File size: 27,311 Bytes
04fe5fd | 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 465 466 467 468 469 470 471 472 473 474 475 476 477 478 479 480 481 482 483 484 485 486 487 488 489 490 491 492 493 494 495 496 497 498 499 500 501 502 503 504 505 506 507 508 509 510 511 512 513 514 515 516 517 518 519 520 521 522 523 524 525 526 527 528 529 530 531 532 533 534 535 536 537 538 539 540 541 542 543 544 545 546 547 548 549 550 551 552 553 554 | """Portable OneForecast model and official checkpoint compatibility helpers.
The parameter hierarchy mirrors the official model. Graph operations use
PyTorch index tensors instead of CUDA-only CuGraph kernels, making the model
usable on CPU, CUDA, and DCU PyTorch builds.
"""
from __future__ import annotations
from dataclasses import dataclass
from pathlib import Path
import sys
import types
from typing import Any, NamedTuple
import numpy as np
import torch
from torch import Tensor, nn
from torch.nn import functional as F
class GraphData:
"""Minimal homogeneous or bipartite graph used by the portable kernels."""
def __init__(self, src: Tensor, dst: Tensor, num_src: int, num_dst: int) -> None:
self.src = src.to(torch.long)
self.dst = dst.to(torch.long)
self.num_src = num_src
self.num_dst = num_dst
def to(self, device: torch.device | str) -> "GraphData":
self.src = self.src.to(device)
self.dst = self.dst.to(device)
return self
def _aggregate(values: Tensor, dst: Tensor, num_dst: int, reduction: str) -> Tensor:
output = values.new_zeros((num_dst,) + values.shape[1:])
index = dst.view((-1,) + (1,) * (values.ndim - 1)).expand_as(values)
output.scatter_add_(0, index, values)
if reduction == "mean":
counts = values.new_zeros(num_dst)
counts.scatter_add_(0, dst, torch.ones_like(dst, dtype=values.dtype))
output = output / counts.clamp_min(1).view((-1,) + (1,) * (values.ndim - 1))
elif reduction != "sum":
raise ValueError(f"Unsupported aggregation: {reduction}")
return output
def _edge_softmax(logits: Tensor, dst: Tensor, num_dst: int) -> Tensor:
index = dst[:, None].expand_as(logits)
maxima = logits.new_full((num_dst, logits.shape[1]), -torch.inf)
maxima.scatter_reduce_(0, index, logits, reduce="amax", include_self=True)
exp = torch.exp(logits - maxima[dst])
denominator = logits.new_zeros((num_dst, logits.shape[1]))
denominator.scatter_add_(0, index, exp)
return exp / denominator[dst].clamp_min(torch.finfo(exp.dtype).tiny)
class MeshGraphMLP(nn.Module):
def __init__(self, input_dim: int, output_dim: int = 512, hidden_dim: int = 512,
hidden_layers: int | None = 1, activation_fn: nn.Module | None = None,
norm_type: str | None = "LayerNorm", recompute_activation: bool = False) -> None:
super().__init__()
del recompute_activation
activation_fn = activation_fn or nn.SiLU()
if hidden_layers is None:
self.model = nn.Identity()
return
layers: list[nn.Module] = [nn.Linear(input_dim, hidden_dim), activation_fn]
for _ in range(hidden_layers - 1):
layers.extend([nn.Linear(hidden_dim, hidden_dim), nn.SiLU()])
layers.append(nn.Linear(hidden_dim, output_dim))
if norm_type is not None:
if norm_type != "LayerNorm":
raise ValueError("The portable model supports LayerNorm only")
layers.append(nn.LayerNorm(output_dim))
self.model = nn.Sequential(*layers)
def forward(self, x: Tensor) -> Tensor:
return self.model(x)
class MeshGraphEdgeMLPSum(nn.Module):
"""Concat-trick edge MLP with the official parameter names and initialization."""
