File size: 17,328 Bytes
6c3f19f | 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 | import torch
import math
import numpy as np
from gpytorch.settings import cholesky_jitter
import matplotlib.pyplot as plt
import torch.optim
import matplotlib.pyplot as plt
from tqdm import tqdm
from onescience.utils.GP_TO import plot_predictions_and_residuals,plot_loss_history,plot_density_and_velocity_fields
plt.rcParams['font.family'] = 'DejaVu Sans'
plt.rcParams['font.size'] = 15
plt.rcParams['figure.dpi'] = 150
checkpoints=[1000,10000,20000,30000,40000,50000]
lambdaa = 0.1
# Determine volume loss based on title
gamma_values = {
'rugby': 0.9,
'pipebend': 0.08 * torch.pi,
'doublepipe': 1 / 3,
'diffuser': 0.5
}
def modified_sigmoid(x, alpha=12.0):
"""Compute the modified sigmoid function."""
return 1 / (1 + torch.exp(-alpha * (x - 0.5)))
def infer_square_grid(num_points):
side = int(math.sqrt(num_points))
if side * side != num_points:
raise ValueError(
"Numerical differentiation expects a square collocation grid. "
f"Got {num_points} points."
)
return side, side
def share_mean_module(model_list):
model_ref = model_list[0]
for model in model_list:
model.mean_module_NN_All = model_ref.mean_module_NN_All
return model_ref
def clear_cached_kernels(model_list):
for model in model_list:
model.k_xX = None
model.chol_decomp = None
def predict_fields(model_list):
model_ref = share_mean_module(model_list)
collocation_x = model_ref.collocation_x
m_col = model_ref.mean_module_NN_All(collocation_x)
if model_ref.k_xX is None:
for model in model_list:
model.k_xX = model.covar_module(model.train_inputs[0], collocation_x).evaluate()
if model_ref.chol_decomp is None:
with cholesky_jitter(1e-7):
for model in model_list:
model.chol_decomp = model.covar_module(model.train_inputs[0]).cholesky()
offsets = []
for i, model in enumerate(model_list):
target_offset = model.train_targets.unsqueeze(-1) - model.mean_module_NN_All(model.train_inputs[0])[:, i].unsqueeze(-1)
offsets.append(model.chol_decomp._cholesky_solve(target_offset))
u = (m_col[:, 0].unsqueeze(-1) + model_list[0].k_xX.t() @ offsets[0]).squeeze(-1)
v = (m_col[:, 1].unsqueeze(-1) + model_list[1].k_xX.t() @ offsets[1]).squeeze(-1)
p = (m_col[:, 2].unsqueeze(-1) + model_list[2].k_xX.t() @ offsets[2]).squeeze(-1)
ro_tensor = (m_col[:, 3].unsqueeze(-1) + model_list[3].k_xX.t() @ offsets[3]).squeeze(-1)
ro = modified_sigmoid(ro_tensor)
return {"x": collocation_x, "u": u, "v": v, "p": p, "ro": ro}
def compute_dynamic_weights_ic(pde_loss, bc_loss, ic_loss, model):
"""
Compute dynamic weights for loss functions based on gradients.
