-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy pathFitClass.py
191 lines (145 loc) · 7.81 KB
/
FitClass.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
import copy
import torch
import torch.nn as nn
import numpy as np
def weight_reset(m):
if isinstance(m, nn.Linear):
m.reset_parameters()
def regularization(model, p):
reg_loss = 0
for name, param in model.named_parameters():
if 'weight' in name or 'bias' in name:
reg_loss = reg_loss + torch.norm(param, p)
return reg_loss
class CustomLoss(torch.nn.Module):
def __init__(self):
super(CustomLoss, self).__init__()
def forward(self, Ec, network_sol, network_test, x_u_train, u_train, x_b_train, u_b_train, x_f_train, epc, verbose=False, minimizing=True):
lambda_res = network_sol.lambda_residual
lambda_reg_sol = network_sol.regularization_param
lambda_reg_test = network_test.regularization_param
u_pred_var_list = list()
u_train_var_list = list()
Ec.apply_bc(network_sol, x_b_train, u_b_train, u_pred_var_list, u_train_var_list)
if x_u_train.shape[0] != 0:
Ec.apply_ic(network_sol, x_u_train, u_train, u_pred_var_list, u_train_var_list)
u_pred_tot_vars = torch.cat(u_pred_var_list, 0).to(Ec.device)
u_train_tot_vars = torch.cat(u_train_var_list, 0).to(Ec.device)
loss_vars = torch.mean(torch.abs(u_pred_tot_vars - u_train_tot_vars) ** Ec.p)
loss_reg_sol = regularization(network_sol, 2)
loss_reg_test = regularization(network_test, 2)
if minimizing:
if verbose:
print("############### MINIMIZING ###############")
loss_pde, loss_pde_no_norm = Ec.compute_res(network_sol, network_test, x_f_train, minimizing)
loss_v = lambda_res * loss_vars.to(Ec.device) + loss_pde.to(Ec.device) + lambda_reg_sol * loss_reg_sol.to(Ec.device) + lambda_reg_test * loss_reg_test.to(Ec.device)
if verbose:
print("###############################################################################################################")
print("Function Loss : ", (loss_vars ** (1 / Ec.p)).detach().cpu().numpy(),
"\nPDE Residual : ", (loss_pde_no_norm ** (1 / Ec.p)).detach().cpu().numpy())
print()
print()
return loss_v, loss_vars, loss_pde, loss_pde_no_norm
else:
if verbose:
print("############### MAXIMIZING ###############")
loss_pde, loss_pde_no_norm = Ec.compute_res(network_sol, network_test, x_f_train, minimizing)
loss_v = - torch.log(loss_pde)
return loss_v, 0, 0, [0, 0, 0], loss_pde_no_norm
def fit(Ec, solution_model, test_function_model, optimizer_min, optimizer_max, training_set_class, verbose=False):
num_epochs = solution_model.num_epochs
iterations_max = test_function_model.iterations
iterations_min = solution_model.iterations
reset_freq = int(solution_model.reset_freq * num_epochs)
best_losses = list([0, 0, 0, 0, 0])
freq = 1000
training_coll = training_set_class.data_coll
training_boundary = training_set_class.data_boundary
training_initial_internal = training_set_class.data_initial_internal
solution_model.train()
test_function_model.train()
if iterations_min != 0:
best_train = 1e+12
else:
best_train = 0
best_solution_model = None
best_test_function_model = None
lambda1 = lambda e: 1 / (1 + (e / num_epochs))
my_lr_scheduler_min = torch.optim.lr_scheduler.LambdaLR(optimizer=optimizer_min, lr_lambda=lambda1)
my_lr_scheduler_max = torch.optim.lr_scheduler.LambdaLR(optimizer=optimizer_max, lr_lambda=lambda1)
for epoch in range(num_epochs): # loop over the dataset multiple times
if epoch % reset_freq == 0 and epoch != 0:
print("Resetting Params")
test_function_model.apply(weight_reset)
current_losses = list([0, 0, 0, 0, 0])
def closure_max():
optimizer_max.zero_grad()
loss_test, _, _, _, res_pde_no_norm \
= CustomLoss().forward(Ec, solution_model, test_function_model, x_u_train_, u_train_, x_b_train_, u_b_train_, x_coll_train_, epoch, verbose, False)
current_losses[4] = current_losses[4] + float(res_pde_no_norm.cpu().detach().numpy())
loss_test.backward()
return loss_test
def closure_min():
optimizer_min.zero_grad()
loss_sol, loss_vars, loss_int, res_pde_no_norm = CustomLoss().forward(Ec, solution_model, test_function_model, x_u_train_, u_train_, x_b_train_, u_b_train_, x_coll_train_, epoch, verbose, True)
current_losses[0] = current_losses[0] + float(loss_sol.cpu().detach().numpy())
current_losses[1] = current_losses[1] + float(loss_vars.cpu().detach().numpy())
current_losses[2] = current_losses[2] + float(loss_int.cpu().detach().numpy())
current_losses[3] = current_losses[3] + float(res_pde_no_norm.cpu().detach().numpy())
loss_sol.backward()
return loss_sol
if epoch % freq == 0:
print("################################################## ", epoch, " ##################################################")
batch = 0
if len(training_boundary) != 0 and len(training_initial_internal) == 0:
for step, ((x_coll_train_, u_coll_train_), (x_b_train_, u_b_train_)) in enumerate(zip(training_coll, training_boundary)):
x_u_train_ = torch.full((0, 1), 0)
u_train_ = torch.full((0, 1), 0)
for _ in range(iterations_max):
optimizer_max.step(closure=closure_max)
for _ in range(iterations_min):
optimizer_min.step(closure=closure_min)
if torch.cuda.is_available():
del x_coll_train_
del x_b_train_
del u_b_train_
del x_u_train_
del u_train_
torch.cuda.empty_cache()
batch = batch + 1
if len(training_boundary) != 0 and len(training_initial_internal) != 0:
for step, ((x_coll_train_, u_coll_train_), (x_u_train_, u_train_), (x_b_train_, u_b_train_)) in enumerate(zip(training_coll, training_initial_internal, training_boundary)):
x_coll_train_ = x_coll_train_.to(Ec.device)
x_b_train_ = x_b_train_.to(Ec.device)
u_b_train_ = u_b_train_.to(Ec.device)
x_u_train_ = x_u_train_.to(Ec.device)
u_train_ = u_train_.to(Ec.device)
for _ in range(iterations_max):
optimizer_max.step(closure=closure_max)
for _ in range(iterations_min):
optimizer_min.step(closure=closure_min)
if torch.cuda.is_available():
del x_coll_train_
del x_b_train_
del u_b_train_
del x_u_train_
del u_train_
torch.cuda.empty_cache()
batch = batch + 1
for l in range(len(current_losses)):
current_losses[l] = current_losses[l] / batch
if np.isnan(current_losses[0]):
print("WARNING: Found NaN")
return best_losses, best_solution_model, best_test_function_model
if current_losses[0] < best_train:
best_solution_model = copy.deepcopy(solution_model)
# best_test_function_model = copy.deepcopy(test_function_model)
best_losses[0] = current_losses[0]
best_losses[1] = current_losses[1]
best_losses[2] = current_losses[2]
best_losses[3] = current_losses[3]
best_losses[4] = current_losses[4]
best_train = current_losses[0]
my_lr_scheduler_min.step()
my_lr_scheduler_max.step()
return best_losses, best_solution_model, best_test_function_model