| from typing import List, Literal, Optional, Any |
| from zencfg import ConfigBase |
|
|
|
|
| class ModelConfig(ConfigBase): |
| arch: str |
| data_channels: int |
| out_channels: int |
|
|
|
|
| class FNOConfig(ModelConfig): |
| model_arch: str = "fno" |
| data_channels: int |
| out_channels: int |
| n_modes: List[int] |
| hidden_channels: int |
| lifting_channel_ratio: int = 2 |
| projection_channel_ratio: int = 4 |
| n_layers: int = 4 |
| domain_padding: float = 0.0 |
| norm: str = "None" |
| fno_skip: str = "linear" |
| implementation: str = "factorized" |
| use_channel_mlp: bool = True |
| channel_mlp_expansion: float = 0.5 |
| channel_mlp_dropout: float = 0 |
| separable: bool = False |
| factorization: str = "None" |
| rank: float = 1.0 |
| fixed_rank_modes: bool = False |
| stabilizer: str = "None" |
|
|
|
|
| class SimpleFNOConfig(FNOConfig): |
| """ |
| SimpleFNOConfig: a basic FNO config that provides access to only |
| the most important FNO parameters. |
| """ |
|
|
| data_channels: int |
| out_channels: int |
| n_modes: List[int] |
| hidden_channels: int |
| projection_channel_ratio: int |
|
|
|
|
| class FNO_Small2d(SimpleFNOConfig): |
| """ |
| FNO_Small2d: a basic, small FNO for 2d problems. |
| """ |
|
|
| data_channels: int = 1 |
| out_channels: int = 1 |
| n_modes: List[int] = [16, 16] |
| hidden_channels: int = 24 |
| projection_channel_ratio: int = 2 |
|
|
|
|
| class FNO_Medium2d(SimpleFNOConfig): |
| """ |
| FNO_Medium2d: a basic, medium-sized FNO for 2d problems. |
| """ |
|
|
| data_channels: int = 1 |
| out_channels: int = 1 |
| n_modes: List[int] = [64, 64] |
| hidden_channels: int = 64 |
| projection_channel_ratio: int = 4 |
|
|
|
|
| class FNO_Large2d(SimpleFNOConfig): |
| """ |
| FNO_Large2d: a large FNO for 2d problems. |
| """ |
|
|
| data_channels: int = 1 |
| out_channels: int = 1 |
| n_modes: List[int] = [64, 64] |
| hidden_channels: int = 128 |
| projection_channel_ratio: int = 2 |
|
|
|
|
| class FNO_Huge2d(SimpleFNOConfig): |
| """ |
| FNO_Huge2d: a giant FNO for 2d problems. |
| |
| Note that this will likely be too large for most single-GPU local settings, |
| and should be paired with tensor factorization, low-rank/sparse optimization |
| via TensorGRaD, or some other means of reducing memory overhead. |
| """ |
|
|
| data_channels: int = 1 |
| out_channels: int = 1 |
| n_modes: List[int] = [128, 128] |
| hidden_channels: int = 128 |
| projection_channel_ratio: int = 2 |
|
|
|
|
| class FNO_Medium3d(SimpleFNOConfig): |
| """ |
| FNO_Medium3d: a medium FNO for 3d problems. |
| """ |
|
|
| data_channels: int = 1 |
| out_channels: int = 1 |
| n_modes: List[int] = [32, 32, 32] |
| hidden_channels: int = 64 |
| projection_channel_ratio: int = 4 |
|
|
|
|
| class FNOGNOConfig(ModelConfig): |
| model_arch: str = "fnogno" |
| data_channels: int |
| out_channels: int |
| gno_coord_dim: int |
| gno_coord_embed_dim: int |
| gno_radius: float |
| gno_transform_type: str |
| fno_n_modes: List[int] |
| fno_hidden_channels: int |
| fno_use_channel_mlp: bool = True |
| fno_norm: str = "instance_norm" |
| fno_ada_in_features: int = 32 |
| fno_factorization: str = "tucker" |
| fno_rank: float = 1.0 |
| fno_domain_padding: float = 0.125 |
| fno_use_channel_mlp: bool = True |
| fno_channel_mlp_expansion: float = 1.0 |
| fno_resolution_scaling_factor: int = 1 |
|
|
|
|