"""
params_to_update = [param for param in model.parameters() if param.requires_grad]
def compute_gradients(loss, scaling_factor):
gradients = torch.autograd.grad(scaling_factor * loss, params_to_update, retain_graph=True, allow_unused=True)
values = [p.reshape(-1).cpu().tolist() for p in gradients if p is not None]
return torch.abs(torch.tensor([v for val in values for v in val]))
delta_pde = compute_gradients(pde_loss, 1.0)
delta_bc = compute_gradients(bc_loss, model.alpha)
delta_ic = compute_gradients(ic_loss, model.beta)
temp_bc = torch.max(delta_pde) / torch.mean(delta_bc)
temp_ic = torch.max(delta_pde) / torch.mean(delta_ic)
return (
(1.0 - lambdaa) * model.alpha + lambdaa * temp_bc,
(1.0 - lambdaa) * model.beta + lambdaa * temp_ic
)
def compute_fvm_residuals_higher_order(u, v, p, f_x, f_y, nu, dx, dy):
# Check the device of the input tensors and move other tensors to the same device
device = u.device
# Extend the fields with ghost cells to handle boundary conditions
u_ext = torch.zeros((u.shape[0] + 4, u.shape[1] + 4), device=device)
v_ext = torch.zeros((v.shape[0] + 4, v.shape[1] + 4), device=device)
p_ext = torch.zeros((p.shape[0] + 4, p.shape[1] + 4), device=device)
# Copy the interior
u_ext[2:-2, 2:-2] = u
v_ext[2:-2, 2:-2] = v
p_ext[2:-2, 2:-2] = p
# Apply ghost cells for no-slip boundary conditions
u_ext[:2, :] = u_ext[2:4, :]
u_ext[-2:, :] = u_ext[-4:-2, :]
u_ext[:, :2] = u_ext[:, 2:4]
u_ext[:, -2:] = u_ext[:, -4:-2]
v_ext[:2, :] = v_ext[2:4, :]
v_ext[-2:, :] = v_ext[-4:-2, :]
v_ext[:, :2] = v_ext[:, 2:4]
v_ext[:, -2:] = v_ext[:, -4:-2]
# Laplacian of u and v (fourth-order central differences for second derivatives)
u_xx = (-u_ext[4:, 2:-2] + 16 * u_ext[3:-1, 2:-2] - 30 * u_ext[2:-2, 2:-2] + 16 * u_ext[1:-3, 2:-2] - u_ext[:-4, 2:-2]) / (12 * dx**2)
u_yy = (-u_ext[2:-2, 4:] + 16 * u_ext[2:-2, 3:-1] - 30 * u_ext[2:-2, 2:-2] + 16 * u_ext[2:-2, 1:-3] - u_ext[2:-2, :-4]) / (12 * dy**2)
laplacian_u = u_xx + u_yy
v_xx = (-v_ext[4:, 2:-2] + 16 * v_ext[3:-1, 2:-2] - 30 * v_ext[2:-2, 2:-2] + 16 * v_ext[1:-3, 2:-2] - v_ext[:-4, 2:-2]) / (12 * dx**2)
v_yy = (-v_ext[2:-2, 4:] + 16 * v_ext[2:-2, 3:-1] - 30 * v_ext[2:-2, 2:-2] + 16 * v_ext[2:-2, 1:-3] - v_ext[2:-2, :-4]) / (12 * dy**2)
laplacian_v = v_xx + v_yy
# Gradients of p using fourth-order central difference
grad_p_x = (-p_ext[4:, 2:-2] + 8 * p_ext[3:-1, 2:-2] - 8 * p_ext[1:-3, 2:-2] + p_ext[:-4, 2:-2]) / (12 * dx)
grad_p_y = (-p_ext[2:-2, 4:] + 8 * p_ext[2:-2, 3:-1] - 8 * p_ext[2:-2, 1:-3] + p_ext[2:-2, :-4]) / (12 * dy)
# Momentum equation residuals
residual_u = -nu * laplacian_u + grad_p_x + f_x
residual_v = -nu * laplacian_v + grad_p_y + f_y
# Mass conservation (divergence of velocity) using central differences
div_u = (u_ext[2:-2, 2:-2] - u_ext[1:-3, 2:-2]) / dx
div_v = (v_ext[2:-2, 2:-2] - v_ext[2:-2, 1:-3]) / dy
u_y = (u_ext[2:-2, 2:-2] - u_ext[2:-2, 1:-3]) / dy
v_x = (v_ext[2:-2, 2:-2] - v_ext[1:-3, 2:-2]) / dx
residual_mass = div_u + div_v
return residual_u.reshape(-1), residual_v.reshape(-1), residual_mass.reshape(-1), div_u.reshape(-1), div_v.reshape(-1), u_y.reshape(-1), v_x.reshape(-1)
def compute_autograd_derivatives(model, z_column, collocation_x, grad_order=1):
"""
Compute first and second-order derivatives using PyTorch autograd.
