Upload 4 files
Browse files- unified_graph_weather_dataloader.py +415 -0
- unified_graph_weather_model.py +366 -0
- unified_graph_weather_model_tp.py +272 -0
- unified_graph_weather_model_uv10.py +229 -0
unified_graph_weather_dataloader.py
ADDED
|
@@ -0,0 +1,415 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# graph_weather_dataloader.py
|
| 2 |
+
from __future__ import annotations
|
| 3 |
+
from pathlib import Path
|
| 4 |
+
from typing import Optional, Tuple, List, Dict
|
| 5 |
+
|
| 6 |
+
import re
|
| 7 |
+
import json
|
| 8 |
+
import math
|
| 9 |
+
import numpy as np
|
| 10 |
+
import torch
|
| 11 |
+
from torch.utils.data import Dataset, DataLoader
|
| 12 |
+
import pytorch_lightning as pl
|
| 13 |
+
|
| 14 |
+
|
| 15 |
+
# ==========================
|
| 16 |
+
# Utils
|
| 17 |
+
# ==========================
|
| 18 |
+
def _parse_ym_from_name(p: Path):
|
| 19 |
+
m = re.search(r"(\d{4})-(\d{2})\.pt$", p.name)
|
| 20 |
+
if not m:
|
| 21 |
+
return None, None
|
| 22 |
+
return int(m.group(1)), int(m.group(2))
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
def _sincos_pos2d_nodes(lat_deg: np.ndarray, lon_deg: np.ndarray) -> np.ndarray:
|
| 26 |
+
"""
|
| 27 |
+
По вузлах графа -> [4, N]: sinφ, cosφ, sin(λ·cosφ̄), cos(λ·cosφ̄).
|
| 28 |
+
φ̄ — середня широта домену (стабільно, як у твоєму build-скрипті).
|
| 29 |
+
"""
|
| 30 |
+
lat = np.deg2rad(lat_deg.astype(np.float32, copy=False))
|
| 31 |
+
lon = np.deg2rad(lon_deg.astype(np.float32, copy=False))
|
| 32 |
+
lat0 = float(lat.mean())
|
| 33 |
+
lon_adj = lon * np.cos(lat0)
|
| 34 |
+
pos = np.stack([np.sin(lat), np.cos(lat), np.sin(lon_adj), np.cos(lon_adj)], axis=0)
|
| 35 |
+
return pos.astype(np.float32, copy=False)
|
| 36 |
+
|
| 37 |
+
|
| 38 |
+
def _load_graph_static(npz_path: Path):
|
| 39 |
+
"""
|
| 40 |
+
Зчитує статичний мультискейл граф:
|
| 41 |
+
обов'язково: coord_lon[N], coord_lat[N], static_geo_mesh[N,4]
|
| 42 |
+
опційно: edge_index_horiz_base[2,Eh] або edge_index_horiz[2, E_total] + horiz_edge_ptr[L+1]
|
| 43 |
+
(fallback на edges_h_src/edges_h_dst або edges_horiz)
|
| 44 |
+
"""
|
| 45 |
+
d = np.load(npz_path, allow_pickle=True)
|
| 46 |
+
|
| 47 |
+
# --- coords
|
| 48 |
+
lon = d["coord_lon"].astype(np.float32, copy=False)
|
| 49 |
+
lat = d["coord_lat"].astype(np.float32, copy=False)
|
| 50 |
+
N = int(lon.shape[0])
|
| 51 |
+
|
| 52 |
+
# --- static geo (4 канали, вже нормалізовані по зонах)
|
| 53 |
+
if "static_geo_mesh" not in d:
|
| 54 |
+
raise KeyError(f"{npz_path} must contain 'static_geo_mesh'")
|
| 55 |
+
static_geo = d["static_geo_mesh"].astype(np.float32, copy=False) # [N,4]
|
| 56 |
+
if static_geo.shape[0] != N:
|
| 57 |
+
raise ValueError("static_geo_mesh first dim must match coord arrays")
|
| 58 |
+
|
| 59 |
+
# --- pos2d: беремо з файла, якщо є; інакше — будуємо
|
| 60 |
+
if "static_pos2d_mesh" in d:
|
| 61 |
+
pos2d = d["static_pos2d_mesh"].astype(np.float32, copy=False).T # [4,N]
|
| 62 |
+
else:
|
| 63 |
+
pos2d = _sincos_pos2d_nodes(lat, lon) # [4,N]
|
| 64 |
+
|
| 65 |
+
# --- levels/meta (не обов'язково, але корисно)
|
| 66 |
+
levels = [str(x) for x in (d["levels"].tolist() if "levels" in d else [])]
|
| 67 |
+
level_offsets = d["level_offsets"].astype(np.int32, copy=False) if "level_offsets" in d else None
|
| 68 |
+
mesh_hash = None
|
| 69 |
+
if "mesh_hash" in d:
|
| 70 |
+
mh = d["mesh_hash"].item()
|
| 71 |
+
mesh_hash = (mh.decode() if isinstance(mh, (bytes, np.bytes_)) else str(mh))
|
| 72 |
+
|
| 73 |
+
# --- edge_index_base: кілька шляхів
|
| 74 |
+
edge_index = None
|
| 75 |
+
|
| 76 |
+
# 1) ідеально — вже готове поле
|
| 77 |
+
if "edge_index_horiz_base" in d:
|
| 78 |
+
ei = d["edge_index_horiz_base"].astype(np.int64, copy=False)
|
| 79 |
+
edge_index = torch.from_numpy(ei)
|
| 80 |
+
|
| 81 |
+
# 2) зібране по рівнях: edge_index_horiz + horiz_edge_ptr → беремо рівень 0
|
| 82 |
+
elif "edge_index_horiz" in d and "horiz_edge_ptr" in d:
|
| 83 |
+
e_all = d["edge_index_horiz"] # [2, E_total] або [E_total, 2]
|
| 84 |
+
if e_all.ndim == 2 and e_all.shape[0] == 2:
|
| 85 |
+
e_all = e_all
|
| 86 |
+
elif e_all.ndim == 2 and e_all.shape[1] == 2:
|
| 87 |
+
e_all = e_all.T
|
| 88 |
+
else:
|
| 89 |
+
raise ValueError("edge_index_horiz must be [2, E] or [E, 2]")
|
| 90 |
+
|
| 91 |
+
ptr = d["horiz_edge_ptr"].astype(np.int64, copy=False) # [L+1]
|
| 92 |
+
if ptr.size < 2:
|
| 93 |
+
raise ValueError("horiz_edge_ptr must have at least 2 entries")
|
| 94 |
+
e0 = e_all[:, int(ptr[0]):int(ptr[1])].astype(np.int64, copy=False) # slice рівня 0
|
| 95 |
+
|
| 96 |
+
# якщо чомусь індекси зсунуто — прибираємо офсет рівня 0
|
| 97 |
+
if level_offsets is not None and int(level_offsets[0]) != 0:
|
| 98 |
+
e0 = e0 - int(level_offsets[0])
|
| 99 |
+
|
| 100 |
+
if e0.min() < 0 or e0.max() >= N:
|
| 101 |
+
raise RuntimeError(f"edge_index_base out of range after slice: min={e0.min()}, max={e0.max()}, N={N}")
|
| 102 |
+
|
| 103 |
+
edge_index = torch.from_numpy(e0.copy())
|
| 104 |
+
|
| 105 |
+
# 3) старі ключі: edges_h_src/edges_h_dst або edges_horiz
|
| 106 |
+
elif "edges_h_src" in d and "edges_h_dst" in d:
|
| 107 |
+
src = torch.from_numpy(d["edges_h_src"].astype(np.int64, copy=False))
|
| 108 |
+
dst = torch.from_numpy(d["edges_h_dst"].astype(np.int64, copy=False))
|
| 109 |
+
edge_index = torch.stack([src, dst], dim=0)
|
| 110 |
+
elif "edges_horiz" in d:
|
| 111 |
+
eh = d["edges_horiz"]
|
| 112 |
+
if eh.ndim == 2 and eh.shape[1] == 2:
|
| 113 |
+
edge_index = torch.from_numpy(eh.astype(np.int64, copy=False)).t().contiguous()
|
| 114 |
+
elif eh.ndim == 2 and eh.shape[0] == 2:
|
| 115 |
+
edge_index = torch.from_numpy(eh.astype(np.int64, copy=False))
|
| 116 |
+
|
| 117 |
+
# --- упакуємо відповіді
|
| 118 |
+
meta = {
|
| 119 |
+
"N_mesh": N,
|
| 120 |
+
"levels": levels,
|
| 121 |
+
"level_offsets": level_offsets,
|
| 122 |
+
"mesh_hash": mesh_hash,
|
| 123 |
+
}
|
| 124 |
+
geo = static_geo.T # [4,N]
|
| 125 |
+
return {
|
| 126 |
+
"lon": lon, "lat": lat, "geo": geo, "pos2d": pos2d, # [4,N] і [4,N]
|
| 127 |
+
"edge_index_base": edge_index, # torch.Long[2,Eh] або None
|
| 128 |
+
"meta": meta,
|
| 129 |
+
}
|
| 130 |
+
|
| 131 |
+
|
| 132 |
+
# ==========================
|
| 133 |
+
# Dataset
|
| 134 |
+
# ==========================
|
| 135 |
+
class GraphShardDataset(Dataset):
|
| 136 |
+
"""
|
| 137 |
+
Зчитує графові шарди і віддає семпли для GAT-моделі.