| class FNOGNO_Small3d(FNOGNOConfig): |
| data_channels: int = 1 |
| out_channels: int = 1 |
| gno_coord_dim: int = 3 |
| gno_coord_embed_dim: int = 16 |
| gno_radius: float = 0.033 |
| gno_transform_type: str = "linear" |
| fno_n_modes: List[int] = [16, 16, 16] |
| fno_hidden_channels: int = 64 |
| fno_use_channel_mlp: bool = True |
| fno_rank: float = 0.4 |
| fno_domain_padding: float = 0.125 |
|
|
|
|
| class GINOConfig(ModelConfig): |
| model_arch: str = "gino" |
| data_channels: int |
| out_channels: int |
| latent_feature_channels: Optional[int] = None |
| gno_coord_dim: int = 3 |
| gno_coord_embed_dim: int = 16 |
| gno_radius: float = 0.033 |
| in_gno_transform_type: str = "linear" |
| out_gno_transform_type: str = "linear" |
| gno_pos_embed_type: str = "nerf" |
| gno_weighting_function: Optional[str] = None |
| gno_weight_function_scale: Optional[float] = None |
| fno_n_modes: List[int] = [16, 16, 16] |
| fno_hidden_channels: int = 64 |
| fno_use_channel_mlp: bool = True |
| fno_norm: str = "instance_norm" |
| fno_ada_in_features: int = 32 |
| fno_factorization: str = "tucker" |
| fno_rank: float = 0.4 |
| fno_domain_padding: float = 0.125 |
| fno_channel_mlp_expansion: float = 1.0 |
| fno_resolution_scaling_factor: int = 1 |
|
|
|
|
| class GINO_Small3d(GINOConfig): |
| data_channels: int = 0 |
| out_channels: int = 1 |
| latent_feature_channels: Optional[int] = 1 |
| gno_coord_dim: int = 3 |
| gno_coord_embed_dim: int = 16 |
| gno_radius: float = 0.033 |
|
|
|
|
| class GINO_Poisson2d(GINOConfig): |
| data_channels: int = 3 |
| out_channels: int = 1 |
| projection_channel_ratio: int = 4 |
| gno_coord_dim: int = 2 |
| in_gno_pos_embed_type: Optional[Any] = None |
| out_gno_pos_embed_type: str = "transformer" |
| gno_embed_channels: int = 16 |
| gno_embed_max_positions: int = 600 |
| gno_use_torch_scatter: bool = True |
| in_gno_radius: float = 0.16 |
| out_gno_radius: float = 0.175 |
| in_gno_transform_type: str = "linear" |
| out_gno_transform_type: str = "linear" |
| gno_reduction: str = "mean" |
| gno_weighting_function: str = "half_cos" |
| gno_weight_function_scale: float = 0.030625 |
| gno_use_open3d: bool = False |
| in_gno_channel_mlp_hidden_layers: List[int] = [256, 512, 256] |
| out_gno_channel_mlp_hidden_layers: List[int] = [512, 1024, 512] |
| in_gno_tanh: Optional[Any] = None |
| out_gno_tanh: Optional[Any] = None |
| fno_n_modes: List[int] = [20, 20] |
| fno_hidden_channels: int = 64 |
| fno_lifting_channel_ratio: int = 4 |
| fno_n_layers: int = 4 |
| fno_use_channel_mlp: bool = True |
| fno_channel_mlp_expansion: float = 0.5 |
| fno_norm: str = "group_norm" |
| fno_ada_in_features: int = 8 |
| fno_factorization: Optional[Any] = None |
| fno_rank: float = 0.8 |
| fno_domain_padding: float = 0.0 |
|
|
|
|
| class OTNOConfig(ModelConfig): |
| model_arch: str = "otno" |
| n_modes: List[int] |
| hidden_channels: int |
| data_channels: int = 9 |
| out_channels: int = 1 |
| lifting_channel_ratio: int = 3 |
| projection_channel_ratio: int = 3 |
| norm: str = "group_norm" |
| use_channel_mlp: bool = True |
| channel_mlp_expansion: float = 1.0 |
| domain_padding: float = 0.125 |
| factorization: str = "tucker" |
| rank: float = 0.4 |
|
|
|
|
| class OTNO_Small2d(OTNOConfig): |
| n_modes: List[int] = [16, 16] |
| hidden_channels: int = 64 |
| data_channels: int = 9 |
| out_channels: int = 1 |
|
|