:param model: Model to train
:param z_column: Column of `z_all` for differentiation
:param collocation_x: Input coordinates
:param grad_order: Order of gradients (1 for first, 2 for second)
:return: First and second-order derivatives as tensors
"""
model.train()
grad_1 = torch.autograd.grad(z_column, collocation_x, torch.ones_like(z_column), create_graph=True)[0]
if grad_order > 1:
grad_2_x = torch.autograd.grad(grad_1[:, 0], collocation_x, torch.ones_like(grad_1[:, 0]), create_graph=True)[0][:, 0]
grad_2_y = torch.autograd.grad(grad_1[:, 1], collocation_x, torch.ones_like(grad_1[:, 1]), create_graph=True)[0][:, 1]
return grad_1[:, 0], grad_1[:, 1], grad_2_x, grad_2_y
return grad_1[:, 0], grad_1[:, 1], None, None
w_1,w_2,w_3, w_4, w_5 =0.01,0.01,1e2,1e5,1e4
def loss_volume(y, gamma=0.5):
"""
Compute the volume loss as the squared difference between the mean of y and gamma.
"""
mean_y = torch.mean(y)
return torch.square(mean_y - gamma)
def dissipated_power(u, u_x, v_y, u_y, v_x):
"""
Compute the total dissipated power based on the input velocity gradients and displacements.
"""
# First part of the dissipated power
p1 = (u_x**2 + v_y**2 + u_y**2 + v_x**2).sum(dim=1, keepdim=True)
# Second part of the dissipated power
u2 = (u[:, :2]**2).sum(dim=1, keepdim=True)
p2 = alpha(u[:, 3:]) * u2
return 0.5 * (p1 + p2)
def alpha(rho):
"""
Compute the alpha parameter based on rho.
"""
alpha_max = 2.5e4
alpha_min = 2.5e-4
q = 0.1
return alpha_max + (alpha_min - alpha_max) * rho * (1 + q) / (rho + q)
def calculate_loss_multioutput(
model_list,
iteration,
diff_method='Numerical',
title="default",
problem="doublepipe",
checkpoint_steps=None,
plot_outputs=True,
):
"""
Calculate loss for multi-output models based on PDE, BC, and IC residuals.
"""
# Clone collocation points
collocation_x = model_list[0].collocation_x.clone()
# Compute mean values at collocation points
m_col = model_list[0].mean_module_NN_All(collocation_x)
# Evaluate covariance matrices if not already done
if model_list[0].k_xX is None:
for i, model in enumerate(model_list):
model.k_xX = model.covar_module(model.train_inputs[0], collocation_x).evaluate()
# Perform Cholesky decomposition if not already done
if model_list[0].chol_decomp is None:
with cholesky_jitter(1e-7):
for model in model_list:
model.chol_decomp = model.covar_module(model.train_inputs[0]).cholesky()
# Solve for offsets using Cholesky decomposition
offsets = []
for i, model in enumerate(model_list):
target_offset = model.train_targets.unsqueeze(-1) - model.mean_module_NN_All(model.train_inputs[0])[:, i].unsqueeze(-1)
offsets.append(model.chol_decomp._cholesky_solve(target_offset))
# Compute values for velocity, pressure, and density
ro_tensor = (m_col[:, 3].unsqueeze(-1) + model_list[3].k_xX.t() @ offsets[3]).squeeze(-1)
ro = modified_sigmoid(ro_tensor)
