-
Notifications
You must be signed in to change notification settings - Fork 1
/
run_train_pdpd.py
323 lines (256 loc) · 15 KB
/
run_train_pdpd.py
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
import h5py
import numpy as np
import paddle
import paddle.nn as nn
from process_data_pdpd import data_norm, data_sampler
from basic_model_pdpd import gradients, DeepModel_single, DeepModel_multi
import visual_data
import matplotlib.pyplot as plt
import time
import os
import argparse
import sys
os.environ["KMP_DUPLICATE_LIB_OK"] = "TRUE"
def get_args():
parser = argparse.ArgumentParser('PINNs for naiver-stokes cylinder with Karman Vortex', add_help=False)
parser.add_argument('-f', type=str, default="external parameters")
parser.add_argument('--Layer_depth', default=5, type=int, help="Number of Layers depth")
parser.add_argument('--Layer_width', default=64, type=int, help="Number of Layers width")
parser.add_argument('--in_norm', default=True, type=bool, help="input feature normalization")
parser.add_argument('--out_norm', default=True, type=bool, help="output fields normalization")
parser.add_argument('--activation', default=nn.Tanh(), help="activation function")
parser.add_argument('--points_name', default="60+8", type=str, help="distribution of supervised points")
parser.add_argument('--Net_pattern', default='single', type=str, help="single or multi networks")
parser.add_argument('--epochs_adam', default=400000, type=int, help='total training epoch')
parser.add_argument('--save_freq', default=5000, type=int, help="frequency to save model and image")
parser.add_argument('--print_freq', default=1000, type=int, help="frequency to print loss")
parser.add_argument('--device', default=0, type=int, help="gpu id")
parser.add_argument('--work_name', default='', type=str, help="work path to save files")
parser.add_argument('--load_pretrained', default=False, type=bool, help="load pretrained model or not")
parser.add_argument('--Nx_EQs', default=30000, type=int, help="xy sampling in for equation loss")
parser.add_argument('--Nt_EQs', default=2, type=int, help="time sampling in for equation loss")
parser.add_argument('--Nt_BCs', default=120, type=int, help="time sampling in for boundary loss")
return parser.parse_args()
def read_data():
data = h5py.File('./data/cyl_Re250.mat', 'r')
nodes = np.array(data['grids_']).squeeze().transpose((3, 2, 1, 0)) # [Nx, Ny, Nf]
field = np.array(data['fields_']).squeeze().transpose((3, 2, 1, 0)) # [Nt, Nx, Ny, Nf]
times = np.array(data['dynamics_']).squeeze().transpose((1, 0))[3::4, (0,)] # (800, 3) -> (200, 1)
nodes = nodes[0]
times = times - times[0, 0]
return times[:120], nodes[:, :, 1:], field[:120, :, :, :] # Nx / 2
def BCS_ICS(nodes, points):
BCS = []
ICS = []
Num_Nodes = nodes.shape[0] * nodes.shape[1]
Index = np.arange(Num_Nodes).reshape((nodes.shape[0], nodes.shape[1]))
BCS.append(np.concatenate((Index[:93, -1], Index[284:, -1]), axis=0)) #in
BCS.append(Index[93:284, -1]) #out
BCS.append(Index[:, 0]) #wall
######监督测点生成
point = np.concatenate((Index[175:220:15, 0:80:4].reshape(-1, 1)), axis=0)
BCS.append(np.concatenate((Index[::48, 1], point), axis=0)) # 60+8
ICS.append(Index.reshape(-1))
INN = np.setdiff1d(ICS[0], np.concatenate(BCS[:-1], axis=0)) #其余点
return INN, BCS, ICS
class Net_single(DeepModel_single):
def __init__(self, planes, data_norm, active):
super(Net_single, self).__init__(planes, data_norm, active)
self.Re = 250.
def equation(self, inn_var, out_var):
# a = grad(psi.sum(), in_var, create_graph=True, retain_graph=True)[0]
p, u, v = out_var[:, 0:1], out_var[:, 1:2], out_var[:, 2:3]
duda = gradients(u, inn_var)
dudx, dudy, dudt = duda[:, 0:1], duda[:, 1:2], duda[:, 2:3]
dvda = gradients(v, inn_var)
dvdx, dvdy, dvdt = dvda[:, 0:1], dvda[:, 1:2], dvda[:, 2:3]
d2udx2 = gradients(dudx, inn_var)[:, 0:1]
d2udy2 = gradients(dudy, inn_var)[:, 1:2]
d2vdx2 = gradients(dvdx, inn_var)[:, 0:1]
d2vdy2 = gradients(dvdy, inn_var)[:, 1:2]
dpda = gradients(p, inn_var)
dpdx, dpdy = dpda[:, 0:1], dpda[:, 1:2]
eq1 = dudt + (u * dudx + v * dudy) + dpdx - 1 / self.Re * (d2udx2 + d2udy2)
eq2 = dvdt + (u * dvdx + v * dvdy) + dpdy - 1 / self.Re * (d2vdx2 + d2vdy2)
eq3 = dudx + dvdy
eqs = paddle.concat((eq1, eq2, eq3), axis=1)
return eqs
class Net_multi(DeepModel_multi):
def __init__(self, planes, data_norm, active):
super(Net_multi, self).__init__(planes, data_norm, active)
self.Re = 250.
