File size: 11,366 Bytes
8b16bae | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 | from __future__ import annotations
from collections.abc import Sequence
import torch
from torch import nn
def _activation(name: str) -> nn.Module:
activations = {
"gelu": nn.GELU,
"relu": nn.ReLU,
"silu": nn.SiLU,
"swish": nn.SiLU,
"tanh": nn.Tanh,
}
try:
return activations[name.lower()]()
except KeyError as error:
raise ValueError(f"Unsupported activation: {name}") from error
class StandardMLP(nn.Module):
"""Pointwise MLP with state-dict names compatible with OneScience."""
def __init__(
self,
input_dim: int,
output_dim: int,
hidden_dims: Sequence[int],
activation: str = "gelu",
use_bias: bool = True,
) -> None:
super().__init__()
dimensions = [int(input_dim), *(int(value) for value in hidden_dims), int(output_dim)]
self.layers = nn.ModuleList(
nn.Linear(in_features, out_features, bias=use_bias)
for in_features, out_features in zip(dimensions[:-1], dimensions[1:])
)
self.activation = _activation(activation)
def forward(self, x: torch.Tensor) -> torch.Tensor:
for layer in self.layers[:-1]:
x = self.activation(layer(x))
return self.layers[-1](x)
class OneMlp(nn.Module):
"""Small compatibility wrapper used by released OneScience checkpoints."""
def __init__(self, **kwargs) -> None:
super().__init__()
self.mlp = StandardMLP(**kwargs)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.mlp(x)
class _PoolingReducer(nn.Module):
def __init__(self, in_dim: int, hidden_dim: int, out_dim: int) -> None:
super().__init__()
self.to_in = nn.Linear(in_dim, hidden_dim, bias=False)
self.out_ffn = nn.Sequential(
nn.LayerNorm(hidden_dim),
nn.Linear(hidden_dim, hidden_dim),
nn.GELU(),
nn.Linear(hidden_dim, out_dim),
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.to_in(x)
if x.ndim > 3:
x = x.mean(dim=tuple(range(2, x.ndim - 1)))
return self.out_ffn(x)
class _SwapGridAxes(nn.Module):
def forward(self, x: torch.Tensor) -> torch.Tensor:
return x.permute(0, 2, 1, 3)
class _FactAttnWeight(nn.Module):
def __init__(self, heads: int, dim_head: int, dropout: float) -> None:
super().__init__()
self.dim_head = int(dim_head)
self.heads = int(heads)
self.scale = self.dim_head**-0.5
self.softmax = nn.Softmax(dim=-1)
self.dropout = nn.Dropout(dropout)
self.to_q = nn.Linear(self.dim_head, self.dim_head, bias=False)
self.to_k = nn.Linear(self.dim_head, self.dim_head, bias=False)
def forward(self, x: torch.Tensor) -> torch.Tensor:
batch, length, _ = x.shape
x = x.reshape(batch, length, self.heads, self.dim_head)
x = x.permute(0, 2, 1, 3).contiguous()
query = self.to_q(x)
key = self.to_k(x)
return self.softmax(torch.matmul(query, key.transpose(-1, -2)) * self.scale)
class FactAttention2D(nn.Module):
"""Factorized axial attention for a structured two-dimensional grid."""
def __init__(
self,
dim: int,
heads: int,
dim_head: int,
dropout: float,
shapelist: Sequence[int],
) -> None:
super().__init__()
if len(shapelist) != 2:
raise ValueError("FactAttention2D requires shapelist=(height, width)")
if dim != heads * dim_head:
raise ValueError("dim must equal heads * dim_head")
self.dim_head = int(dim_head)
self.heads = int(heads)
self.H, self.W = (int(value) for value in shapelist)
self.attn_x = _FactAttnWeight(self.heads, self.dim_head, dropout)
self.attn_y = _FactAttnWeight(self.heads, self.dim_head, dropout)
self.to_v = nn.Linear(self.dim_head, self.dim_head, bias=False)
self.to_x = nn.Sequential(_PoolingReducer(dim, dim, dim))
self.to_y = nn.Sequential(_SwapGridAxes(), _PoolingReducer(dim, dim, dim))
self.to_out = nn.Sequential(nn.Linear(dim, dim), nn.Dropout(dropout))
def forward(self, x: torch.Tensor) -> torch.Tensor:
batch, length, channels = x.shape
if length != self.H * self.W:
raise ValueError(
f"Sequence length {length} does not match grid {self.H}x{self.W}"
)
grid = x.reshape(batch, self.H, self.W, channels).contiguous()
values = self.to_v(
grid.reshape(batch, self.H, self.W, self.heads, self.dim_head)
)
values = values.permute(0, 3, 1, 2, 4).contiguous()
result_x = torch.einsum(
"bhij,bhjmc->bhimc", self.attn_x(self.to_x(grid)), values
)
result_y = torch.einsum(
"bhlm,bhimc->bhilc", self.attn_y(self.to_y(grid)), result_x
)
result = result_y.permute(0, 2, 3, 1, 4).contiguous()
result = result.reshape(batch, length, channels)
return self.to_out(result)
class OneAttention(nn.Module):
"""Compatibility wrapper retaining the original checkpoint hierarchy."""