u = (m_col[:, 0].unsqueeze(-1) + model_list[0].k_xX.t() @ offsets[0]).squeeze(-1)
v = (m_col[:, 1].unsqueeze(-1) + model_list[1].k_xX.t() @ offsets[1]).squeeze(-1)
p = (m_col[:, 2].unsqueeze(-1) + model_list[2].k_xX.t() @ offsets[2]).squeeze(-1)
z_all = torch.cat((u.unsqueeze(1), v.unsqueeze(1), p.unsqueeze(1), ro.unsqueeze(1)), dim=1)
f = alpha(z_all[:, 3:]) * z_all[:, :2]
fx, fy = f[:, :1], f[:, 1:]
# Compute residuals
if diff_method == 'Numerical':
nx, ny = infer_square_grid(collocation_x.shape[0])
dx, dy = 1.0 / nx, 1.0 / ny
u, v, p = [arr.reshape(nx, ny) for arr in (u, v, p)]
fx, fy = [arr.reshape(nx, ny) for arr in (fx, fy)]
residuals = compute_fvm_residuals_higher_order(u, v, p, fx, fy, nu=1, dx=dx, dy=dy)
residual_pde1, residual_pde2, residual_pde3, u_x, v_y, u_y, v_x = residuals
residual_pde1, residual_pde2, residual_pde3 = [res * w for res, w in zip(residuals[:3], (w_1, w_2, w_3))]
u, v, p = [arr.reshape(-1) for arr in (u, v, p)]
fx, fy = [arr.reshape(-1) for arr in (fx, fy)]
else:
u_x, u_y, u_xx, u_yy = compute_autograd_derivatives(model_list[0], z_all[:, 0], collocation_x, grad_order=2)
v_x, v_y, v_xx, v_yy = compute_autograd_derivatives(model_list[1], z_all[:, 1], collocation_x, grad_order=2)
p_x, p_y, _, _ = compute_autograd_derivatives(model_list[2], z_all[:, 2], collocation_x, grad_order=1)
residual_pde1 = (- (u_xx + u_yy) + p_x + fx[:, 0]) * w_1
residual_pde2 = (- (v_xx + v_yy) + p_y + fy[:, 0]) * w_2
residual_pde3 = (u_x + v_y) * w_3
# Plot fields at checkpoints
active_checkpoints = checkpoints if checkpoint_steps is None else checkpoint_steps
if plot_outputs and iteration in active_checkpoints:
plot_density_and_velocity_fields(u, v, p, ro, collocation_x, iteration, problem)
plot_predictions_and_residuals(u, v, p, ro, collocation_x, iteration, residual_pde1, residual_pde2, residual_pde3, 1, 1, 1, problem)
# Calculate losses
loss_pde1 = torch.mean(residual_pde1**2)
loss_pde2 = torch.mean(residual_pde2**2)
loss_pde3 = torch.mean(residual_pde3**2)
dp_loss = dissipated_power(z_all, u_x.reshape(-1, 1), v_y.reshape(-1, 1), u_y.reshape(-1, 1), v_x.reshape(-1, 1))
gamma = next((val for key, val in gamma_values.items() if key in title), 0.5)
vol_loss = loss_volume(z_all[:, 3:], gamma=gamma)* w_4
# Return final losses
return loss_pde1, loss_pde2, loss_pde3, torch.sum(dp_loss), vol_loss
def find_TO(
model_list,
lr_default: float = 0.01,
num_iter: int = 500,
title: str = 'default',
problem: str = 'doublepipe', # 设置默认值
diff_method: str = 'Numerical',
checkpoint_steps=None,
plot_outputs=True,
) -> float:
"""
Train models to minimize total loss using dynamic weights and record progress.
Args:
model_list: List of models to be trained.
lr_default: Default learning rate for the optimizer.
num_iter: Number of training iterations.
title: Title for specific settings.
diff_method: Differentiation method ('Numerical' or 'Autograd').