def __init__(self, efeat_dim: int, src_dim: int, dst_dim: int,
output_dim: int = 512, hidden_dim: int = 512,
hidden_layers: int = 1, activation_fn: nn.Module | None = None,
norm_type: str | None = "LayerNorm", recompute_activation: bool = False) -> None:
super().__init__()
del recompute_activation
activation_fn = activation_fn or nn.SiLU()
initial = nn.Linear(efeat_dim + src_dim + dst_dim, hidden_dim)
weights = torch.split(initial.weight, [efeat_dim, src_dim, dst_dim], dim=1)
self.lin_efeat = nn.Parameter(weights[0])
self.lin_src = nn.Parameter(weights[1])
self.lin_dst = nn.Parameter(weights[2])
self.bias = initial.bias
layers: list[nn.Module] = [activation_fn]
for _ in range(hidden_layers - 1):
layers.extend([nn.Linear(hidden_dim, hidden_dim), nn.SiLU()])
layers.append(nn.Linear(hidden_dim, output_dim))
if norm_type is not None:
if norm_type != "LayerNorm":
raise ValueError("The portable model supports LayerNorm only")
layers.append(nn.LayerNorm(output_dim))
self.model = nn.Sequential(*layers)
def forward(self, efeat: Tensor, nfeat: Tensor | tuple[Tensor, Tensor],
graph: GraphData) -> Tensor:
src_feat, dst_feat = (nfeat, nfeat) if isinstance(nfeat, Tensor) else nfeat
hidden = F.linear(efeat, self.lin_efeat)
hidden = hidden + F.linear(src_feat[graph.src], self.lin_src)
hidden = hidden + F.linear(dst_feat[graph.dst], self.lin_dst, self.bias)
return self.model(hidden)
class OneForecastEncoderEmbedder(nn.Module):
def __init__(self, input_dim_grid_nodes: int = 69, input_dim_mesh_nodes: int = 3,
input_dim_edges: int = 4, output_dim: int = 512,
hidden_dim: int = 512, hidden_layers: int = 1) -> None:
super().__init__()
kwargs = dict(output_dim=output_dim, hidden_dim=hidden_dim, hidden_layers=hidden_layers)
self.grid_node_mlp = MeshGraphMLP(input_dim_grid_nodes, **kwargs)
self.mesh_node_mlp = MeshGraphMLP(input_dim_mesh_nodes, **kwargs)
self.mesh_edge_mlp = MeshGraphMLP(input_dim_edges, **kwargs)
self.grid2mesh_edge_mlp = MeshGraphMLP(input_dim_edges, **kwargs)
def forward(self, grid: Tensor, mesh: Tensor, g2m: Tensor,
mesh_edges: Tensor) -> tuple[Tensor, Tensor, Tensor, Tensor]:
return (self.grid_node_mlp(grid), self.mesh_node_mlp(mesh),
self.grid2mesh_edge_mlp(g2m), self.mesh_edge_mlp(mesh_edges))
class OneForecastDecoderEmbedder(nn.Module):
def __init__(self, input_dim_edges: int = 4, output_dim: int = 512,
hidden_dim: int = 512, hidden_layers: int = 1) -> None:
super().__init__()
self.mesh2grid_edge_mlp = MeshGraphMLP(
input_dim_edges, output_dim, hidden_dim, hidden_layers)
def forward(self, edges: Tensor) -> Tensor:
return self.mesh2grid_edge_mlp(edges)
class MeshGraphEncoder(nn.Module):
def __init__(self, hidden_dim: int = 512, hidden_layers: int = 1,
aggregation: str = "sum") -> None:
super().__init__()
self.aggregation = aggregation
self.edge_mlp = MeshGraphEdgeMLPSum(hidden_dim, hidden_dim, hidden_dim,
hidden_dim, hidden_dim, hidden_layers)
self.src_node_mlp = MeshGraphMLP(hidden_dim, hidden_dim, hidden_dim, hidden_layers)
self.dst_node_mlp = MeshGraphMLP(hidden_dim * 2, hidden_dim, hidden_dim, hidden_layers)
def forward(self, edges: Tensor, grid: Tensor, mesh: Tensor,
graph: GraphData) -> tuple[Tensor, Tensor]:
edges = self.edge_mlp(edges, (grid, mesh), graph)
aggregated = _aggregate(edges, graph.dst, graph.num_dst, self.aggregation)
return grid + self.src_node_mlp(grid), mesh + self.dst_node_mlp(torch.cat((aggregated, mesh), -1))