def __init__(self, **kwargs) -> None:
super().__init__()
self.attentioner = FactAttention2D(**kwargs)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.attentioner(x)
class FactformerBlock(nn.Module):
def __init__(
self,
num_heads: int,
hidden_dim: int,
dropout: float,
activation: str,
mlp_ratio: int,
spatial_shape: Sequence[int],
) -> None:
super().__init__()
self.ln_1 = nn.LayerNorm(hidden_dim)
self.Attn = OneAttention(
dim=hidden_dim,
heads=num_heads,
dim_head=hidden_dim // num_heads,
dropout=dropout,
shapelist=spatial_shape,
)
self.ln_2 = nn.LayerNorm(hidden_dim)
self.mlp = OneMlp(
input_dim=hidden_dim,
output_dim=hidden_dim,
hidden_dims=[hidden_dim * mlp_ratio],
activation=activation,
use_bias=True,
)
def forward(self, x: torch.Tensor) -> torch.Tensor:
x = self.Attn(self.ln_1(x)) + x
return self.mlp(self.ln_2(x)) + x
class OneTransformer(nn.Module):
"""Compatibility wrapper retaining the original checkpoint hierarchy."""
def __init__(self, **kwargs) -> None:
super().__init__()
self.transformer = FactformerBlock(**kwargs)
def forward(self, x: torch.Tensor) -> torch.Tensor:
return self.transformer(x)
class FactFormer2D(nn.Module):
"""FactFormer neural operator for structured 2D flow fields.
``pos`` has shape ``(B, H*W, 2)`` and ``fx`` has shape
``(B, H*W, t_in*out_dim)``. The output contains ``latent_steps`` future
fields concatenated along the final dimension.
"""
def __init__(
self,
in_dim: int,
out_dim: int,
spatial_shape: Sequence[int],
hidden_dim: int = 128,
depth: int = 4,
heads: int = 8,
mlp_ratio: int = 2,
dropout: float = 0.0,
activation: str = "gelu",
include_pos: bool = True,
space_dim: int = 2,
latent_multiplier: float = 2.0,
max_latent_steps: int = 4,
) -> None:
super().__init__()
self.in_dim = int(in_dim)
self.out_dim = int(out_dim)
self.spatial_shape = tuple(int(value) for value in spatial_shape)
self.hidden_dim = int(hidden_dim)
self.depth = int(depth)
self.heads = int(heads)
self.include_pos = bool(include_pos)
self.space_dim = int(space_dim)
self.latent_dim = int(self.hidden_dim * float(latent_multiplier))
self.max_latent_steps = int(max_latent_steps)
if len(self.spatial_shape) != 2:
raise ValueError("FactFormer2D requires a two-dimensional spatial shape")
if self.hidden_dim % self.heads:
raise ValueError("hidden_dim must be divisible by heads")
if self.max_latent_steps < 1:
raise ValueError("max_latent_steps must be positive")
input_dim = self.in_dim + (self.space_dim if self.include_pos else 0)
self.preprocess = OneMlp(
input_dim=input_dim,
output_dim=self.hidden_dim,
hidden_dims=[self.hidden_dim * 2],
activation=activation,
use_bias=True,
)
self.blocks = nn.ModuleList(
OneTransformer(
num_heads=self.heads,
hidden_dim=self.hidden_dim,
dropout=float(dropout),
activation=activation,
mlp_ratio=int(mlp_ratio),
spatial_shape=self.spatial_shape,
)
for _ in range(self.depth)
)
self.expand_latent = nn.Linear(self.hidden_dim, self.latent_dim, bias=False)
self.latent_time_embedding = nn.Parameter(
torch.randn(1, self.max_latent_steps, 1, self.latent_dim) * 0.02
)
self.propagator = OneMlp(
input_dim=self.latent_dim,
output_dim=self.latent_dim,
hidden_dims=[self.hidden_dim],
activation=activation,
use_bias=True,
)
self.to_out = OneMlp(
input_dim=self.latent_dim,
output_dim=self.out_dim,
hidden_dims=[self.hidden_dim],
activation=activation,
use_bias=True,
)
self.initialize_weights()
def initialize_weights(self) -> None:
for module in self.modules():
if isinstance(module, nn.Linear):
nn.init.trunc_normal_(module.weight, std=0.02)
if module.bias is not None:
nn.init.zeros_(module.bias)
elif isinstance(module, nn.LayerNorm):
nn.init.ones_(module.weight)
nn.init.zeros_(module.bias)
def forward(
self,
pos: torch.Tensor,
fx: torch.Tensor,
latent_steps: int = 1,
) -> torch.Tensor:
if fx.ndim != 3:
raise ValueError(f"fx must have shape (B, N, C), got {tuple(fx.shape)}")
if pos.ndim == 2:
pos = pos.unsqueeze(0)
if pos.shape[0] == 1 and fx.shape[0] > 1:
pos = pos.expand(fx.shape[0], -1, -1)
if pos.shape[:2] != fx.shape[:2]:
raise ValueError(f"Incompatible pos and fx shapes: {pos.shape}, {fx.shape}")
if not 1 <= int(latent_steps) <= self.max_latent_steps:
raise ValueError(
f"latent_steps must be in [1, {self.max_latent_steps}], got {latent_steps}"
)
hidden = torch.cat((pos, fx), dim=-1) if self.include_pos else fx
hidden = self.preprocess(hidden)
for block in self.blocks:
hidden = block(hidden)
latent = self.expand_latent(hidden)
outputs = []
for step in range(int(latent_steps)):
latent = latent + self.latent_time_embedding[:, step]
latent = self.propagator(latent) + latent
outputs.append(self.to_out(latent))
return torch.cat(outputs, dim=-1)
__all__ = ["FactFormer2D"]
|