|
| 138 |
+
|
| 139 |
+
__getitem__ -> ((x_all, geo, pos2d), y_all)
|
| 140 |
+
x_all: [Tin, C, N]
|
| 141 |
+
y_all: [Tout, C, N]
|
| 142 |
+
geo: [G, N] (4)
|
| 143 |
+
pos2d: [P, N] (4)
|
| 144 |
+
"""
|
| 145 |
+
def __init__(self,
|
| 146 |
+
shards_root: str,
|
| 147 |
+
years: range,
|
| 148 |
+
static_graph_npz: str = "./data/graph/graph_static_from_stats_vgeo_aligned.npz",
|
| 149 |
+
allowed_months: Optional[List[int]] = None, # None → всі 1..12
|
| 150 |
+
force_tin_tout: Optional[Tuple[int, int]] = None, # None → брати з шару
|
| 151 |
+
dtype_out: str = "float32",
|
| 152 |
+
shuffle_within_file: bool = True):
|
| 153 |
+
super().__init__()
|
| 154 |
+
self.root = Path(shards_root)
|
| 155 |
+
self.years = set(int(y) for y in years)
|
| 156 |
+
self.allowed_months = None if allowed_months is None else {int(m) for m in allowed_months}
|
| 157 |
+
self.force_tin_tout = force_tin_tout
|
| 158 |
+
assert dtype_out in {"float16", "float32"}
|
| 159 |
+
self.dtype_out = torch.float16 if dtype_out == "float16" else torch.float32
|
| 160 |
+
self.shuffle_within_file = bool(shuffle_within_file)
|
| 161 |
+
|
| 162 |
+
# ---- зчитуємо статичні фічі графа один раз ----
|
| 163 |
+
sg = _load_graph_static(Path(static_graph_npz))
|
| 164 |
+
self.geo = torch.from_numpy(sg["geo"].copy()) # [4,N]
|
| 165 |
+
self.pos2d = torch.from_numpy(sg["pos2d"].copy()) # [4,N]
|
| 166 |
+
self.N = int(sg["geo"].shape[1])
|
| 167 |
+
self.edge_index_base = sg["edge_index_base"] # може бути None
|
| 168 |
+
self.static_meta = sg["meta"]
|
| 169 |
+
|
| 170 |
+
# ---- зібрати список файлів і entries ----
|
| 171 |
+
files = []
|
| 172 |
+
for p in sorted(self.root.glob("graph_*.pt")):
|
| 173 |
+
y, m = _parse_ym_from_name(p)
|
| 174 |
+
if y is None:
|
| 175 |
+
continue
|
| 176 |
+
if (y in self.years) and (self.allowed_months is None or m in self.allowed_months):
|
| 177 |
+
files.append(p)
|
| 178 |
+
if not files:
|
| 179 |
+
raise FileNotFoundError(f"No graph_YYYY-MM.pt shards in {self.root}")
|
| 180 |
+
|
| 181 |
+
self.entries: List[Tuple[str, int]] = [] # (path, s_start)
|
| 182 |
+
self._file_meta: Dict[str, Dict] = {} # path -> {T,C,N,Tin,Tout,stride,S}
|
| 183 |
+
self._file_counts: Dict[str, int] = {}
|
| 184 |
+
|
| 185 |
+
for p in files:
|
| 186 |
+
d = torch.load(p, map_location="cpu")
|
| 187 |
+
values = d["values"] # [T,C,N]
|
| 188 |
+
T, C, N = values.shape
|
| 189 |
+
if N != self.N:
|
| 190 |
+
raise ValueError(f"{p.name}: N({N}) != static N({self.N})")
|
| 191 |
+
|
| 192 |
+
Tin = int(d["Tin"])
|
| 193 |
+
Tout = int(d["Tout"])
|
| 194 |
+
stride = int(d["stride"])
|
| 195 |
+
starts = d["sample_starts"].cpu().numpy().astype(np.int64) # [S]
|
| 196 |
+
if self.force_tin_tout is not None:
|
| 197 |
+
# Перерахувати sample_starts під інші Tin/Tout (залишимо stride=1)
|
| 198 |
+
fTin, fTout = self.force_tin_tout
|
| 199 |
+
max_start = T - (fTin + fTout)
|
| 200 |
+
if max_start < 0:
|
| 201 |
+
continue
|
| 202 |
+
starts = np.arange(0, max_start + 1, 1, dtype=np.int64)
|
| 203 |
+
Tin, Tout = int(fTin), int(fTout)
|
| 204 |
+
|
| 205 |
+
if self.shuffle_within_file:
|
| 206 |
+
rng = np.random.default_rng(int(y) * 100 + int(m))
|
| 207 |
+
rng.shuffle(starts)
|
| 208 |
+
|
| 209 |
+
for s in starts:
|
| 210 |
+
self.entries.append((str(p), int(s)))
|
| 211 |
+
|
| 212 |
+
self._file_meta[str(p)] = {
|
| 213 |
+
"T": T, "C": C, "N": N, "Tin": Tin, "Tout": Tout, "stride": stride, "S": int(len(starts))
|
| 214 |
+
}
|
| 215 |
+
self._file_counts[str(p)] = int(len(starts))
|
| 216 |
+
|
| 217 |
+
# cache last shard
|
| 218 |
+
self._last_path: Optional[str] = None
|
| 219 |
+
self._last_data = None
|
| 220 |
+
|
| 221 |
+
# зручні поля
|
| 222 |
+
self.C = next(iter(self._file_meta.values()))["C"]
|
| 223 |
+
self.Tin = next(iter(self._file_meta.values()))["Tin"]
|
| 224 |
+
self.Tout = next(iter(self._file_meta.values()))["Tout"]
|
| 225 |
+
|
| 226 |
+
def __len__(self) -> int:
|
| 227 |
+
return len(self.entries)
|
| 228 |
+
|
| 229 |
+
def _load_shard(self, path: str):
|
| 230 |
+
if self._last_path != path:
|
| 231 |
+
self._last_data = torch.load(path, map_location="cpu")
|
| 232 |
+
self._last_path = path
|
| 233 |
+
return self._last_data
|
| 234 |
+
|
| 235 |
+
def __getitem__(self, i: int):
|
| 236 |
+
path, s = self.entries[i]
|
| 237 |
+
meta = self._file_meta[path]
|
| 238 |
+
Tin, Tout = meta["Tin"], meta["Tout"]
|
| 239 |
+
|
| 240 |
+
d = self._load_shard(path)
|
| 241 |
+
vals = d["values"] # torch Tensor [T,C,N], може бути float16
|
| 242 |
+
# збираємо x/y як float32 (рекомендовано для стабільності)
|
| 243 |
+
x_all = vals[s:s+Tin].to(dtype=self.dtype_out) # [Tin,C,N]
|
| 244 |
+
y_all = vals[s+Tin:s+Tin+Tout].to(dtype=self.dtype_out) # [Tout,C,N]
|
| 245 |
+
|
| 246 |
+
# статичні — копії не потрібні, вони read-only, але distinct тензори бажані
|
| 247 |
+
geo = self.geo.to(dtype=self.dtype_out) # [4,N]
|
| 248 |
+
pos2d = self.pos2d.to(dtype=self.dtype_out) # [4,N]
|
| 249 |
+
|
| 250 |
+
return (x_all, geo, pos2d), y_all
|
| 251 |
+
|
| 252 |
+
# ---------- diagnostics ----------
|
| 253 |
+
@property
|
| 254 |
+
def file_counts(self) -> Dict[str, int]:
|
| 255 |
+
return dict(self._file_counts)
|
| 256 |
+
|
| 257 |
+
@property
|
| 258 |
+
def N_mesh(self) -> int:
|
| 259 |
+
return self.N
|
| 260 |
+
|
| 261 |
+
|
| 262 |
+
# ==========================
|
| 263 |
+
# Prefetch (GPU)
|
| 264 |
+
# ==========================
|
| 265 |
+
class PrefetchLoader:
|
| 266 |
+
"""GPU-prefetch з відомим __len__."""
|
| 267 |
+
def __init__(self, loader: DataLoader, device: str = 'cuda'):
|
| 268 |
+
self.loader = loader
|
| 269 |
+
self.device = device
|
| 270 |
+
|
| 271 |
+
def __len__(self):
|
| 272 |
+
return len(self.loader)
|
| 273 |
+
|
| 274 |
+
def __iter__(self):
|
| 275 |
+
if self.device.startswith("cuda"):
|
| 276 |
+
stream = torch.cuda.Stream()
|
| 277 |
+
first = True
|
| 278 |
+
for (x_all, geo, pos2d), y_all in self.loader:
|
| 279 |
+
with torch.cuda.stream(stream):
|
| 280 |
+
x_all = x_all.to(self.device, non_blocking=True)
|
| 281 |
+
geo = geo.to(self.device, non_blocking=True)
|
| 282 |
+
pos2d = pos2d.to(self.device, non_blocking=True)
|
| 283 |
+
y_all = y_all.to(self.device, non_blocking=True)
|
| 284 |
+
next_in, next_tgt = (x_all, geo, pos2d), y_all
|
| 285 |
+
if not first:
|
| 286 |
+
yield cur_in, cur_tgt
|
| 287 |
+
else:
|
| 288 |
+
first = False
|
| 289 |
+
torch.cuda.current_stream().wait_stream(stream)
|
| 290 |
+
cur_in, cur_tgt = next_in, next_tgt
|
| 291 |
+
yield cur_in, cur_tgt
|
| 292 |
+
else:
|
| 293 |
+
# CPU → прозора прокладка
|
| 294 |
+
for batch in self.loader:
|
| 295 |
+
yield batch
|
| 296 |
+
|
| 297 |
+
|
| 298 |
+
# ==========================
|
| 299 |
+
# Lightning DataModule
|
| 300 |
+
# ==========================
|
| 301 |
+
class GraphWeatherDataModule(pl.LightningDataModule):
|
| 302 |
+
"""
|
| 303 |
+
Створює train/val лоадери для графових шард-файлів.
|
| 304 |
+
Виносить edge_index_base і N_mesh як властивості, щоб інітити модель.