class MeshGraphDecoder(nn.Module):
def __init__(self, hidden_dim: int = 512, hidden_layers: int = 1,
aggregation: str = "sum") -> None:
super().__init__()
self.aggregation = aggregation
self.edge_mlp = MeshGraphEdgeMLPSum(hidden_dim, hidden_dim, hidden_dim,
hidden_dim, hidden_dim, hidden_layers)
self.node_mlp = MeshGraphMLP(hidden_dim * 2, hidden_dim, hidden_dim, hidden_layers)
def forward(self, edges: Tensor, grid: Tensor, mesh: Tensor, graph: GraphData) -> Tensor:
edges = self.edge_mlp(edges, (mesh, grid), graph)
aggregated = _aggregate(edges, graph.dst, graph.num_dst, self.aggregation)
return grid + self.node_mlp(torch.cat((aggregated, grid), -1))
class MeshEdgeBlockMultiHeadGated(nn.Module):
def __init__(self, hidden_dim: int = 512, hidden_layers: int = 1,
num_heads: int = 4) -> None:
super().__init__()
self.num_heads = num_heads
self.edge_mlp = MeshGraphEdgeMLPSum(hidden_dim, hidden_dim, hidden_dim,
hidden_dim, hidden_dim, hidden_layers)
gating_hidden = max(16, hidden_dim // 8)
self.gate_net = nn.Sequential(nn.Linear(hidden_dim * 3, gating_hidden), nn.SiLU(),
nn.Linear(gating_hidden, 3 * num_heads), nn.Sigmoid())
def forward(self, edges: Tensor, nodes: Tensor, graph: GraphData) -> tuple[Tensor, Tensor]:
raw = torch.cat((edges, nodes[graph.src], nodes[graph.dst]), -1)
gates = self.gate_net(raw).view(-1, self.num_heads, 3).mean(1)
updated = self.edge_mlp(edges, nodes, graph)
return edges + updated * gates.mean(-1, keepdim=True), nodes
class MeshNodeBlockMultiHeadAttn(nn.Module):
def __init__(self, hidden_dim: int = 512, hidden_layers: int = 1,
aggregation: str = "sum", num_heads: int = 4) -> None:
super().__init__()
self.num_heads = num_heads
self.aggregation = aggregation
self.node_mlp = MeshGraphMLP(hidden_dim * (num_heads + 1), hidden_dim,
hidden_dim, hidden_layers)
attention_hidden = max(16, hidden_dim // 8)
self.attn_net = nn.Sequential(nn.Linear(hidden_dim, attention_hidden), nn.SiLU(),
nn.Linear(attention_hidden, num_heads))
def forward(self, edges: Tensor, nodes: Tensor, graph: GraphData) -> tuple[Tensor, Tensor]:
scores = _edge_softmax(self.attn_net(edges), graph.dst, graph.num_dst)
messages = edges[:, None, :].expand(-1, self.num_heads, -1) * scores[:, :, None]
aggregated = _aggregate(messages, graph.dst, graph.num_dst, self.aggregation).flatten(1)
return edges, nodes + self.node_mlp(torch.cat((aggregated, nodes), -1))
class OneForecastProcessor(nn.Module):
def __init__(self, processor_layers: int, hidden_dim: int = 512,
hidden_layers: int = 1, aggregation: str = "sum",
num_heads_edge: int = 4, num_heads_node: int = 4) -> None:
super().__init__()
layers: list[nn.Module] = []
for _ in range(processor_layers):
layers.append(MeshEdgeBlockMultiHeadGated(hidden_dim, hidden_layers, num_heads_edge))
layers.append(MeshNodeBlockMultiHeadAttn(hidden_dim, hidden_layers, aggregation, num_heads_node))
self.processor_layers = nn.ModuleList(layers)
def forward(self, edges: Tensor, nodes: Tensor, graph: GraphData) -> tuple[Tensor, Tensor]:
for layer in self.processor_layers:
edges, nodes = layer(edges, nodes, graph)
return edges, nodes
class TriangularMesh(NamedTuple):
vertices: np.ndarray
faces: np.ndarray
def _icosahedron() -> TriangularMesh:
from scipy.spatial.transform import Rotation
phi = (1 + np.sqrt(5)) / 2
vertices = []
for first in (1.0, -1.0):
