File size: 9,174 Bytes
ff0fadf | 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 |
import torch
import math
import torch.nn as nn
import numpy as np
import torch.nn.functional as F
from onescience.modules.mlp.MLP import StandardMLP
from onescience.modules.transformer.Neural_Spectral_Block import (
NeuralSpectralBlock1D,
NeuralSpectralBlock2D,
NeuralSpectralBlock3D,
)
from onescience.modules.embedding import timestep_embedding, unified_pos_embedding
from onescience.modules.fourier.geo_spectral import GeoSpectralConv2d, IPHI
from onescience.modules.layer.unet_layer import (
DoubleConv1D, Down1D, Up1D, OutConv1D,
DoubleConv2D, Down2D, Up2D, OutConv2D,
DoubleConv3D, Down3D, Up3D, OutConv3D,
)
ConvList = [None, DoubleConv1D, DoubleConv2D, DoubleConv3D]
DownList = [None, Down1D, Down2D, Down3D]
UpList = [None, Up1D, Up2D, Up3D]
OutList = [None, OutConv1D, OutConv2D, OutConv3D]
NeuralSpectralBlockList = [
None,
NeuralSpectralBlock1D,
NeuralSpectralBlock2D,
NeuralSpectralBlock3D,
]
class Model(nn.Module):
"""
LSM (Latent Spectral Model) 模型。
结合了 U-Net 的多尺度特征提取与 Neural Spectral Block (Latent Transformer) 的全局谱处理能力。
"""
def __init__(
self, args, device, bilinear=True, num_token=4, num_basis=12, s1=96, s2=96
):
super(Model, self).__init__()
self.__name__ = "LSM"
self.args = args
if args.task == "steady":
normtype = "bn"
else:
normtype = "in"
# 1. Embedding & Preprocessing
if args.unified_pos and args.geotype != "unstructured":
self.pos = unified_pos_embedding(args.shapelist, args.ref, device=device)
in_dim = args.fun_dim + args.ref ** len(args.shapelist)
else:
in_dim = args.fun_dim + args.space_dim
self.preprocess = StandardMLP(
input_dim=in_dim,
output_dim=args.n_hidden,
hidden_dims=[args.n_hidden * 2],
activation=args.act,
use_bias=True
)
if args.time_input:
self.time_fc = nn.Sequential(
nn.Linear(args.n_hidden, args.n_hidden),
nn.SiLU(),
nn.Linear(args.n_hidden, args.n_hidden),
)
else:
self.time_fc = None
# 2. Geometry Projection
if self.args.geotype == "unstructured":
self.fftproject_in = GeoSpectralConv2d(
in_channels=args.n_hidden,
out_channels=args.n_hidden,
modes1=args.modes,
modes2=args.modes,
s1=s1,
s2=s2
)
self.fftproject_out = GeoSpectralConv2d(
in_channels=args.n_hidden,
out_channels=args.n_hidden,
modes1=args.modes,
modes2=args.modes,
s1=s1,
s2=s2
)
self.iphi = IPHI()
patch_size = [(size + (16 - size % 16) % 16) // 16 for size in [s1, s2]]
self.padding = [(16 - size % 16) % 16 for size in [s1, s2]]
else:
patch_size = [
(size + (16 - size % 16) % 16) // 16 for size in args.shapelist
]
self.padding = [(16 - size % 16) % 16 for size in args.shapelist]
# 3. Multiscale U-Net Modules
# -----------------------------------------------------------
dim = len(patch_size)
self.inc = ConvList[dim](args.n_hidden, args.n_hidden, normtype=normtype)
self.down1 = DownList[dim](args.n_hidden, args.n_hidden * 2, normtype=normtype)
self.down2 = DownList[dim](args.n_hidden * 2, args.n_hidden * 4, normtype=normtype)
self.down3 = DownList[dim](args.n_hidden * 4, args.n_hidden * 8, normtype=normtype)
factor = 2 if bilinear else 1
self.down4 = DownList[dim](args.n_hidden * 8, args.n_hidden * 16 // factor, normtype=normtype)
self.up1 = UpList[dim](args.n_hidden * 16, args.n_hidden * 8 // factor, bilinear, normtype=normtype)
self.up2 = UpList[dim](args.n_hidden * 8, args.n_hidden * 4 // factor, bilinear, normtype=normtype)
self.up3 = UpList[dim](args.n_hidden * 4, args.n_hidden * 2 // factor, bilinear, normtype=normtype)