|
| 305 |
+
|
| 306 |
+
Batch: ((x_all, geo, pos2d), y_all)
|
| 307 |
+
x_all: [B, Tin, C, N]
|
| 308 |
+
y_all: [B, Tout, C, N]
|
| 309 |
+
geo: [B, 4, N]
|
| 310 |
+
pos2d: [B, 4, N]
|
| 311 |
+
"""
|
| 312 |
+
def __init__(self,
|
| 313 |
+
shards_root_train: str = "./data/shards_graph",
|
| 314 |
+
shards_root_val: str = "./data/shards_graph",
|
| 315 |
+
years_train: range = range(1980, 2011),
|
| 316 |
+
years_val: range = range(2011, 2013),
|
| 317 |
+
static_graph_npz: str = "./data/graph/graph_static_from_stats_vgeo_aligned.npz",
|
| 318 |
+
allowed_months: Optional[List[int]] = None,
|
| 319 |
+
force_tin_tout: Optional[Tuple[int,int]] = None,
|
| 320 |
+
dtype_out: str = "float32",
|
| 321 |
+
shuffle_within_file_train: bool = True,
|
| 322 |
+
shuffle_within_file_val: bool = False,
|
| 323 |
+
batch_size_train: int = 1,
|
| 324 |
+
batch_size_val: int = 1,
|
| 325 |
+
num_workers: Optional[int] = None,
|
| 326 |
+
pin_memory: bool = True,
|
| 327 |
+
persistent_workers: bool = True,
|
| 328 |
+
device: str = "cuda"):
|
| 329 |
+
super().__init__()
|
| 330 |
+
self.paths = {"train": shards_root_train, "val": shards_root_val}
|
| 331 |
+
self.years = {"train": years_train, "val": years_val}
|
| 332 |
+
self.kw_ds = dict(
|
| 333 |
+
static_graph_npz=static_graph_npz,
|
| 334 |
+
allowed_months=allowed_months,
|
| 335 |
+
force_tin_tout=force_tin_tout,
|
| 336 |
+
dtype_out=dtype_out,
|
| 337 |
+
)
|
| 338 |
+
self.shuffle = {"train": shuffle_within_file_train, "val": shuffle_within_file_val}
|
| 339 |
+
self.bs = {"train": batch_size_train, "val": batch_size_val}
|
| 340 |
+
self.num_workers = num_workers if num_workers is not None else max((torch.get_num_threads() or 1) - 1, 0)
|
| 341 |
+
self.pin_memory = pin_memory
|
| 342 |
+
self.persistent_workers = persistent_workers
|
| 343 |
+
self.device = device
|
| 344 |
+
|
| 345 |
+
self._train_ds = None
|
| 346 |
+
self._val_ds = None
|
| 347 |
+
self._train_loader = None
|
| 348 |
+
self._val_loader = None
|
| 349 |
+
|
| 350 |
+
# public: для ініціалізації моделі
|
| 351 |
+
self.edge_index_base: Optional[torch.Tensor] = None # Long[2,Eh] або None
|
| 352 |
+
self.N_mesh: Optional[int] = None
|
| 353 |
+
self.C: Optional[int] = None
|
| 354 |
+
self.Tin: Optional[int] = None
|
| 355 |
+
self.Tout: Optional[int] = None
|
| 356 |
+
|
| 357 |
+
def setup(self, stage=None):
|
| 358 |
+
self._train_ds = GraphShardDataset(
|
| 359 |
+
shards_root=self.paths["train"],
|
| 360 |
+
years=self.years["train"],
|
| 361 |
+
shuffle_within_file=self.shuffle["train"],
|
| 362 |
+
**self.kw_ds
|
| 363 |
+
)
|
| 364 |
+
self._val_ds = GraphShardDataset(
|
| 365 |
+
shards_root=self.paths["val"],
|
| 366 |
+
years=self.years["val"],
|
| 367 |
+
shuffle_within_file=self.shuffle["val"],
|
| 368 |
+
**self.kw_ds
|
| 369 |
+
)
|
| 370 |
+
|
| 371 |
+
# заповнимо публічні поля із train DS (ідентичні у val при тих самих даних)
|
| 372 |
+
self.N_mesh = self._train_ds.N_mesh
|
| 373 |
+
self.edge_index_base = self._train_ds.edge_index_base # може бути None
|
| 374 |
+
self.C = self._train_ds.C
|
| 375 |
+
self.Tin = self._train_ds.Tin
|
| 376 |
+
self.Tout = self._train_ds.Tout
|
| 377 |
+
|
| 378 |
+
def _make_loader(self, ds: Dataset, bs: int, shuffle: bool):
|
| 379 |
+
base = DataLoader(
|
| 380 |
+
ds, batch_size=bs, shuffle=shuffle,
|
| 381 |
+
num_workers=self.num_workers,
|
| 382 |
+
pin_memory=self.pin_memory,
|
| 383 |
+
persistent_workers=self.persistent_workers,
|
| 384 |
+
drop_last=False,
|
| 385 |
+
)
|
| 386 |
+
return PrefetchLoader(base, device=self.device)
|
| 387 |
+
|
| 388 |
+
def train_dataloader(self):
|
| 389 |
+
self._train_loader = self._make_loader(self._train_ds, self.bs["train"], shuffle=True)
|
| 390 |
+
return self._train_loader
|
| 391 |
+
|
| 392 |
+
def val_dataloader(self):
|
| 393 |
+
self._val_loader = self._make_loader(self._val_ds, self.bs["val"], shuffle=False)
|
| 394 |
+
return self._val_loader
|
| 395 |
+
|
| 396 |
+
# handy stats
|
| 397 |
+
@property
|
| 398 |
+
def train_samples(self):
|
| 399 |
+
return len(self._train_ds) if self._train_ds is not None else None
|
| 400 |
+
|
| 401 |
+
@property
|
| 402 |
+
def val_samples(self):
|
| 403 |
+
return len(self._val_ds) if self._val_ds is not None else None
|
| 404 |
+
|
| 405 |
+
@property
|
| 406 |
+
def train_steps_per_epoch(self):
|
| 407 |
+
if self._train_ds is None:
|
| 408 |
+
return None
|
| 409 |
+
return math.ceil(len(self._train_ds) / self.bs["train"])
|
| 410 |
+
|
| 411 |
+
@property
|
| 412 |
+
def val_steps(self):
|
| 413 |
+
if self._val_ds is None:
|
| 414 |
+
return None
|
| 415 |
+
return math.ceil(len(self._val_ds) / self.bs["val"])
|
unified_graph_weather_model.py
ADDED
|
@@ -0,0 +1,366 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# unified_graph_weather_model_gat.py
|
| 2 |
+
# Графова версія: Vertical (C) attention + Horizontal GAT over edges
|
| 3 |
+
from __future__ import annotations
|
| 4 |
+
from typing import Tuple, Optional, Literal
|
| 5 |
+
|
| 6 |
+
import torch
|
| 7 |
+
import torch.nn as nn
|
| 8 |
+
import torch.nn.functional as F
|
| 9 |
+
import pytorch_lightning as pl
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
# -----------------------------
|
| 13 |
+
# SDPA helpers (без Flash)
|
| 14 |
+
# -----------------------------
|
| 15 |
+
def _sdpa_no_flash_ctx():
|
| 16 |
+
try:
|
| 17 |
+
from torch.nn.attention import sdpa_kernel, SDPBackend
|
| 18 |
+
return sdpa_kernel(backends=[SDPBackend.MATH, SDPBackend.EFFICIENT_ATTENTION])
|
| 19 |
+
except Exception:
|
| 20 |
+
return torch.backends.cuda.sdp_kernel(
|
| 21 |
+
enable_flash=False, enable_math=True, enable_mem_efficient=True
|
| 22 |
+
)
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
class EfficientScaledDotProductAttention(nn.Module):
|
| 26 |
+
"""Self-attention по послідовності (torch SDPA)."""
|
| 27 |
+
def __init__(self, embed_dim: int, num_heads: int):
|
| 28 |
+
super().__init__()
|
| 29 |
+
assert embed_dim % num_heads == 0
|
| 30 |
+
self.num_heads = num_heads
|
| 31 |
+
self.embed_dim = embed_dim
|
| 32 |
+
self.head_dim = embed_dim // num_heads
|
| 33 |
+
self.q_proj = nn.Linear(embed_dim, embed_dim)
|
| 34 |
+
self.k_proj = nn.Linear(embed_dim, embed_dim)
|
| 35 |
+
self.v_proj = nn.Linear(embed_dim, embed_dim)
|
| 36 |
+
self.out_proj = nn.Linear(embed_dim, embed_dim)
|
| 37 |
+
self.norm = nn.LayerNorm(embed_dim)
|
| 38 |
+
|
| 39 |
+
def forward(self, x: torch.Tensor) -> torch.Tensor:
|
| 40 |
+
# x: [B, N, E]
|
| 41 |
+
B, N, _ = x.shape
|
| 42 |
+
q = self.q_proj(x).view(B, N, self.num_heads, self.head_dim).transpose(1, 2) # [B,h,N,d]
|
| 43 |
+
k = self.k_proj(x).view(B, N, self.num_heads, self.head_dim).transpose(1, 2)
|
| 44 |
+
v = self.v_proj(x).view(B, N, self.num_heads, self.head_dim).transpose(1, 2)
|
| 45 |
+
with _sdpa_no_flash_ctx():
|
| 46 |
+
attn = F.scaled_dot_product_attention(q, k, v, dropout_p=0.0, is_causal=False)
|
| 47 |
+
attn = attn.transpose(1, 2).contiguous().view(B, N, self.embed_dim) # [B,N,E]
|
| 48 |
+
out = self.out_proj(attn)
|
| 49 |
+
return self.norm(out + x)
|
| 50 |
+
|
| 51 |
+
|
| 52 |
+
# -----------------------------
|
| 53 |
+
# GAT mixer (multi-head, без сторонніх бібліотек)
|
| 54 |
+
# -----------------------------
|
| 55 |
+
class GraphGATMixer(nn.Module):
|
| 56 |
+
"""
|
| 57 |
+
Multi-head GAT по горизонтальних ребрах.
|
| 58 |
+
Працює окремо для кожного вертикального токена (каналу) і кожного елемента батчу.
|
| 59 |
+
|
| 60 |
+
Вхід: H [B, C, N, E]
|
| 61 |
+
Вихід: H' того ж розміру
|
| 62 |
+
|
| 63 |
+
edge_index: Long[2, Eh] (індекси 0..N-1)
|
| 64 |
+
Порада: бажано додати самопетлі в edge_index поза моделлю (i->i), або використовується skip-резідюал.