def equation(self, inn_var, out_var):
# a = grad(psi.sum(), in_var, create_graph=True, retain_graph=True)[0]
p, u, v = out_var[:, 0:1], out_var[:, 1:2], out_var[:, 2:3]
duda = gradients(u, inn_var)
dudx, dudy, dudt = duda[:, 0:1], duda[:, 1:2], duda[:, 2:3]
dvda = gradients(v, inn_var)
dvdx, dvdy, dvdt = dvda[:, 0:1], dvda[:, 1:2], dvda[:, 2:3]
d2udx2 = gradients(dudx, inn_var)[:, 0:1]
d2udy2 = gradients(dudy, inn_var)[:, 1:2]
d2vdx2 = gradients(dvdx, inn_var)[:, 0:1]
d2vdy2 = gradients(dvdy, inn_var)[:, 1:2]
dpda = gradients(p, inn_var)
dpdx, dpdy = dpda[:, 0:1], dpda[:, 1:2]
eq1 = dudt + (u * dudx + v * dudy) + dpdx - 1 / self.Re * (d2udx2 + d2udy2)
eq2 = dvdt + (u * dvdx + v * dvdy) + dpdy - 1 / self.Re * (d2vdx2 + d2vdy2)
eq3 = dudx + dvdy
eqs = paddle.concat((eq1, eq2, eq3), axis=1)
return eqs
def train(inn_var, BCs, ICs, out_true, model, Loss, optimizer, scheduler, log_loss, opts):
inn = BCs[0].sampling(Nx=opts.Nx_EQs, Nt=opts.Nt_EQs) #随机抽取守恒损失计算点
BC_in = BCs[1].sampling(Nx='all', Nt=opts.Nt_BCs); ind_BC_in = BC_in.shape[0] #入口
BC_out = BCs[2].sampling(Nx='all', Nt=opts.Nt_BCs); ind_BC_out = BC_out.shape[0] + ind_BC_in #出口
BC_wall = BCs[3].sampling(Nx='all', Nt=opts.Nt_BCs); ind_BC_wall = BC_wall.shape[0] + ind_BC_out #圆柱
BC_meas = BCs[4].sampling(Nx='all', Nt=opts.Nt_BCs); ind_BC_meas = BC_meas.shape[0] + ind_BC_wall
IC_0 = ICs[0].sampling(Nx='all') #初始场
inn_BCs = paddle.concat((inn_var[BC_in], inn_var[BC_out], inn_var[BC_wall], inn_var[BC_meas], inn_var[IC_0]), axis=0)
out_BCs = paddle.concat((out_true[BC_in], out_true[BC_out], out_true[BC_wall], out_true[BC_meas], out_true[IC_0]), axis=0)
inn_EQs = inn_var[inn]
out_EQs = out_true[inn]
optimizer.clear_grad()
inn_EQs.stop_gradient = False
out_EQs_ = model(inn_EQs, in_norm=opts.in_norm, out_norm=opts.out_norm)
res_EQs = model.equation(inn_EQs, out_EQs_)
inn_BCs.stop_gradient = True
out_BCs_ = model(inn_BCs, in_norm=opts.in_norm, out_norm=opts.out_norm)
bcs_loss_1 = Loss(out_BCs_[:ind_BC_in, 1:], out_BCs[:ind_BC_in, 1:]) #进口速度
bcs_loss_2 = Loss(out_BCs_[ind_BC_in:ind_BC_out, 0], out_BCs[ind_BC_in:ind_BC_out, 0]) #出口压力
bcs_loss_3 = Loss(out_BCs_[ind_BC_out:ind_BC_wall, 1:], out_BCs[ind_BC_out:ind_BC_wall, 1:]) #壁面速度
bcs_loss_4 = Loss(out_BCs_[ind_BC_wall:ind_BC_meas, :], out_BCs[ind_BC_wall:ind_BC_meas, :]) #监督测点
ics_loss_0 = Loss(out_BCs_[ind_BC_meas:, :], out_BCs[ind_BC_meas:, :]) #初始条件损失
eqs_loss = (res_EQs**2).mean() #方程损失
loss_batch = bcs_loss_1 + bcs_loss_2 + bcs_loss_3 + bcs_loss_4 + ics_loss_0 + eqs_loss
loss_batch.backward()
data_loss = Loss(out_EQs_, out_EQs) #全部点的data loss 没有用来训练
log_loss.append([eqs_loss.item(), bcs_loss_1.item(), bcs_loss_2.item(), bcs_loss_3.item(), bcs_loss_4.item(),
ics_loss_0.item(), data_loss.item()])
# return loss_batch
# optimizer.step(closure)
optimizer.step()
scheduler.step()
def inference(inn_var, model, opts):
with paddle.no_grad():
out_pred = model(inn_var, in_norm=opts.in_norm, out_norm=opts.out_norm)
return out_pred
if __name__ == '__main__':