self.up4 = UpList[dim](args.n_hidden * 2, args.n_hidden, bilinear, normtype=normtype)
self.outc = OutList[dim](args.n_hidden, args.n_hidden)
# 4. Patchified Neural Spectral Blocks
block_class = NeuralSpectralBlockList[dim]
self.process1 = block_class(
width=args.n_hidden,
num_basis=num_basis,
patch_size=patch_size,
num_token=num_token,
n_heads=args.n_heads
)
self.process2 = block_class(
width=args.n_hidden * 2,
num_basis=num_basis,
patch_size=patch_size,
num_token=num_token,
n_heads=args.n_heads
)
self.process3 = block_class(
width=args.n_hidden * 4,
num_basis=num_basis,
patch_size=patch_size,
num_token=num_token,
n_heads=args.n_heads
)
self.process4 = block_class(
width=args.n_hidden * 8,
num_basis=num_basis,
patch_size=patch_size,
num_token=num_token,
n_heads=args.n_heads
)
self.process5 = block_class(
width=args.n_hidden * 16 // factor,
num_basis=num_basis,
patch_size=patch_size,
num_token=num_token,
n_heads=args.n_heads
)
# 5. Projectors
# -----------------------------------------------------------
self.fc1 = nn.Linear(args.n_hidden, args.n_hidden * 2)
self.fc2 = nn.Linear(args.n_hidden * 2, args.out_dim)
def structured_geo(self, x, fx, T=None):
B, N, _ = x.shape
if self.args.unified_pos:
x = self.pos.repeat(x.shape[0], 1, 1)
if fx is not None:
fx = torch.cat((x, fx), -1)
fx = self.preprocess(fx)
else:
fx = self.preprocess(x)
if T is not None:
Time_emb = timestep_embedding(T, self.args.n_hidden)
Time_emb = self.time_fc(Time_emb)
if Time_emb.ndim == 2:
Time_emb = Time_emb.unsqueeze(1)
fx = fx + Time_emb
x = fx.permute(0, 2, 1).reshape(B, self.args.n_hidden, *self.args.shapelist)
if not all(item == 0 for item in self.padding):
pad_arg = []
for p in reversed(self.padding):
pad_arg.extend([0, p])
x = F.pad(x, pad_arg)
# LSM 核心处理流
x1 = self.inc(x)
x2 = self.down1(x1)
x3 = self.down2(x2)
x4 = self.down3(x3)
x5 = self.down4(x4)
x = self.up1(self.process5(x5), self.process4(x4))
x = self.up2(x, self.process3(x3))
x = self.up3(x, self.process2(x2))
x = self.up4(x, self.process1(x1))
x = self.outc(x)
if not all(item == 0 for item in self.padding):
if len(self.args.shapelist) == 1:
x = x[..., : -self.padding[0]]
elif len(self.args.shapelist) == 2:
x = x[..., : -self.padding[0], : -self.padding[1]]
elif len(self.args.shapelist) == 3:
x = x[..., : -self.padding[0], : -self.padding[1], : -self.padding[2]]
x = x.reshape(B, self.args.n_hidden, -1).permute(0, 2, 1)
x = self.fc1(x)
x = F.gelu(x)
x = self.fc2(x)
return x
def unstructured_geo(self, x, fx, T=None):
original_pos = x
if fx is not None:
fx = torch.cat((x, fx), -1)
fx = self.preprocess(fx)
else:
fx = self.preprocess(x)
if T is not None:
Time_emb = timestep_embedding(T, self.args.n_hidden)
Time_emb = self.time_fc(Time_emb)
if Time_emb.ndim == 2:
Time_emb = Time_emb.unsqueeze(1)
fx = fx + Time_emb
x = self.fftproject_in(
fx.permute(0, 2, 1), x_in=original_pos, iphi=self.iphi, code=None
)
# LSM 核心处理流
x1 = self.inc(x)
x2 = self.down1(x1)
x3 = self.down2(x2)
x4 = self.down3(x3)
x5 = self.down4(x4)
x = self.up1(self.process5(x5), self.process4(x4))
x = self.up2(x, self.process3(x3))
x = self.up3(x, self.process2(x2))
x = self.up4(x, self.process1(x1))
x = self.outc(x)
x = self.fftproject_out(
x, x_out=original_pos, iphi=self.iphi, code=None
).permute(0, 2, 1)
x = self.fc1(x)
x = F.gelu(x)
x = self.fc2(x)
return x
def forward(self, x, fx, T=None, geo=None):
if self.args.geotype == "unstructured":
return self.unstructured_geo(x, fx, T)
else:
return self.structured_geo(x, fx, T)
|