|
| 65 |
+
"""
|
| 66 |
+
def __init__(self, embed_dim: int, num_heads: int = 4, attn_drop: float = 0.0, neg_slope: float = 0.2):
|
| 67 |
+
super().__init__()
|
| 68 |
+
assert embed_dim % num_heads == 0, "embed_dim must be divisible by num_heads"
|
| 69 |
+
self.E = embed_dim
|
| 70 |
+
self.H = num_heads
|
| 71 |
+
self.Dh = embed_dim // num_heads
|
| 72 |
+
self.attn_drop = attn_drop
|
| 73 |
+
self.leaky = nn.LeakyReLU(neg_slope)
|
| 74 |
+
|
| 75 |
+
# Лінійна проєкція у head-простір
|
| 76 |
+
self.lin = nn.Linear(self.E, self.E, bias=False)
|
| 77 |
+
|
| 78 |
+
# a_src / a_dst параметри на кожну голову: [H, Dh]
|
| 79 |
+
self.a_src = nn.Parameter(torch.empty(self.H, self.Dh))
|
| 80 |
+
self.a_dst = nn.Parameter(torch.empty(self.H, self.Dh))
|
| 81 |
+
|
| 82 |
+
# вихідна проєкція + нормалізація
|
| 83 |
+
self.out_proj = nn.Linear(self.E, self.E, bias=True)
|
| 84 |
+
self.norm = nn.LayerNorm(self.E)
|
| 85 |
+
self.act = nn.GELU()
|
| 86 |
+
self.reset_parameters()
|
| 87 |
+
|
| 88 |
+
def reset_parameters(self):
|
| 89 |
+
nn.init.xavier_uniform_(self.lin.weight)
|
| 90 |
+
nn.init.xavier_uniform_(self.out_proj.weight)
|
| 91 |
+
nn.init.zeros_(self.out_proj.bias)
|
| 92 |
+
nn.init.xavier_uniform_(self.a_src.unsqueeze(-1))
|
| 93 |
+
nn.init.xavier_uniform_(self.a_dst.unsqueeze(-1))
|
| 94 |
+
|
| 95 |
+
def forward(self, H: torch.Tensor, edge_index: torch.Tensor) -> torch.Tensor:
|
| 96 |
+
"""
|
| 97 |
+
H: [B, C, N, E]
|
| 98 |
+
edge_index: [2, Eh]
|
| 99 |
+
"""
|
| 100 |
+
B, C, N, E = H.shape
|
| 101 |
+
src = edge_index[0].to(H.device, non_blocking=True)
|
| 102 |
+
dst = edge_index[1].to(H.device, non_blocking=True)
|
| 103 |
+
Eh = src.numel()
|
| 104 |
+
|
| 105 |
+
# Перетворення у [B*C, N, H, Dh]
|
| 106 |
+
X = H.view(B * C, N, E)
|
| 107 |
+
X = self.lin(X) # [BC, N, E]
|
| 108 |
+
X = X.view(B * C, N, self.H, self.Dh) # [BC, N, H, Dh]
|
| 109 |
+
|
| 110 |
+
# За один прохід пройдемо BC-семпли
|
| 111 |
+
out = torch.zeros(B * C, N, self.H, self.Dh, device=H.device, dtype=H.dtype)
|
| 112 |
+
|
| 113 |
+
# Попередньо підготуємо a_src/dst у потрібному dtype/пристрої
|
| 114 |
+
a_src = self.a_src.to(device=H.device, dtype=H.dtype)
|
| 115 |
+
a_dst = self.a_dst.to(device=H.device, dtype=H.dtype)
|
| 116 |
+
|
| 117 |
+
for bc in range(B * C):
|
| 118 |
+
Xbc = X[bc] # [N, H, Dh]
|
| 119 |
+
# Витягуємо ознаки джерел і приймачів по ребрах
|
| 120 |
+
x_i = Xbc.index_select(0, src) # [Eh, H, Dh] (source features)
|
| 121 |
+
x_j = Xbc.index_select(0, dst) # [Eh, H, Dh] (target features)
|
| 122 |
+
|
| 123 |
+
# e_ij = Leaky(a_src·x_i + a_dst·x_j), по головах
|
| 124 |
+
e_src = torch.einsum("ehd,hd->eh", x_i, a_src) # [Eh, H]
|
| 125 |
+
e_dst = torch.einsum("ehd,hd->eh", x_j, a_dst) # [Eh, H]
|
| 126 |
+
e = self.leaky(e_src + e_dst) # [Eh, H]
|
| 127 |
+
|
| 128 |
+
# стабілізуємо softmax: обмежимо значення
|
| 129 |
+
e = torch.clamp(e, -10.0, 10.0)
|
| 130 |
+
exp_e = torch.exp(e) # [Eh, H]
|
| 131 |
+
if self.attn_drop > 0:
|
| 132 |
+
exp_e = F.dropout(exp_e, p=self.attn_drop, training=self.training)
|
| 133 |
+
|
| 134 |
+
# ∑_{k in N(dst)} exp_e; нормалізація по вхідних ребрах кожного dst (per head)
|
| 135 |
+
# Використовуємо index_add по другій осі (уздовж вузлів)
|
| 136 |
+
sum_dst = torch.zeros(self.H, N, device=H.device, dtype=H.dtype) # [H, N]
|
| 137 |
+
# transpose -> [H, Eh], add along dim=1 at indices 'dst'
|
| 138 |
+
sum_dst.index_add_(1, dst, exp_e.transpose(0, 1))
|
| 139 |
+
# вибрати суму для кожного ребра/голови
|
| 140 |
+
denom = sum_dst[:, dst].transpose(0, 1) + 1e-12 # [Eh, H]
|
| 141 |
+
alpha = exp_e / denom # [Eh, H]
|
| 142 |
+
|
| 143 |
+
# повідомлення від джерела до приймача: alpha * x_i (можна x_j; беремо x_i як у GAT v1)
|
| 144 |
+
m = alpha.unsqueeze(-1) * x_i # [Eh, H, Dh]
|
| 145 |
+
|
| 146 |
+
# агрегуємо у вузли-приймачі (per head)
|
| 147 |
+
# робимо head-цикл — H невелике, зате пам’ять економна
|
| 148 |
+
for h in range(self.H):
|
| 149 |
+
out[bc, :, h, :].index_add_(0, dst, m[:, h, :]) # [N, Dh]
|
| 150 |
+
|
| 151 |
+
# concat голови → [BC, N, E]
|
| 152 |
+
out = out.reshape(B * C, N, self.H * self.Dh)
|
| 153 |
+
out = self.out_proj(out) # [BC, N, E]
|
| 154 |
+
out = self.act(out)
|
| 155 |
+
# назад у [B, E, C, N]
|
| 156 |
+
out = out.view(B, C, N, E).permute(0, 3, 1, 2).contiguous() # [B, E, C, N]
|
| 157 |
+
# skip connection + norm опрацьовуємо у блоці
|
| 158 |
+
return out
|
| 159 |
+
|
| 160 |
+
|
| 161 |
+
# -----------------------------
|
| 162 |
+
# Один блок: Vertical SDPA + GAT + FFN
|
| 163 |
+
# -----------------------------
|
| 164 |
+
class GraphAxialGATBlock(nn.Module):
|
| 165 |
+
"""
|
| 166 |
+
Працює у представленні [B, E, C, N].
|
| 167 |
+
1) Vertical attention (по C у кожній вершині)
|
| 168 |
+
2) GAT по графу (по горизонталі) окремо для кожного каналу
|
| 169 |
+
3) FFN (1x1 Conv2d) + residual + norm
|
| 170 |
+
"""
|
| 171 |
+
def __init__(self, E: int, heads_v: int = 4, heads_xy: int = 4, attn_drop_xy: float = 0.0):
|
| 172 |
+
super().__init__()
|
| 173 |
+
self.attn_v = EfficientScaledDotProductAttention(E, heads_v)
|
| 174 |
+
self.gat = GraphGATMixer(E, num_heads=heads_xy, attn_drop=attn_drop_xy)
|
| 175 |
+
self.ffn = nn.Sequential(
|
| 176 |
+
nn.Conv2d(E, 2 * E, kernel_size=1), nn.GELU(),
|
| 177 |
+
nn.Conv2d(2 * E, E, kernel_size=1),
|
| 178 |
+
)
|
| 179 |
+
self.norm = nn.GroupNorm(8, E)
|
| 180 |
+
|
| 181 |
+
def forward(self, F: torch.Tensor, edge_index: torch.Tensor) -> torch.Tensor:
|
| 182 |
+
# (1) Vertical attention
|
| 183 |
+
B, E, C, N = F.shape
|
| 184 |
+
V = F.permute(0, 3, 2, 1).reshape(B * N, C, E) # [B*N, C, E]
|
| 185 |
+
V = self.attn_v(V) # [B*N, C, E]
|
| 186 |
+
V = V.reshape(B, N, C, E).permute(0, 3, 2, 1) # [B, E, C, N]
|
| 187 |
+
|
| 188 |
+
# (2) GAT mixing (per channel)
|
| 189 |
+
G_in = V.permute(0, 2, 3, 1).contiguous() # [B, C, N, E]
|
| 190 |
+
G_out = self.gat(G_in, edge_index=edge_index) # [B, E, C, N]
|
| 191 |
+
|
| 192 |
+
S = V + G_out # residual після GAT
|
| 193 |
+
|
| 194 |
+
# (3) FFN + residual + norm
|
| 195 |
+
U = self.ffn(S)
|
| 196 |
+
return self.norm(U + S)
|
| 197 |
+
|
| 198 |
+
|
| 199 |
+
# -----------------------------
|
| 200 |
+
# Core network
|
| 201 |
+
# -----------------------------
|
| 202 |
+
class UnifiedGraphSpatioVerticalGATNet(nn.Module):
|
| 203 |
+
"""
|
| 204 |
+
input: (x_all, geo, pos2d), diffusion_step(optional)
|
| 205 |
+
x_all: [B, Tin, C, N]
|
| 206 |
+
geo: [B, G, N]
|
| 207 |
+
pos2d: [B, P, N]
|
| 208 |
+
output: [B, Tout, C, N]
|
| 209 |
+
|
| 210 |
+
Потрібно подати edge_index_base: Long[2, Eh] (горизонтальні ребра базового рівня).
|
| 211 |
+
"""
|
| 212 |
+
def __init__(self,
|
| 213 |
+
C: int,
|
| 214 |
+
Tin: int,
|
| 215 |
+
Tout: int,
|
| 216 |
+
edge_index_base: torch.Tensor,
|
| 217 |
+
N_mesh: int,
|
| 218 |
+
static_geo_ch: int = 4,
|
| 219 |
+
static_pos2d_ch: int = 4,
|
| 220 |
+
embed_dim: int = 192,
|
| 221 |
+
blocks: int = 4,
|
| 222 |
+
heads_v: int = 4,
|
| 223 |
+
heads_xy: int = 4,
|
| 224 |
+
attn_drop_xy: float = 0.0):
|
| 225 |
+
super().__init__()
|
| 226 |
+
self.C = int(C)
|
| 227 |
+
self.Tin = int(Tin)
|
| 228 |
+
self.Tout = int(Tout)
|
| 229 |
+
self.N = int(N_mesh)
|
| 230 |
+
self.static_in = int(static_geo_ch + static_pos2d_ch)
|
| 231 |
+
E = int(embed_dim)
|
| 232 |
+
|
| 233 |
+
# зафіксуємо ребра як буфери
|
| 234 |
+
if edge_index_base.dtype != torch.long:
|
| 235 |
+
edge_index_base = edge_index_base.long()
|
| 236 |
+
self.register_buffer("edge_src", edge_index_base[0].contiguous())
|
| 237 |
+
self.register_buffer("edge_dst", edge_index_base[1].contiguous())