opts = get_args()
opts.load_pretrained = True
print(opts)
if paddle.fluid.is_compiled_with_cuda():
paddle.set_device("gpu:" + str(opts.device)) # 指定第一块gpu
device = "gpu:" + str(opts.device)
else:
paddle.set_device('cpu')
device = 'cpu'
points_name = opts.points_name
work_name = 'NS-cylinder-2d-t_pdpd_' + points_name + '-' + opts.work_name
work_path = os.path.join('work', work_name)
tran_path = os.path.join('work', work_name, 'train')
isCreated = os.path.exists(tran_path)
if not isCreated:
os.makedirs(tran_path)
# 将控制台的结果输出到a.log文件，可以改成a.txt
sys.stdout = visual_data.Logger(os.path.join(work_path, 'train.log'), sys.stdout)
times, nodes, field = read_data()
INN, BCS, ICS = BCS_ICS(nodes, points_name)
Nt, Nx, Ny, Nf = field.shape[0], field.shape[1], field.shape[2], field.shape[3]
times = np.tile(times[:, None, None, :], (1, Nx, Ny, 1))
nodes = np.tile(nodes[None, :, :, :], (Nt, 1, 1, 1))
times = times.reshape(-1, 1)
nodes = nodes.reshape(-1, 2)
field = field.reshape(-1, Nf)
input = np.concatenate((nodes, times), axis=-1)
input_norm = data_norm(input, method='mean-std')
field_norm = data_norm(field, method='mean-std')
input_visual = input.reshape((Nt, Nx, Ny, 3))
add_input = input_visual[:, 0, :, :].reshape((Nt, -1, Ny, 3))
input_visual = np.concatenate((input_visual, add_input), axis=1)
field_visual = field.reshape((Nt, Nx, Ny, Nf))
add_field = field_visual[:, 0, :, :].reshape((Nt, -1, Ny, Nf))
field_visual = np.concatenate((field_visual, add_field), axis=1)
# Training Data
input = paddle.to_tensor(input, dtype='float32')
field = paddle.to_tensor(field, dtype='float32')
NumNodes = Nx * Ny
BC_in = data_sampler(BCS[0], NumNodes, time=Nt)
BC_out = data_sampler(BCS[1], NumNodes, time=Nt)
BC_cyl = data_sampler(BCS[2], NumNodes, time=Nt)
BC_meas = data_sampler(BCS[3], NumNodes, time=Nt)
IC_cyl = data_sampler(ICS[0], NumNodes, time=0)
IN_cyl = data_sampler(INN, NumNodes, time=Nt)
BCs = [IN_cyl, BC_in, BC_out, BC_cyl, BC_meas]
ICs = [IC_cyl,]
L1Loss = nn.L1Loss()
HBLoss = nn.SmoothL1Loss()
L2Loss = nn.MSELoss()
planes = [3,] + [opts.Layer_width] * opts.Layer_depth + [3,]
if opts.Net_pattern == "single":
Net_model = Net_single(planes=planes, data_norm=(input_norm, field_norm), active=opts.activation).to(device)
elif opts.Net_pattern == "multi":
Net_model = Net_multi(planes=planes, data_norm=(input_norm, field_norm), active=opts.activation).to(device)
Boundary_epoch1 = [opts.epochs_adam*8/10, opts.epochs_adam*9/10]
Boundary_epoch2 = [opts.epochs_adam*11/10, opts.epochs_adam*12/10]
Scheduler1 = paddle.optimizer.lr.MultiStepDecay(learning_rate=0.001, milestones=Boundary_epoch1, gamma=0.1)
Optimizer1 = paddle.optimizer.Adam(parameters=Net_model.parameters(), learning_rate=Scheduler1, beta1=0.8,
beta2=0.9)
# Optimizer2 = paddle.incubate.optimizer.functional.minimize_lbfgs(parameters=Net_model.parameters(), learning_rate=.1, max_iter=100)
# Scheduler2 = paddle.optimizer.lr.MultiStepDecay(Optimizer2, milestones=Boundary_epoch2, gamma=0.1)
Visual = visual_data.matplotlib_vision('/', field_name=('p', 'u', 'v'), input_name=('x', 'y'))