for second in (phi, -phi):
vertices.extend(((first, second, 0.0), (0.0, first, second), (second, 0.0, first)))
vertices = np.asarray(vertices, dtype=np.float32) / np.linalg.norm([1.0, phi])
faces = np.asarray(((0,1,2),(0,6,1),(8,0,2),(8,4,0),(3,8,2),(3,2,7),(7,2,1),
(0,4,6),(4,11,6),(6,11,5),(1,5,7),(4,10,11),(4,8,10),(10,8,3),
(10,3,9),(11,10,9),(11,9,5),(5,9,7),(9,3,7),(1,6,5)), dtype=np.int32)
angle = (np.pi - 2 * np.arcsin(phi / np.sqrt(3))) / 2
vertices = vertices @ Rotation.from_euler("y", angle).as_matrix()
return TriangularMesh(vertices.astype(np.float32), faces)
def _split_mesh(mesh: TriangularMesh) -> TriangularMesh:
vertices = list(mesh.vertices)
children: dict[tuple[int, int], int] = {}
faces = []
for a, b, c in mesh.faces:
mids = []
for pair in ((a, b), (b, c), (c, a)):
key = tuple(sorted(map(int, pair)))
if key not in children:
position = mesh.vertices[list(pair)].mean(0)
position /= np.linalg.norm(position)
children[key] = len(vertices)
vertices.append(position)
mids.append(children[key])
ab, bc, ca = mids
faces.extend(((a, ab, ca), (ab, b, bc), (ca, bc, c), (ab, bc, ca)))
return TriangularMesh(np.asarray(vertices, dtype=np.float32), np.asarray(faces, dtype=np.int32))
def _mesh_hierarchy(level: int) -> list[TriangularMesh]:
meshes = [_icosahedron()]
for _ in range(level):
meshes.append(_split_mesh(meshes[-1]))
return meshes
def _faces_to_edges(faces: np.ndarray) -> tuple[np.ndarray, np.ndarray]:
return (np.concatenate((faces[:, 0], faces[:, 1], faces[:, 2])),
np.concatenate((faces[:, 1], faces[:, 2], faces[:, 0])))
def _latlon_to_xyz(latlon: Tensor) -> Tensor:
values = torch.deg2rad(latlon)
lat, lon = values[:, 0], values[:, 1]
return torch.stack((torch.cos(lat) * torch.cos(lon), torch.cos(lat) * torch.sin(lon), torch.sin(lat)), 1)
def _node_features(xyz: Tensor) -> Tensor:
# The official implementation applies trigonometric functions to the
# degree-valued xyz2latlon output; retain that behavior for parity.
lat = torch.rad2deg(torch.asin(xyz[:, 2]))
lon = torch.rad2deg(torch.atan2(xyz[:, 1], xyz[:, 0]))
return torch.stack((torch.cos(lat), torch.sin(lon), torch.cos(lon)), -1)
def _edge_features(src_pos: Tensor, dst_pos: Tensor, src: Tensor, dst: Tensor) -> Tensor:
source, target = src_pos[src], dst_pos[dst]
lat = torch.asin(target[:, 2])
lon = torch.atan2(target[:, 1], target[:, 0])
cos_lon, sin_lon = torch.cos(-lon), torch.sin(-lon)
source = torch.stack((cos_lon * source[:, 0] - sin_lon * source[:, 1],
sin_lon * source[:, 0] + cos_lon * source[:, 1], source[:, 2]), -1)
target = torch.stack((cos_lon * target[:, 0] - sin_lon * target[:, 1],
sin_lon * target[:, 0] + cos_lon * target[:, 1], target[:, 2]), -1)
cos_lat, sin_lat = torch.cos(lat), torch.sin(lat)
source = torch.stack((cos_lat * source[:, 0] + sin_lat * source[:, 2], source[:, 1],
-sin_lat * source[:, 0] + cos_lat * source[:, 2]), -1)
target = torch.stack((cos_lat * target[:, 0] + sin_lat * target[:, 2], target[:, 1],
-sin_lat * target[:, 0] + cos_lat * target[:, 2]), -1)
displacement = source - target
norm = torch.linalg.norm(displacement, dim=-1, keepdim=True)
maximum = norm.max()
return torch.cat((displacement / maximum, norm / maximum), -1)
def _local_refine(mesh: TriangularMesh, lat_min: float, lat_max: float,
lon_min: float, lon_max: float) -> TriangularMesh:
centroids = mesh.vertices[mesh.faces].mean(axis=1)