|
| 238 |
+
|
| 239 |
+
# diffusion-step embedding (опц.)
|
| 240 |
+
self.diffusion_embed = nn.Sequential(
|
| 241 |
+
nn.Linear(1, E), nn.ReLU(inplace=True),
|
| 242 |
+
nn.Linear(E, E)
|
| 243 |
+
)
|
| 244 |
+
|
| 245 |
+
# Temporal projection: [B, Tin, C, N] -> [B, E, C, N]
|
| 246 |
+
self.temporal_proj = nn.Conv2d(self.Tin, E, kernel_size=1)
|
| 247 |
+
|
| 248 |
+
# Static encoder: [B, static_in, 1, N] -> [B, E, 1, N] -> broadcast по C
|
| 249 |
+
self.static_encoder = nn.Sequential(
|
| 250 |
+
nn.Conv2d(self.static_in, 64, kernel_size=1),
|
| 251 |
+
nn.GroupNorm(8, 64), nn.ReLU(inplace=True),
|
| 252 |
+
nn.Conv2d(64, E, kernel_size=1),
|
| 253 |
+
nn.GroupNorm(8, E), nn.ReLU(inplace=True),
|
| 254 |
+
)
|
| 255 |
+
|
| 256 |
+
# Blocks
|
| 257 |
+
self.blocks = nn.ModuleList([
|
| 258 |
+
GraphAxialGATBlock(E, heads_v=heads_v, heads_xy=heads_xy, attn_drop_xy=attn_drop_xy)
|
| 259 |
+
for _ in range(blocks)
|
| 260 |
+
])
|
| 261 |
+
|
| 262 |
+
# Decoder: [B, E, C, N] -> [B, Tout, C, N]
|
| 263 |
+
self.decoder = nn.Sequential(
|
| 264 |
+
nn.Conv2d(E, E, kernel_size=1), nn.GELU(),
|
| 265 |
+
nn.Conv2d(E, self.Tout, kernel_size=1)
|
| 266 |
+
)
|
| 267 |
+
|
| 268 |
+
def forward(self,
|
| 269 |
+
inputs: Tuple[torch.Tensor, torch.Tensor, torch.Tensor],
|
| 270 |
+
diffusion_step: Optional[torch.Tensor] = None) -> torch.Tensor:
|
| 271 |
+
x_all, geo, pos2d = inputs
|
| 272 |
+
x_all = x_all.float() # [B, Tin, C, N]
|
| 273 |
+
geo = geo.float() # [B, G, N]
|
| 274 |
+
pos2d = pos2d.float() # [B, P, N]
|
| 275 |
+
|
| 276 |
+
B, Tin, C, N = x_all.shape
|
| 277 |
+
assert Tin == self.Tin and C == self.C and N == self.N, \
|
| 278 |
+
f"Shape mismatch: x_all={x_all.shape} expected Tin={self.Tin}, C={self.C}, N={self.N}"
|
| 279 |
+
|
| 280 |
+
# (1) Temporal projection
|
| 281 |
+
F = self.temporal_proj(x_all) # [B, E, C, N]
|
| 282 |
+
|
| 283 |
+
# (2) Diffusion step embedding (optional)
|
| 284 |
+
if diffusion_step is None:
|
| 285 |
+
diffusion_step = torch.zeros(B, device=x_all.device, dtype=x_all.dtype)
|
| 286 |
+
d_emb = self.diffusion_embed(diffusion_step.unsqueeze(-1)) # [B, E]
|
| 287 |
+
d_map = d_emb.view(B, -1, 1, 1).expand(-1, -1, C, N) # [B, E, C, N]
|
| 288 |
+
F = F + d_map
|
| 289 |
+
|
| 290 |
+
# (3) Static features (broadcast per channel)
|
| 291 |
+
static = torch.cat([geo, pos2d], dim=1) # [B, static_in, N]
|
| 292 |
+
static_e = self.static_encoder(static.unsqueeze(2)) # [B, E, 1, N]
|
| 293 |
+
F = F + static_e.expand(-1, -1, C, -1) # [B, E, C, N]
|
| 294 |
+
|
| 295 |
+
# (4) Axial GAT blocks
|
| 296 |
+
edge_index = torch.stack([self.edge_src, self.edge_dst], dim=0)
|
| 297 |
+
for blk in self.blocks:
|
| 298 |
+
F = blk(F, edge_index=edge_index)
|
| 299 |
+
|
| 300 |
+
# (5) Decode to Tout
|
| 301 |
+
y = self.decoder(F) # [B, Tout, C, N]
|
| 302 |
+
return y
|
| 303 |
+
|
| 304 |
+
|
| 305 |
+
# -----------------------------
|
| 306 |
+
# Lightning wrapper
|
| 307 |
+
# -----------------------------
|
| 308 |
+
class UnifiedGraphWeatherGATLightning(pl.LightningModule):
|
| 309 |
+
"""
|
| 310 |
+
Очікує batches у форматі:
|
| 311 |
+
((x_all, geo, pos2d), y_all)
|
| 312 |
+
де:
|
| 313 |
+
x_all: [B, Tin, C, N]
|
| 314 |
+
y_all: [B, Tout, C, N]
|
| 315 |
+
"""
|
| 316 |
+
def __init__(self,
|
| 317 |
+
C: int,
|
| 318 |
+
Tin: int,
|
| 319 |
+
Tout: int,
|
| 320 |
+
edge_index_base: torch.Tensor,
|
| 321 |
+
N_mesh: int,
|
| 322 |
+
embed_dim: int = 192,
|
| 323 |
+
blocks: int = 4,
|
| 324 |
+
heads_v: int = 4,
|
| 325 |
+
heads_xy: int = 4,
|
| 326 |
+
attn_drop_xy: float = 0.0,
|
| 327 |
+
static_geo_ch: int = 4,
|
| 328 |
+
static_pos2d_ch: int = 4,
|
| 329 |
+
lr: float = 3e-4,
|
| 330 |
+
weight_decay: float = 1e-4,
|
| 331 |
+
loss: Literal["mse", "huber"] = "mse"):
|
| 332 |
+
super().__init__()
|
| 333 |
+
self.save_hyperparameters(ignore=["edge_index_base"])
|
| 334 |
+
self.net = UnifiedGraphSpatioVerticalGATNet(
|
| 335 |
+
C=C, Tin=Tin, Tout=Tout, N_mesh=N_mesh,
|
| 336 |
+
edge_index_base=edge_index_base,
|
| 337 |
+
static_geo_ch=static_geo_ch, static_pos2d_ch=static_pos2d_ch,
|
| 338 |
+
embed_dim=embed_dim, blocks=blocks,
|
| 339 |
+
heads_v=heads_v, heads_xy=heads_xy, attn_drop_xy=attn_drop_xy,
|
| 340 |
+
)
|
| 341 |
+
self.criterion = nn.MSELoss() if loss == "mse" else nn.HuberLoss()
|
| 342 |
+
self.lr = lr
|
| 343 |
+
self.weight_decay = weight_decay
|
| 344 |
+
|
| 345 |
+
def forward(self, x_tuple, diffusion_step: Optional[torch.Tensor] = None):
|
| 346 |
+
return self.net(x_tuple, diffusion_step=diffusion_step)
|
| 347 |
+
|
| 348 |
+
def _step(self, batch, stage: str):
|
| 349 |
+
(x_all, geo, pos2d), y_true = batch
|
| 350 |
+
B = x_all.size(0)
|
| 351 |
+
t = torch.zeros(B, device=x_all.device, dtype=x_all.dtype) # нульовий diffusion-step
|
| 352 |
+
y_pred = self.forward((x_all, geo, pos2d), diffusion_step=t)
|
| 353 |
+
loss = self.criterion(y_pred, y_true)
|
| 354 |
+
self.log(f"{stage}_loss", loss, prog_bar=True, on_step=(stage=="train"), on_epoch=True, batch_size=B)
|
| 355 |
+
return loss
|
| 356 |
+
|
| 357 |
+
def training_step(self, batch, batch_idx):
|
| 358 |
+
return self._step(batch, "train")
|
| 359 |
+
|
| 360 |
+
def validation_step(self, batch, batch_idx):
|
| 361 |
+
self._step(batch, "val")
|
| 362 |
+
|
| 363 |
+
def configure_optimizers(self):
|
| 364 |
+
opt = torch.optim.AdamW(self.parameters(), lr=self.lr, weight_decay=self.weight_decay)
|
| 365 |
+
sch = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=self.trainer.max_epochs)
|
| 366 |
+
return {"optimizer": opt, "lr_scheduler": {"scheduler": sch, "interval": "epoch"}}
|
unified_graph_weather_model_tp.py
ADDED
|
@@ -0,0 +1,272 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# model/unified_graph_weather_model_tp.py
|
| 2 |
+
import math
|
| 3 |
+
from typing import Optional, Sequence, Tuple
|
| 4 |
+
|
| 5 |
+
import torch
|
| 6 |
+
import torch.nn as nn
|
| 7 |
+
import pytorch_lightning as pl
|
| 8 |
+
|
| 9 |
+
# базова загальна мережа
|
| 10 |
+
from model.unified_graph_weather_model import UnifiedGraphSpatioVerticalGATNet
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
# ----------------------- допоміжні утиліти -----------------------
|
| 14 |
+
def _coerce_list_str(x) -> list[str]:
|
| 15 |
+
if isinstance(x, (list, tuple)):
|
| 16 |
+
return [str(v.decode("utf-8")) if isinstance(v, (bytes, bytearray)) else str(v) for v in x]
|
| 17 |
+
if isinstance(x, str):
|
| 18 |
+
return [x]
|
| 19 |
+
return [str(x)]
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
def _find_tp_index(order: Sequence[str]) -> int:
|
| 23 |
+
"""
|
| 24 |
+
Знаходимо індекс каналу опадів у shard order.
|
| 25 |
+
Пробуємо 'tp_log@sfc', 'tp@sfc', 'tp_log', 'tp' (в порядку пріоритету).
|
| 26 |
+
"""
|
| 27 |
+
order = [str(v) for v in order]
|
| 28 |
+
low = [s.lower() for s in order]
|
| 29 |
+
for cand in ("tp_log@sfc", "tp@sfc", "tp_log", "tp"):
|
| 30 |
+
if cand.lower() in low:
|
| 31 |
+
return low.index(cand.lower())
|
| 32 |
+
raise KeyError(f"Не знайшов 'tp' каналу у channel_order. Приклади: {order[:6]} ...")
|
| 33 |
+
|
| 34 |
+
|
| 35 |
+
def _stats_key_for_tp(name: str) -> str:
|
| 36 |
+
# у статистиці ми використовуємо лог-опади
|
| 37 |
+
return "tp_log"
|
| 38 |
+
|
| 39 |
+
|
| 40 |
+
# ---------- зручні перетворення (нормалізований лог ↔ фізичні мм) ----------
|
| 41 |
+
class TPDenormHelper:
|
| 42 |
+
"""
|
| 43 |
+
Конвертації між нормалізованим 'tp_log' і фізичними мм.
