dssim: float = 0.2 |
| 42 | forder: int = 1 |
| 43 | dynamic_type: str = "poly" |
| 44 | image_height: int = 288 |
| 45 | image_width: int = 512 |
| 46 | down_resolution: Tuple[int, int] = (288, 512) |
| 47 | input_frames: int = 1 |
| 48 | output_frames: int = 1 |
| 49 | background_color: Tuple[float, float, float] = (0.0, 0.0, 0.0) |
| 50 | lambda_lpips: float = 0.05 |
| 51 | lambda_depth: float = 0.05 |
| 52 | lambda_reg: float = 0.0 |
| 53 | ignore_large_loss: float = 0.0 |
| 54 | lpips_start_epoch : int = 50 |
| 55 | depth_start_epoch : int = 0 |
| 56 | dino_name: str = 'dinov2_vitb14_reg' |
| 57 | workspace: str = './workspace' |
| 58 | |
| 59 | # GS Encoder |
| 60 | in_channels: int = 3 |
| 61 | patch_size: int = 8 |
| 62 | checkpointing: bool = True |
| 63 | num_layers: int = 10 |
| 64 | hidden_dim: int = 768 |
| 65 | bwindow_size: Optional[int] = None # encoder window size |
| 66 | window_size: int = 2304 # upsampler window size |
| 67 | decoder_ratio: float = 2.0 |
| 68 | |
| 69 | # GS Decoder |
| 70 | decoder_dim: int = 64 |