# Match the official xyz2latlon call, which assumes radius=1 for centroids.
latitudes = np.rad2deg(np.arcsin(centroids[:, 2]))
longitudes = np.rad2deg(np.arctan2(centroids[:, 1], centroids[:, 0]))
selected = ((latitudes >= lat_min) & (latitudes <= lat_max)
& (longitudes >= lon_min) & (longitudes <= lon_max))
refined = _split_mesh(TriangularMesh(mesh.vertices, mesh.faces[selected]))
combined_vertices = np.concatenate((refined.vertices, mesh.vertices), axis=0)
combined_faces = np.concatenate((refined.faces, mesh.faces[~selected] + len(refined.vertices)), axis=0)
rounded = np.round(combined_vertices, decimals=6)
unique: dict[tuple[float, float, float], int] = {}
remap = np.empty(len(rounded), dtype=np.int64)
vertices = []
for index, coordinates in enumerate(rounded):
key = tuple(coordinates.tolist())
if key not in unique:
unique[key] = len(vertices)
vertices.append(combined_vertices[index])
remap[index] = unique[key]
return TriangularMesh(np.asarray(vertices, dtype=np.float32), remap[combined_faces].astype(np.int32))
def _build_graphs(height: int, width: int, mesh_level: int) -> tuple[GraphData, GraphData, GraphData, Tensor, Tensor, Tensor, Tensor]:
from scipy.spatial import cKDTree
latitudes = torch.linspace(-90, 90, height + 1)[:-1]
longitudes = torch.linspace(-180, 180, width + 1)[1:]
latlon = torch.stack(torch.meshgrid(latitudes, longitudes, indexing="ij"), -1).reshape(-1, 2)
grid_xyz = _latlon_to_xyz(latlon)
hierarchy = _mesh_hierarchy(mesh_level)
finest = hierarchy[-1]
refined = _local_refine(finest, 0.0, 30.0, 105.0, 160.0)
refined = _local_refine(refined, 10.0, 30.0, -95.0, -35.0)
mesh_vertices = refined.vertices
mesh_faces = np.concatenate([mesh.faces for mesh in hierarchy] + [refined.faces], axis=0)
mesh_src, mesh_dst = _faces_to_edges(mesh_faces)
mesh_src = np.concatenate((mesh_src, mesh_dst))
mesh_dst = np.concatenate((mesh_dst, mesh_src[:len(mesh_dst)]))
pairs = np.unique(np.stack((mesh_src, mesh_dst), 1), axis=0)
mesh_src_t = torch.from_numpy(pairs[:, 0])
mesh_dst_t = torch.from_numpy(pairs[:, 1])
mesh_xyz = torch.from_numpy(mesh_vertices)
mesh_graph = GraphData(mesh_src_t, mesh_dst_t, len(mesh_vertices), len(mesh_vertices))
finest_src, finest_dst = _faces_to_edges(finest.faces)
max_edge = np.linalg.norm(finest.vertices[finest_src] - finest.vertices[finest_dst], axis=1).max()
distances, neighbors = cKDTree(mesh_vertices).query(grid_xyz.numpy(), k=4)
valid = distances <= 0.6 * max_edge
g2m_src, neighbor_slot = np.nonzero(valid)
g2m_dst = neighbors[g2m_src, neighbor_slot]
g2m_graph = GraphData(torch.from_numpy(g2m_src), torch.from_numpy(g2m_dst), len(grid_xyz), len(mesh_vertices))
centroids = mesh_vertices[mesh_faces].mean(axis=1)
face_indices = cKDTree(centroids).query(grid_xyz.numpy(), k=1)[1]
m2g_src = mesh_faces[face_indices].reshape(-1)
m2g_dst = np.repeat(np.arange(len(grid_xyz)), 3)
m2g_graph = GraphData(torch.from_numpy(m2g_src), torch.from_numpy(m2g_dst), len(mesh_vertices), len(grid_xyz))
mesh_nodes = _node_features(mesh_xyz)
mesh_edges = _edge_features(mesh_xyz, mesh_xyz, mesh_graph.src, mesh_graph.dst)
g2m_edges = _edge_features(grid_xyz, mesh_xyz, g2m_graph.src, g2m_graph.dst)
m2g_edges = _edge_features(mesh_xyz, grid_xyz, m2g_graph.src, m2g_graph.dst)
return mesh_graph, g2m_graph, m2g_graph, mesh_nodes, mesh_edges, g2m_edges, m2g_edges
class OneForecast(nn.Module):
"""Official OneForecast message-passing architecture with portable graph kernels."""