|
| 44 |
+
Використовує ROI-статистику: mean/std для 'tp_log'.
|
| 45 |
+
"""
|
| 46 |
+
def __init__(self, stats_order: Sequence[str], mu: torch.Tensor, sd: torch.Tensor):
|
| 47 |
+
if isinstance(mu, (list, tuple)):
|
| 48 |
+
mu = torch.tensor(mu, dtype=torch.float32)
|
| 49 |
+
if isinstance(sd, (list, tuple)):
|
| 50 |
+
sd = torch.tensor(sd, dtype=torch.float32)
|
| 51 |
+
|
| 52 |
+
key = _stats_key_for_tp("tp")
|
| 53 |
+
idx = _coerce_list_str(stats_order).index(key)
|
| 54 |
+
self.mu = float(mu[idx])
|
| 55 |
+
self.sd = float(sd[idx])
|
| 56 |
+
|
| 57 |
+
def lognorm_to_mm(self, arr_norm: torch.Tensor) -> torch.Tensor:
|
| 58 |
+
# arr_norm → arr_log → мм
|
| 59 |
+
arr_log = arr_norm * self.sd + self.mu
|
| 60 |
+
mm = torch.expm1(arr_log)
|
| 61 |
+
return torch.clamp(mm, min=0.0)
|
| 62 |
+
|
| 63 |
+
def mm_to_lognorm(self, mm: torch.Tensor) -> torch.Tensor:
|
| 64 |
+
arr_log = torch.log1p(torch.clamp(mm, min=0.0))
|
| 65 |
+
return (arr_log - self.mu) / max(self.sd, 1e-6)
|
| 66 |
+
|
| 67 |
+
|
| 68 |
+
# ----------------------- лосс під опади -----------------------
|
| 69 |
+
class SparsePrecipLoss(nn.Module):
|
| 70 |
+
"""
|
| 71 |
+
Комбінований лосс для опадів:
|
| 72 |
+
L = huber(log_norm) + α * w * |mm_pred - mm_true|
|
| 73 |
+
де w підсилює "мокрі" пікселі (mm_true >= wet_thr_mm), а для дуже малих опадів
|
| 74 |
+
можна тримати меншу вагу.
|
| 75 |
+
|
| 76 |
+
Параметри:
|
| 77 |
+
huber_delta: δ для Huber в нормалізованому лог-просторі
|
| 78 |
+
wet_thr_mm: поріг "мокро" в міліметрах
|
| 79 |
+
wet_weight: множник ваги для мокрих пікселів
|
| 80 |
+
dry_weight: вага для сухих/майже сухих пікселів
|
| 81 |
+
alpha_mm: коефіцієнт при терміні в мм
|
| 82 |
+
gamma_focal: опційно фокалізація для мокрих (підсилює великі значення)
|
| 83 |
+
"""
|
| 84 |
+
def __init__(self,
|
| 85 |
+
huber_delta: float = 0.5,
|
| 86 |
+
wet_thr_mm: float = 0.1,
|
| 87 |
+
wet_weight: float = 5.0,
|
| 88 |
+
dry_weight: float = 0.5,
|
| 89 |
+
alpha_mm: float = 0.5,
|
| 90 |
+
gamma_focal: float = 0.0):
|
| 91 |
+
super().__init__()
|
| 92 |
+
self.delta = float(huber_delta)
|
| 93 |
+
self.wet_thr = float(wet_thr_mm)
|
| 94 |
+
self.w_wet = float(wet_weight)
|
| 95 |
+
self.w_dry = float(dry_weight)
|
| 96 |
+
self.alpha = float(alpha_mm)
|
| 97 |
+
self.gamma = float(gamma_focal)
|
| 98 |
+
self.huber = nn.SmoothL1Loss(beta=self.delta, reduction="none") # PyTorch Huber v2
|
| 99 |
+
|
| 100 |
+
def forward(self,
|
| 101 |
+
y_pred_norm: torch.Tensor, # [B, N] (tp_log нормалізований)
|
| 102 |
+
y_true_norm: torch.Tensor, # [B, N]
|
| 103 |
+
denorm_helper: TPDenormHelper) -> torch.Tensor:
|
| 104 |
+
|
| 105 |
+
# 1) Huber у нормалізованому лог-просторі
|
| 106 |
+
L_h = self.huber(y_pred_norm, y_true_norm) # [B, N]
|
| 107 |
+
|
| 108 |
+
if self.alpha <= 0.0:
|
| 109 |
+
return L_h.mean()
|
| 110 |
+
|
| 111 |
+
# 2) Додатковий термін у фізичних мм
|
| 112 |
+
mm_pred = denorm_helper.lognorm_to_mm(y_pred_norm) # [B,N]
|
| 113 |
+
mm_true = denorm_helper.lognorm_to_mm(y_true_norm) # [B,N]
|
| 114 |
+
L_mm = torch.abs(mm_pred - mm_true)
|
| 115 |
+
|
| 116 |
+
# Ваги: мокре/сухе
|
| 117 |
+
wet_mask = (mm_true >= self.wet_thr).float()
|
| 118 |
+
dry_mask = 1.0 - wet_mask
|
| 119 |
+
w = self.w_wet * wet_mask + self.w_dry * dry_mask
|
| 120 |
+
|
| 121 |
+
if self.gamma > 0.0:
|
| 122 |
+
# фокалізація: підсилити великі значення (не обов’язково)
|
| 123 |
+
w = w * torch.pow(1.0 + mm_true, self.gamma)
|
| 124 |
+
|
| 125 |
+
L = L_h + self.alpha * (w * L_mm)
|
| 126 |
+
return L.mean()
|
| 127 |
+
|
| 128 |
+
|
| 129 |
+
# ------------------- TP-адаптер над загальною моделлю -------------------
|
| 130 |
+
class UnifiedGraphSpatioVerticalGATNetTP(nn.Module):
|
| 131 |
+
"""
|
| 132 |
+
Обгортка над загальною моделлю, що повертає тільки 1 канал (tp_log@sfc).
|
| 133 |
+
Внутрішня вартість майже як у повної, але head узятий з бази (ми лише обрізаємо вихід).
|
| 134 |
+
"""
|
| 135 |
+
def __init__(self,
|
| 136 |
+
C: int, Tin: int, Tout: int,
|
| 137 |
+
edge_index_base: torch.Tensor,
|
| 138 |
+
N_mesh: Optional[int] = None,
|
| 139 |
+
embed_dim: int = 192, blocks: int = 4, heads_v: int = 4, heads_xy: int = 4,
|
| 140 |
+
attn_drop_xy: float = 0.0,
|
| 141 |
+
channel_order: Optional[Sequence[str]] = None):
|
| 142 |
+
super().__init__()
|
| 143 |
+
self.base = UnifiedGraphSpatioVerticalGATNet(
|
| 144 |
+
C=C, Tin=Tin, Tout=Tout,
|
| 145 |
+
edge_index_base=edge_index_base, N_mesh=N_mesh,
|
| 146 |
+
embed_dim=embed_dim, blocks=blocks, heads_v=heads_v, heads_xy=heads_xy,
|
| 147 |
+
attn_drop_xy=attn_drop_xy,
|
| 148 |
+
)
|
| 149 |
+
self.channel_order = list(channel_order) if channel_order is not None else None
|
| 150 |
+
self.idx_tp = None
|
| 151 |
+
if self.channel_order is not None:
|
| 152 |
+
self.idx_tp = _find_tp_index(self.channel_order)
|
| 153 |
+
|
| 154 |
+
def set_channel_order(self, order: Sequence[str]):
|
| 155 |
+
self.channel_order = list(order)
|
| 156 |
+
self.idx_tp = _find_tp_index(self.channel_order)
|
| 157 |
+
|
| 158 |
+
def forward(self, x_tuple, diffusion_step: Optional[torch.Tensor] = None,
|
| 159 |
+
channel_order: Optional[Sequence[str]] = None):
|
| 160 |
+
"""
|
| 161 |
+
x_tuple: (x_all, geo, pos2d)
|
| 162 |
+
x_all: [B, Tin, C, N]
|
| 163 |
+
geo: [B, 4, N]
|
| 164 |
+
pos2d: [B, 4, N]
|
| 165 |
+
return: [B, Tout, 1, N] (tp_log нормалізований)
|
| 166 |
+
"""
|
| 167 |
+
if (self.idx_tp is None) and (channel_order is not None):
|
| 168 |
+
self.set_channel_order(channel_order)
|
| 169 |
+
|
| 170 |
+
y_full = self.base(x_tuple, diffusion_step=diffusion_step) # [B,Tout,C,N]
|
| 171 |
+
if self.idx_tp is None:
|
| 172 |
+
raise RuntimeError("Не визначено індекс tp у channel_order. Передай order у forward або set_channel_order().")
|
| 173 |
+
tp = y_full[:, :, self.idx_tp].unsqueeze(2) # [B, Tout, 1, N]
|
| 174 |
+
return tp
|
| 175 |
+
|
| 176 |
+
|
| 177 |
+
# ------------------------- Lightning під опади -------------------------
|
| 178 |
+
class PrecipTPLightning(pl.LightningModule):
|
| 179 |
+
"""
|
| 180 |
+
1-кроковий TP (tp_log@sfc), лосс адаптований до розрідженості.
|
| 181 |
+
Підтримує файнтюн з загального чекпойнта (через .net.base.load_state_dict()).