Returns:
loss history
"""
# Use the first model as a reference for shared NN
model_ref = share_mean_module(model_list)
for model in model_list:
model.train()
# Initialize variables
loss_total, loss_hist, loss_hist_total = [], [], []
loss_pde1_hist, loss_pde2_hist, loss_pde3_hist = [], [], []
loss_dp_hist, vol_loss_hist = [], []
scaled_loss_pde1_hist, scaled_loss_pde2_hist = [], []
scaled_loss_pde3_hist, scaled_loss_dp_hist, scaled_vol_loss_hist = [], [], []
weights = {'alpha': [], 'beta': [], 'mu_p': []}
result_loss_thetas = {'largest_eigval_hist': [], 'condition_number_hist': [], 'Hessian': [], 'grads': []}
GP_NN_hist, NN_hist, time_hist = [], [], []
dynamic_weights = True
sigma_1, beta_1, alph_1 = 10, 10, 0.1
f_inc, mu_F = math.inf, 1
# Initialize model-specific parameters
model1 = model_list[0]
model1.alpha, model1.beta, model1.theta = 1, 1, 1
# Set up optimizer and scheduler
optimizer = torch.optim.Adam(model_ref.parameters(), lr=lr_default)
milestones = sorted({int(x) for x in np.linspace(0, max(num_iter, 1), 4).tolist() if int(x) > 0})
scheduler = torch.optim.lr_scheduler.MultiStepLR(
optimizer,
milestones=milestones,
gamma=0.75
)
# Training loop
with tqdm(range(num_iter + 1), desc='Epoch', position=0, leave=True) as pbar:
for j in pbar:
optimizer.zero_grad()
# Calculate losses
loss_pde1, loss_pde2, loss_pde3, loss_dp, vol_loss = calculate_loss_multioutput(
model_list,
j,
diff_method=diff_method,
title=title,
problem=problem,
checkpoint_steps=checkpoint_steps,
plot_outputs=plot_outputs,
)
loss_pde = loss_pde1 + loss_pde3
# Dynamic weight adjustments
if dynamic_weights:
alpha, beta = compute_dynamic_weights_ic(
loss_pde1 + loss_pde2, loss_pde3, vol_loss, model1
)
if all(torch.is_tensor(val) and not (torch.isnan(val).any() or torch.isinf(val).any()) for val in [alpha, beta]):
model1.alpha, model1.beta = alpha, beta
weights['alpha'].append(alpha.detach().cpu().item())
weights['beta'].append(beta.detach().cpu().item())
weights['mu_p'].append(mu_F)
loss = loss_dp + mu_F * (
loss_pde1 + loss_pde2 + model1.alpha * loss_pde3 + model1.beta * vol_loss
)
if (j + 1) % 50 == 0:
mu_F = min(mu_F * 1.05, 5e2)
else:
alph_1 = 1e-6 / (j + 1)
beta_1, sigma_1 = 2 * (j + 1), 5 * (j + 1)
loss = sigma_1 * loss_pde + alph_1 * loss_dp + beta_1 * vol_loss
# Record loss history
loss_total.append(loss.item())
loss_pde1_hist.append((1/w_1) * loss_pde1.item())
loss_pde2_hist.append((1/w_2) * loss_pde2.item())
loss_pde3_hist.append((1/w_3) * loss_pde3.item())
loss_dp_hist.append((1/w_5) * loss_dp.item())
vol_loss_hist.append((1/w_4) * vol_loss.item())
scaled_loss_pde1_hist.append(loss_pde1.item())
scaled_loss_pde2_hist.append(loss_pde2.item())
scaled_loss_pde3_hist.append(model1.alpha * loss_pde3.item())
scaled_loss_dp_hist.append(loss_dp.item())
scaled_vol_loss_hist.append(model1.beta * vol_loss.item())
# Plot loss history at checkpoints
active_checkpoints = checkpoints if checkpoint_steps is None else checkpoint_steps
if plot_outputs and j in active_checkpoints:
plot_loss_history(
loss_total,
[loss_pde1_hist, scaled_loss_pde1_hist],
[loss_pde2_hist, scaled_loss_pde2_hist],
[loss_pde3_hist, scaled_loss_pde3_hist],
[loss_dp_hist, scaled_loss_dp_hist],
[vol_loss_hist, scaled_vol_loss_hist],
j,
problem
)
# Backpropagation and optimizer step
loss.backward(retain_graph=True)
optimizer.step()
scheduler.step()
# Update progress description
pbar.set_postfix(loss=f"{loss.item():.6f}")
loss_hist.append(loss.item())
loss_hist_total = loss_hist
return loss_hist_total
|