def __init__(self, input_res: tuple[int, int] = (120, 240), input_dim_grid_nodes: int = 69,
output_dim_grid_nodes: int = 69, mesh_level: int = 5,
processor_layers: int = 16, hidden_layers: int = 1,
hidden_dim: int = 512, aggregation: str = "sum",
num_heads_edge: int = 4, num_heads_node: int = 4,
build_graph: bool = True) -> None:
super().__init__()
if processor_layers <= 2:
raise ValueError("Expected at least 3 processor layers")
self.register_buffer("device_buffer", torch.empty(0))
self.input_res = tuple(input_res)
self.input_dim_grid_nodes = input_dim_grid_nodes
self.output_dim_grid_nodes = output_dim_grid_nodes
self.mesh_level = mesh_level
self.encoder_embedder = OneForecastEncoderEmbedder(
input_dim_grid_nodes, 3, 4, hidden_dim, hidden_dim, hidden_layers)
self.decoder_embedder = OneForecastDecoderEmbedder(4, hidden_dim, hidden_dim, hidden_layers)
self.encoder = MeshGraphEncoder(hidden_dim, hidden_layers, aggregation)
self.processor_encoder = OneForecastProcessor(
1, hidden_dim, hidden_layers, aggregation, num_heads_edge, num_heads_node)
self.processor = OneForecastProcessor(
processor_layers - 2, hidden_dim, hidden_layers, aggregation, num_heads_edge, num_heads_node)
self.processor_decoder = OneForecastProcessor(
1, hidden_dim, hidden_layers, aggregation, num_heads_edge, num_heads_node)
self.decoder = MeshGraphDecoder(hidden_dim, hidden_layers, aggregation)
self.finale = MeshGraphMLP(hidden_dim, output_dim_grid_nodes, hidden_dim, hidden_layers, norm_type=None)
self._graph_ready = False
if build_graph:
self.build_graph()
def build_graph(self) -> None:
values = _build_graphs(*self.input_res, self.mesh_level)
self.mesh_graph, self.g2m_graph, self.m2g_graph = values[:3]
for name, value in zip(("mesh_ndata", "mesh_edata", "g2m_edata", "m2g_edata"), values[3:]):
self.register_buffer(name, value, persistent=False)
self._graph_ready = True
def forward(self, grid_nfeat: Tensor) -> Tensor:
if not self._graph_ready:
raise RuntimeError("Graph construction was disabled for this model instance")
if grid_nfeat.shape != (1, self.input_dim_grid_nodes, *self.input_res):
raise ValueError(f"Expected input shape (1, {self.input_dim_grid_nodes}, {self.input_res[0]}, {self.input_res[1]}), got {tuple(grid_nfeat.shape)}")
grid = grid_nfeat[0].reshape(self.input_dim_grid_nodes, -1).T
grid, mesh, g2m, mesh_edges = self.encoder_embedder(
grid, self.mesh_ndata, self.g2m_edata, self.mesh_edata)
grid, mesh = self.encoder(g2m, grid, mesh, self.g2m_graph)
mesh_edges, mesh = self.processor_encoder(mesh_edges, mesh, self.mesh_graph)
mesh_edges, mesh = self.processor(mesh_edges, mesh, self.mesh_graph)
_, mesh = self.processor_decoder(mesh_edges, mesh, self.mesh_graph)
grid = self.decoder(self.decoder_embedder(self.m2g_edata), grid, mesh, self.m2g_graph)
output = self.finale(grid).T.reshape(self.output_dim_grid_nodes, *self.input_res)
return output.unsqueeze(0)
def to(self, *args: Any, **kwargs: Any) -> "OneForecast":
super().to(*args, **kwargs)
if self._graph_ready:
device = self.device_buffer.device
self.mesh_graph.to(device)
self.g2m_graph.to(device)
self.m2g_graph.to(device)
return self
@dataclass(frozen=True)
class CheckpointReport:
checkpoint_path: str
checkpoint_keys: int
model_keys: int
missing_keys: tuple[str, ...]
unexpected_keys: tuple[str, ...]
shape_mismatches: tuple[str, ...]