|
| 182 |
+
"""
|
| 183 |
+
def __init__(self,
|
| 184 |
+
C: int, Tin: int, Tout: int,
|
| 185 |
+
edge_index_base: torch.Tensor,
|
| 186 |
+
N_mesh: Optional[int] = None,
|
| 187 |
+
embed_dim: int = 192, blocks: int = 4, heads_v: int = 4, heads_xy: int = 4,
|
| 188 |
+
attn_drop_xy: float = 0.0,
|
| 189 |
+
lr: float = 3e-4, weight_decay: float = 1e-4,
|
| 190 |
+
# лосс-параметри
|
| 191 |
+
huber_delta: float = 0.5,
|
| 192 |
+
wet_thr_mm: float = 0.1,
|
| 193 |
+
wet_weight: float = 5.0,
|
| 194 |
+
dry_weight: float = 0.5,
|
| 195 |
+
alpha_mm: float = 0.5,
|
| 196 |
+
gamma_focal: float = 0.0,
|
| 197 |
+
diffusion_max_step: int = 0,
|
| 198 |
+
channel_order: Optional[Sequence[str]] = None,
|
| 199 |
+
stats_order: Optional[Sequence[str]] = None,
|
| 200 |
+
stats_mu: Optional[Sequence[float]] = None,
|
| 201 |
+
stats_sd: Optional[Sequence[float]] = None):
|
| 202 |
+
super().__init__()
|
| 203 |
+
self.save_hyperparameters(ignore=["edge_index_base", "channel_order", "stats_order", "stats_mu", "stats_sd"])
|
| 204 |
+
self.net = UnifiedGraphSpatioVerticalGATNetTP(
|
| 205 |
+
C=C, Tin=Tin, Tout=Tout,
|
| 206 |
+
edge_index_base=edge_index_base, N_mesh=N_mesh,
|
| 207 |
+
embed_dim=embed_dim, blocks=blocks, heads_v=heads_v, heads_xy=heads_xy,
|
| 208 |
+
attn_drop_xy=attn_drop_xy,
|
| 209 |
+
channel_order=channel_order,
|
| 210 |
+
)
|
| 211 |
+
self.lr = lr
|
| 212 |
+
self.weight_decay = weight_decay
|
| 213 |
+
self.diffusion_max_step = int(diffusion_max_step)
|
| 214 |
+
|
| 215 |
+
self.crit = SparsePrecipLoss(
|
| 216 |
+
huber_delta=huber_delta,
|
| 217 |
+
wet_thr_mm=wet_thr_mm,
|
| 218 |
+
wet_weight=wet_weight,
|
| 219 |
+
dry_weight=dry_weight,
|
| 220 |
+
alpha_mm=alpha_mm,
|
| 221 |
+
gamma_focal=gamma_focal,
|
| 222 |
+
)
|
| 223 |
+
|
| 224 |
+
# для другого терміну лоссу потрібні ROI-статистики
|
| 225 |
+
if (stats_order is None) or (stats_mu is None) or (stats_sd is None):
|
| 226 |
+
self.denorm_helper = None
|
| 227 |
+
else:
|
| 228 |
+
# збережемо в torch.Tensors для зручності на GPU
|
| 229 |
+
self.register_buffer("_stats_mu", torch.tensor(stats_mu, dtype=torch.float32))
|
| 230 |
+
self.register_buffer("_stats_sd", torch.tensor(stats_sd, dtype=torch.float32))
|
| 231 |
+
self.stats_order = list(stats_order)
|
| 232 |
+
self.denorm_helper = TPDenormHelper(self.stats_order, self._stats_mu, self._stats_sd)
|
| 233 |
+
|
| 234 |
+
# ----------- shared step -----------
|
| 235 |
+
def _shared_step(self, batch, stage: str):
|
| 236 |
+
(x_all, geo, pos2d), y_full = batch # y_full: [B, Tout, C, N]
|
| 237 |
+
B = x_all.size(0)
|
| 238 |
+
# дифузійний крок як у загальній
|
| 239 |
+
steps = torch.zeros((B,), device=x_all.device, dtype=torch.float32)
|
| 240 |
+
if self.diffusion_max_step > 0:
|
| 241 |
+
steps = torch.randint(0, self.diffusion_max_step, (B,), device=x_all.device, dtype=torch.float32)
|
| 242 |
+
|
| 243 |
+
# канал-ордер беремо з DM
|
| 244 |
+
order = getattr(self.trainer.datamodule, "channel_order", None)
|
| 245 |
+
if order is None:
|
| 246 |
+
raise RuntimeError("DataModule не заповнив channel_order.")
|
| 247 |
+
# прогноз [B,Tout,1,N]
|
| 248 |
+
y_hat = self.net((x_all, geo, pos2d), diffusion_step=steps, channel_order=order)
|
| 249 |
+
y_true_tp = y_full[:, :, self.net.idx_tp].unsqueeze(2) # [B,Tout,1,N]
|
| 250 |
+
|
| 251 |
+
y_pred = y_hat[:, 0, 0] # [B,N] нормалізований tp_log
|
| 252 |
+
y_true = y_true_tp[:, 0, 0]
|
| 253 |
+
|
| 254 |
+
if self.denorm_helper is None:
|
| 255 |
+
loss = nn.SmoothL1Loss(beta=0.5)(y_pred, y_true) # fallback, якщо нема стат
|
| 256 |
+
else:
|
| 257 |
+
loss = self.crit(y_pred, y_true, self.denorm_helper)
|
| 258 |
+
|
| 259 |
+
self.log_dict({f"{stage}/loss": loss}, prog_bar=True, on_step=(stage == "train"), on_epoch=True)
|
| 260 |
+
return loss
|
| 261 |
+
|
| 262 |
+
def training_step(self, batch, batch_idx):
|
| 263 |
+
return self._shared_step(batch, "train")
|
| 264 |
+
|
| 265 |
+
def validation_step(self, batch, batch_idx):
|
| 266 |
+
self._shared_step(batch, "val")
|
| 267 |
+
|
| 268 |
+
# ----------- optim -----------
|
| 269 |
+
def configure_optimizers(self):
|
| 270 |
+
opt = torch.optim.AdamW(self.parameters(), lr=self.lr, weight_decay=self.weight_decay)
|
| 271 |
+
sch = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=max(self.trainer.max_epochs, 1))
|
| 272 |
+
return {"optimizer": opt, "lr_scheduler": {"scheduler": sch, "interval": "epoch"}}
|
unified_graph_weather_model_uv10.py
ADDED
|
@@ -0,0 +1,229 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# model_unified_graph_uv10.py
|
| 2 |
+
import math
|
| 3 |
+
from typing import Optional, Tuple, Sequence
|
| 4 |
+
|
| 5 |
+
import torch
|
| 6 |
+
import torch.nn as nn
|
| 7 |
+
import pytorch_lightning as pl
|
| 8 |
+
|
| 9 |
+
# ВАЖЛИВО: імпортуємо вашу загальну модель
|
| 10 |
+
from model.unified_graph_weather_model import UnifiedGraphSpatioVerticalGATNet
|
| 11 |
+
|
| 12 |
+
|
| 13 |
+
# ------------------------- Допоміжний векторний лосс -------------------------
|
| 14 |
+
class WindVectorLoss(nn.Module):
|
| 15 |
+
"""
|
| 16 |
+
Лосс для вітру: компоненти + (опційно) модуль і напрям.
|
| 17 |
+
y_pred, y_true: [B, 2, N] (u, v) у НОРМАЛІЗОВАНОМУ масштабі (як у тренінгу).
|
| 18 |
+
Зазвичай достатньо component MSE. Якщо хочеться трохи "фізики", вмикаємо модуль/кут.
|
| 19 |
+
"""
|
| 20 |
+
def __init__(self,
|
| 21 |
+
w_components: float = 1.0,
|
| 22 |
+
w_magnitude: float = 0.0,
|
| 23 |
+
w_direction: float = 0.0,
|
| 24 |
+
robust: bool = False,
|
| 25 |
+
eps: float = 1e-6):
|
| 26 |
+
super().__init__()
|
| 27 |
+
self.wc = float(w_components)
|
| 28 |
+
self.wm = float(w_magnitude)
|
| 29 |
+
self.wd = float(w_direction)
|
| 30 |
+
self.eps = float(eps)
|
| 31 |
+
self.comp_loss = nn.SmoothL1Loss(reduction="mean") if robust else nn.MSELoss(reduction="mean")
|
| 32 |
+
self.mag_loss = nn.SmoothL1Loss(reduction="mean") if robust else nn.MSELoss(reduction="mean")
|
| 33 |
+
|
| 34 |
+
def forward(self, y_pred: torch.Tensor, y_true: torch.Tensor) -> torch.Tensor:
|
| 35 |
+
# компоненти
|
| 36 |
+
Lc = self.comp_loss(y_pred, y_true)
|
| 37 |
+
|
| 38 |
+
if (self.wm == 0.0) and (self.wd == 0.0):
|
| 39 |
+
return self.wc * Lc
|
| 40 |
+
|
| 41 |
+
# модуль
|
| 42 |
+
mag_p = torch.sqrt(y_pred[:, 0]**2 + y_pred[:, 1]**2 + self.eps) # [B, N]
|
| 43 |
+
mag_t = torch.sqrt(y_true[:, 0]**2 + y_true[:, 1]**2 + self.eps)
|
| 44 |
+
Lm = self.mag_loss(mag_p, mag_t)
|
| 45 |
+
|
| 46 |
+
# напрям (1 - cos(Δθ)) через косинусну схожість
|
| 47 |
+
dot = (y_pred[:, 0]*y_true[:, 0] + y_pred[:, 1]*y_true[:, 1]) # [B, N]
|
| 48 |
+
cos_sim = dot / (mag_p * mag_t + self.eps)
|
| 49 |
+
# clamp щоб уникнути NaN при рідкісних виродженнях
|
| 50 |
+
cos_sim = torch.clamp(cos_sim, -1.0, 1.0)
|
| 51 |
+
Ld = (1.0 - cos_sim).mean()
|
| 52 |
+
|
| 53 |
+
return self.wc * Lc + self.wm * Lm + self.wd * Ld
|
| 54 |
+
|
| 55 |
+
|
| 56 |
+
# ----------------- UV-адаптер над загальною Graph GAT моделлю ----------------
|
| 57 |
+
class UnifiedGraphSpatioVerticalGATNetUV(nn.Module):
|
| 58 |
+
"""
|
| 59 |
+
Тонкий адаптер: всередині — Ваша загальна мережа.
|
| 60 |
+
На виході повертаємо лише 2 канали (u10, v10).
|
| 61 |
+
Обчислювальні витрати майже ті самі (всі внутрішні блоки працюють як і раніше),
|
| 62 |
+
економія — лише на невеликій «голові». Це компроміс заради сумісності з чекпойнтом.