Visual.font['size'] = 20
star_time = time.time()
start_epoch=0
log_loss=[]
"""load a pre-trained model"""
if opts.load_pretrained:
print('load pretrained model')
start_epoch, log_loss = Net_model.loadmodel(os.path.join(work_path, 'pretrained_model.pth'))
else:
start_epoch, log_loss = Net_model.loadmodel(os.path.join(work_path, 'latest_model.pth'))
for i in range(start_epoch):
# update the learning rate for start_epoch times
Scheduler1.step()
# Training
for iter in range(start_epoch, opts.epochs_adam):
if iter < opts.epochs_adam:
train(input, BCs, ICs, field, Net_model, L2Loss, Optimizer1, Scheduler1, log_loss, opts)
learning_rate = Optimizer1.get_lr()
if iter > 0 and iter % opts.print_freq == 0:
print('iter: {:6d}/{:6d}, lr: {:.1e}, cost: {:.2f}, dat_loss: {:.2e} \n'
'eqs_loss: {:.2e}, BCS_loss_in: {:.2e}, BCS_loss_out: {:.2e}, '
'BCS_loss_wall: {:.2e}, BCS_loss_meas: {:.2e}, ICS_loss_0: {:.2e}'.
format(iter, opts.epochs_adam, learning_rate, time.time() - star_time, log_loss[-1][-1],
log_loss[-1][0], log_loss[-1][1], log_loss[-1][2],
log_loss[-1][3], log_loss[-1][4], log_loss[-1][5]))
plt.figure(1, figsize=(20, 15))
plt.clf()
plt.subplot(2, 1, 1)
Visual.plot_loss(np.arange(len(log_loss)), np.array(log_loss)[:, -1], 'dat_loss')
Visual.plot_loss(np.arange(len(log_loss)), np.array(log_loss)[:, 0], 'eqs_loss')
Visual.plot_loss(np.arange(len(log_loss)), np.array(log_loss)[:, 5], 'ICS_loss_0')
plt.subplot(2, 1, 2)
Visual.plot_loss(np.arange(len(log_loss)), np.array(log_loss)[:, 1], 'BCS_loss_in')
Visual.plot_loss(np.arange(len(log_loss)), np.array(log_loss)[:, 2], 'BCS_loss_out')
Visual.plot_loss(np.arange(len(log_loss)), np.array(log_loss)[:, 3], 'BCS_loss_wall')
Visual.plot_loss(np.arange(len(log_loss)), np.array(log_loss)[:, 4], 'BCS_loss_meas')
plt.savefig(os.path.join(tran_path, 'log_loss.svg'))
star_time = time.time()
if iter > 0 and iter % opts.save_freq == 0:
input_visual_p = paddle.to_tensor(input_visual[:100:10], dtype='float32', place='gpu:0')
field_visual_p = inference(input_visual_p, Net_model, opts)
field_visual_t = field_visual[:100:10]
field_visual_p = field_visual_p.cpu().numpy()
for t in range(field_visual_p.shape[0]):
plt.figure(2, figsize=(20, 10))
plt.clf()
Visual.plot_fields_ms(field_visual_t[t], field_visual_p[t], input_visual_p[0, :, :, :2].numpy(),
cmin_max=[[-4, -4], [10, 4]], field_name=['p', 'u', 'v'])
plt.subplots_adjust(wspace=0.2, hspace=0.3) # left=0.05, bottom=0.05, right=0.95, top=0.95
plt.savefig(os.path.join(tran_path, 'loca_' + str(t) + '.jpg'))
plt.figure(3, figsize=(30, 20))
plt.clf()
Visual.plot_fields_ms(field_visual_t[t], field_visual_p[t], input_visual_p[0, :, :, :2].numpy())
plt.subplots_adjust(wspace=0.2, hspace=0.3)
plt.savefig(os.path.join(tran_path, 'full_' + str(t) + '.jpg'))
paddle.save({'epoch': iter, 'model': Net_model.state_dict(), 'log_loss': log_loss},
os.path.join(work_path, 'latest_model.pth'))