@property
def compatible(self) -> bool:
return not (self.missing_keys or self.unexpected_keys or self.shape_mismatches)
def _install_scalarfloat_safe_global() -> type[float]:
"""Allow the known ruamel ScalarFloat metadata type without importing ruamel."""
module_name = "ruamel.yaml.scalarfloat"
module = sys.modules.get(module_name)
if module is not None and hasattr(module, "ScalarFloat"):
scalar_float = module.ScalarFloat
else:
ruamel = sys.modules.setdefault("ruamel", types.ModuleType("ruamel"))
yaml_module = sys.modules.setdefault("ruamel.yaml", types.ModuleType("ruamel.yaml"))
module = types.ModuleType(module_name)
scalar_float = type("ScalarFloat", (float,), {})
scalar_float.__module__ = module_name
module.ScalarFloat = scalar_float
yaml_module.scalarfloat = module
ruamel.yaml = yaml_module
sys.modules[module_name] = module
torch.serialization.add_safe_globals([scalar_float])
return scalar_float
def read_official_checkpoint(path: str | Path) -> tuple[dict[str, Tensor], dict[str, Any]]:
path = Path(path).expanduser().resolve()
_install_scalarfloat_safe_global()
checkpoint = torch.load(path, map_location="cpu", weights_only=True, mmap=True)
if not isinstance(checkpoint, dict) or "model_state" not in checkpoint:
raise ValueError(f"{path} does not contain an official model_state")
state = checkpoint["model_state"]
if not isinstance(state, dict):
raise TypeError("checkpoint model_state must be a mapping")
cleaned = {key.removeprefix("module."): value for key, value in state.items()}
metadata = {key: value for key, value in checkpoint.items() if key not in {"model_state", "optimizer_state_dict"}}
return cleaned, metadata
def _compare_checkpoint_state(model: nn.Module, checkpoint_state: dict[str, Tensor],
path: str | Path) -> CheckpointReport:
model_state = model.state_dict()
missing = tuple(sorted(set(model_state) - set(checkpoint_state)))
unexpected = tuple(sorted(set(checkpoint_state) - set(model_state)))
mismatches = tuple(sorted(
f"{key}: checkpoint={tuple(checkpoint_state[key].shape)} model={tuple(model_state[key].shape)}"
for key in set(model_state) & set(checkpoint_state)
if model_state[key].shape != checkpoint_state[key].shape
))
return CheckpointReport(str(Path(path).expanduser().resolve()), len(checkpoint_state),
len(model_state), missing, unexpected, mismatches)
def check_checkpoint_compatibility(model: nn.Module, path: str | Path) -> CheckpointReport:
checkpoint_state, _ = read_official_checkpoint(path)
return _compare_checkpoint_state(model, checkpoint_state, path)
def load_official_checkpoint(model: nn.Module, path: str | Path, strict: bool = True) -> CheckpointReport:
state, _ = read_official_checkpoint(path)
report = _compare_checkpoint_state(model, state, path)
if strict and not report.compatible:
raise RuntimeError(f"Official checkpoint is incompatible: {report}")
compatible = {key: value for key, value in state.items()
if key in model.state_dict() and value.shape == model.state_dict()[key].shape}
model.load_state_dict(compatible, strict=strict)
return report
def build_model(config: dict[str, Any], build_graph: bool = True) -> OneForecast:
settings = config["model"]
model = OneForecast(
input_res=(settings["grid_height"], settings["grid_width"]),
input_dim_grid_nodes=settings["input_channels"],
output_dim_grid_nodes=settings["output_channels"],
mesh_level=settings.get("mesh_level", 5),
processor_layers=settings.get("processor_layers", 16),
hidden_layers=settings.get("hidden_layers", 1),
hidden_dim=settings.get("hidden_dim", 512),
num_heads_edge=settings.get("num_heads_edge", 4),
num_heads_node=settings.get("num_heads_node", 4),
build_graph=build_graph,
)
initialization = settings.get("weight_init", "scratch")
if initialization == "official":
load_official_checkpoint(model, settings["checkpoint_path"])
elif initialization != "scratch":
raise ValueError("model.weight_init must be 'scratch' or 'official'")
return model
__all__ = ["CheckpointReport", "OneForecast", "build_model", "check_checkpoint_compatibility",
"load_official_checkpoint", "read_official_checkpoint"]
|