|
| 63 |
+
|
| 64 |
+
Параметри:
|
| 65 |
+
- chan_u_name, chan_v_name: як канали називаються у shard'ах (звично "u10@sfc", "v10@sfc")
|
| 66 |
+
- channel_order: список з шарду для пошуку індексів
|
| 67 |
+
"""
|
| 68 |
+
def __init__(self,
|
| 69 |
+
C: int, Tin: int, Tout: int,
|
| 70 |
+
edge_index_base: torch.Tensor,
|
| 71 |
+
N_mesh: Optional[int] = None,
|
| 72 |
+
embed_dim: int = 192, blocks: int = 4, heads_v: int = 4, heads_xy: int = 4,
|
| 73 |
+
attn_drop_xy: float = 0.0,
|
| 74 |
+
chan_u_name: str = "u10@sfc",
|
| 75 |
+
chan_v_name: str = "v10@sfc",
|
| 76 |
+
channel_order: Optional[Sequence[str]] = None):
|
| 77 |
+
super().__init__()
|
| 78 |
+
self.chan_u_name = chan_u_name
|
| 79 |
+
self.chan_v_name = chan_v_name
|
| 80 |
+
self.channel_order = list(channel_order) if channel_order is not None else None
|
| 81 |
+
|
| 82 |
+
# базова мережа як у загальній моделі (C_out = C)
|
| 83 |
+
self.base = UnifiedGraphSpatioVerticalGATNet(
|
| 84 |
+
C=C, Tin=Tin, Tout=Tout,
|
| 85 |
+
edge_index_base=edge_index_base, N_mesh=N_mesh,
|
| 86 |
+
embed_dim=embed_dim, blocks=blocks, heads_v=heads_v, heads_xy=heads_xy,
|
| 87 |
+
attn_drop_xy=attn_drop_xy,
|
| 88 |
+
)
|
| 89 |
+
|
| 90 |
+
# індекси каналів у повному виході
|
| 91 |
+
if self.channel_order is not None:
|
| 92 |
+
self.idx_u, self.idx_v = self._find_uv_indices(self.channel_order)
|
| 93 |
+
else:
|
| 94 |
+
# якщо не дали під час ініт — довизначимо при першому forward (по наданому channel_order)
|
| 95 |
+
self.idx_u, self.idx_v = None, None
|
| 96 |
+
|
| 97 |
+
@staticmethod
|
| 98 |
+
def _find_uv_indices(order: Sequence[str],
|
| 99 |
+
u_candidates=("u10@sfc", "u10"),
|
| 100 |
+
v_candidates=("v10@sfc", "v10")) -> Tuple[int, int]:
|
| 101 |
+
lower = [str(x).lower() for x in order]
|
| 102 |
+
iu, iv = None, None
|
| 103 |
+
for cand in u_candidates:
|
| 104 |
+
if cand.lower() in lower:
|
| 105 |
+
iu = lower.index(cand.lower()); break
|
| 106 |
+
for cand in v_candidates:
|
| 107 |
+
if cand.lower() in lower:
|
| 108 |
+
iv = lower.index(cand.lower()); break
|
| 109 |
+
if iu is None or iv is None:
|
| 110 |
+
raise KeyError(f"Не знайшов індекси u/v у channel_order. order приклад: {order[:5]} ...")
|
| 111 |
+
return int(iu), int(iv)
|
| 112 |
+
|
| 113 |
+
def set_channel_order(self, order: Sequence[str]):
|
| 114 |
+
self.channel_order = list(order)
|
| 115 |
+
self.idx_u, self.idx_v = self._find_uv_indices(self.channel_order)
|
| 116 |
+
|
| 117 |
+
@torch.no_grad()
|
| 118 |
+
def _lazy_indices_from_order(self, order: Sequence[str]):
|
| 119 |
+
if (self.idx_u is None) or (self.idx_v is None):
|
| 120 |
+
self.set_channel_order(order)
|
| 121 |
+
|
| 122 |
+
def forward(self, x_tuple, diffusion_step: Optional[torch.Tensor] = None,
|
| 123 |
+
channel_order: Optional[Sequence[str]] = None):
|
| 124 |
+
"""
|
| 125 |
+
x_tuple: (x_all, geo, pos2d)
|
| 126 |
+
x_all: [B, Tin, C, N]
|
| 127 |
+
geo: [B, 4, N]
|
| 128 |
+
pos2d: [B, 4, N]
|
| 129 |
+
return: [B, Tout, 2, N] (u, v)
|
| 130 |
+
"""
|
| 131 |
+
if (self.idx_u is None or self.idx_v is None) and (channel_order is not None):
|
| 132 |
+
self._lazy_indices_from_order(channel_order)
|
| 133 |
+
|
| 134 |
+
y_full = self.base(x_tuple, diffusion_step=diffusion_step) # [B, Tout, C, N]
|
| 135 |
+
if self.idx_u is None or self.idx_v is None:
|
| 136 |
+
raise RuntimeError("UV індекси не визначені (передайте channel_order у forward або set_channel_order()).")
|
| 137 |
+
uv = torch.stack([y_full[:, :, self.idx_u], y_full[:, :, self.idx_v]], dim=2) # [B, Tout, 2, N]
|
| 138 |
+
return uv
|
| 139 |
+
|
| 140 |
+
|
| 141 |
+
# ----------------------- Lightning: UV-файнтюнінг ----------------------------
|
| 142 |
+
class WindUV10Lightning(pl.LightningModule):
|
| 143 |
+
"""
|
| 144 |
+
Lightning-клас для файнтюнінгу під (u10, v10).
|
| 145 |
+
- Повертає лише 2-канальний вихід
|
| 146 |
+
- Лосс: компоненти + (опц.) модуль/напрям
|
| 147 |
+
- Може підвантажувати «знання» з загальної моделі (ckpt)
|
| 148 |
+
"""
|
| 149 |
+
def __init__(self,
|
| 150 |
+
C: int, Tin: int, Tout: int,
|
| 151 |
+
edge_index_base: torch.Tensor,
|
| 152 |
+
N_mesh: Optional[int] = None,
|
| 153 |
+
embed_dim: int = 192, blocks: int = 4, heads_v: int = 4, heads_xy: int = 4,
|
| 154 |
+
attn_drop_xy: float = 0.0,
|
| 155 |
+
lr: float = 3e-4, weight_decay: float = 1e-4,
|
| 156 |
+
loss_components: float = 1.0,
|
| 157 |
+
loss_magnitude: float = 0.0,
|
| 158 |
+
loss_direction: float = 0.0,
|
| 159 |
+
robust_loss: bool = False,
|
| 160 |
+
chan_u_name: str = "u10@sfc",
|
| 161 |
+
chan_v_name: str = "v10@sfc",
|
| 162 |
+
channel_order: Optional[Sequence[str]] = None,
|
| 163 |
+
diffusion_max_step: int = 100):
|
| 164 |
+
super().__init__()
|
| 165 |
+
self.save_hyperparameters(ignore=["edge_index_base", "channel_order"])
|
| 166 |
+
self.net = UnifiedGraphSpatioVerticalGATNetUV(
|
| 167 |
+
C=C, Tin=Tin, Tout=Tout,
|
| 168 |
+
edge_index_base=edge_index_base, N_mesh=N_mesh,
|
| 169 |
+
embed_dim=embed_dim, blocks=blocks, heads_v=heads_v, heads_xy=heads_xy,
|
| 170 |
+
attn_drop_xy=attn_drop_xy,
|
| 171 |
+
chan_u_name=chan_u_name, chan_v_name=chan_v_name,
|
| 172 |
+
channel_order=channel_order,
|
| 173 |
+
)
|
| 174 |
+
self.lr = lr
|
| 175 |
+
self.weight_decay = weight_decay
|
| 176 |
+
self.diffusion_max_step = int(diffusion_max_step)
|
| 177 |
+
self.criterion = WindVectorLoss(
|
| 178 |
+
w_components=loss_components,
|
| 179 |
+
w_magnitude=loss_magnitude,
|
| 180 |
+
w_direction=loss_direction,
|
| 181 |
+
robust=robust_loss,
|
| 182 |
+
)
|
| 183 |
+
|
| 184 |
+
# -------- helper: вирізати GT (u,v) із повного y ----------
|
| 185 |
+
def _select_uv_from_full(self, y_full: torch.Tensor, order: Sequence[str]) -> torch.Tensor:
|
| 186 |
+
"""
|
| 187 |
+
y_full: [B, Tout, C, N] -> [B, Tout, 2, N]
|
| 188 |
+
"""
|
| 189 |
+
self.net._lazy_indices_from_order(order)
|
| 190 |
+
iu, iv = self.net.idx_u, self.net.idx_v
|
| 191 |
+
return torch.stack([y_full[:, :, iu], y_full[:, :, iv]], dim=2)
|
| 192 |
+
|
| 193 |
+
# ----------------------- train/val step -----------------------
|
| 194 |
+
def _shared_step(self, batch, stage: str):
|
| 195 |
+
(x_all, geo, pos2d), y_full = batch # y_full: [B, Tout, C, N]
|
| 196 |
+
# дифузійний крок (як у загальній моделі)
|
| 197 |
+
steps = torch.randint(
|
| 198 |
+
low=0, high=max(self.diffusion_max_step, 1),
|
| 199 |
+
size=(x_all.size(0),), device=x_all.device, dtype=torch.float32
|
| 200 |
+
)
|
| 201 |
+
# forward -> [B, Tout, 2, N]
|
| 202 |
+
y_hat = self.net((x_all, geo, pos2d), diffusion_step=steps,
|
| 203 |
+
channel_order=getattr(self.trainer.datamodule, "channel_order", None))
|
| 204 |
+
|
| 205 |
+
# GT тільки UV
|
| 206 |
+
order = getattr(self.trainer.datamodule, "channel_order", None)
|
| 207 |
+
if order is None:
|
| 208 |
+
raise RuntimeError("DataModule не надав channel_order.")
|
| 209 |
+
y_uv = self._select_uv_from_full(y_full, order) # [B, Tout, 2, N]
|
| 210 |
+
|
| 211 |
+
# lead-1 (Tout=1)
|
| 212 |
+
y_pred = y_hat[:, 0] # [B, 2, N]
|
| 213 |
+
y_true = y_uv[:, 0] # [B, 2, N]
|
| 214 |
+
|
| 215 |
+
loss = self.criterion(y_pred, y_true)
|
| 216 |
+
self.log_dict({f"{stage}/loss": loss}, prog_bar=True, on_step=(stage=="train"), on_epoch=True)
|
| 217 |
+
return loss
|
| 218 |
+
|
| 219 |
+
def training_step(self, batch, batch_idx):
|
| 220 |
+
return self._shared_step(batch, "train")
|
| 221 |
+
|
| 222 |
+
def validation_step(self, batch, batch_idx):
|
| 223 |
+
self._shared_step(batch, "val")
|
| 224 |
+
|
| 225 |
+
# ----------------------- optim -----------------------
|
| 226 |
+
def configure_optimizers(self):
|
| 227 |
+
opt = torch.optim.AdamW(self.parameters(), lr=self.lr, weight_decay=self.weight_decay)
|
| 228 |
+
sch = torch.optim.lr_scheduler.CosineAnnealingLR(opt, T_max=max(self.trainer.max_epochs, 1))
|
| 229 |
+
return {"optimizer": opt, "lr_scheduler": {"scheduler": sch, "interval": "epoch"}}
|