-
Notifications
You must be signed in to change notification settings - Fork 3
/
train_student.py
183 lines (148 loc) · 8.37 KB
/
train_student.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
from torch.utils.data import DataLoader
import loader2 as lo
import torch
from torch import nn
import torch.optim as optim
import torch.nn.functional as F
from torch.optim.lr_scheduler import ExponentialLR
from tqdm import tqdm
import os
from evaluate_teacher import Evaluate_teacher
from config import *
from student_model import student_model as stu_model
from teacher_model import teacher_model as tea_model
from utils import sigmoid, maskedNLL, maskedMSE, MSELoss2, CELoss, distillation_loss
import warnings
import math
warnings.filterwarnings("ignore", category=UserWarning)
def main():
args['train_flag'] = True
l_path = args['path']
l_pre_path = args['pre_path']
name = args['name']
print(args['path'])
highwayNet_name = '/epoch' + name + '_g.pth'
highwayNet_path = os.path.join(l_pre_path, highwayNet_name)
log_var_fut = nn.Parameter(torch.zeros(1).to(device), requires_grad=True)
log_var_lat = nn.Parameter(torch.zeros(1).to(device), requires_grad=True)
log_var_lon = nn.Parameter(torch.zeros(1).to(device), requires_grad=True)
log_var = nn.Parameter(torch.zeros(1).to(device), requires_grad=True)
log_var_distill = nn.Parameter(torch.zeros(1).to(device), requires_grad=True)
evaluate_teacher = Evaluate_teacher()
gdEncoder_stu = stu_model.GDEncoder_stu(args)
generator_stu = stu_model.Generator_stu(args)
gdEncoder_stu = gdEncoder_stu.to(device)
generator_stu = generator_stu.to(device)
gdEncoder_stu.train()
generator_stu.train()
trSet = lo.ngsimDataset('./data/dataset_t_v_t/TrainSet.mat')
trDataloader = DataLoader(trSet, batch_size=args['batch_size'], shuffle=True, num_workers=8,
drop_last=True, persistent_workers=True, prefetch_factor=4,
collate_fn=trSet.collate_fn, pin_memory=True)
params_gdEncoder = list(gdEncoder_stu.parameters()) + [log_var, log_var_distill, log_var_fut, log_var_lat, log_var_lon]
params_generator = list(generator_stu.parameters()) + [log_var, log_var_distill, log_var_fut, log_var_lat, log_var_lon]
optimizer_gd = optim.Adam(params_gdEncoder, lr=learning_rate)
optimizer_g = optim.Adam(params_generator, lr=learning_rate)
scheduler_gd = ExponentialLR(optimizer_gd, gamma=0.6)
scheduler_g = ExponentialLR(optimizer_g, gamma=0.6)
for epoch in range(args['epoch']):
print("epoch:", epoch + 1, 'lr', optimizer_g.param_groups[0]['lr']) # 打印训练轮次信息
loss_gi1 = 0
loss_gix = 0
loss_gx_2i = 0
loss_gx_3i = 0
distil_loss = 0
distil_loss2 = 0
distil_loss3 = 0
for idx, data in enumerate(tqdm(trDataloader)):
hist_batch_stu, nbrs_batch_stu, lane_batch_stu, nbrslane_batch_stu, class_batch_stu, nbrsclass_batch_stu, va_batch_stu, nbrsva_batch_stu, fut_batch_stu, \
hist_batch, nbrs_batch, mask_batch, lat_enc_batch, lon_enc_batch, lane_batch, nbrslane_batch, class_batch, nbrsclass_batch, va_batch, nbrsva_batch, \
fut_batch, op_mask_batch, edge_index_batch, ve_matrix_batch, ac_matrix_batch, man_matrix_batch, view_grip_batch, graph_matrix = data
hist_batch = hist_batch.to(device)
nbrs_batch = nbrs_batch.to(device)
mask_batch = mask_batch.to(device)
lat_enc_batch = lat_enc_batch.to(device)
lon_enc_batch = lon_enc_batch.to(device)
lane_batch = lane_batch.to(device)
nbrslane_batch = nbrslane_batch.to(device)
class_batch = class_batch.to(device)
nbrsclass_batch = nbrsclass_batch.to(device)
fut_batch = fut_batch.to(device)
op_mask_batch = op_mask_batch.to(device)
va_batch = va_batch.to(device)
nbrsva_batch = nbrsva_batch.to(device)
hist_batch_stu = hist_batch_stu.to(device)
nbrs_batch_stu = nbrs_batch_stu.to(device)
lane_batch_stu = lane_batch_stu.to(device)
nbrslane_batch_stu = nbrslane_batch_stu.to(device)
class_batch_stu = class_batch_stu.to(device)
nbrsclass_batch_stu = nbrsclass_batch_stu.to(device)
fut_batch_stu = fut_batch_stu.to(device)
va_batch_stu = va_batch_stu.to(device)
nbrsva_batch_stu = nbrsva_batch_stu.to(device)
edge_index_batch = edge_index_batch.to(device)
ve_matrix_batch = ve_matrix_batch.to(device)
ac_matrix_batch = ac_matrix_batch.to(device)
man_matrix_batch = man_matrix_batch.to(device)
view_grip_batch = view_grip_batch.to(device)
graph_matrix = graph_matrix.to(device)
fut_pred_tea, lat_pred_tea, lon_pred_tea = evaluate_teacher.main(args['name'], hist_batch, nbrs_batch, mask_batch, lat_enc_batch, lon_enc_batch, lane_batch, nbrslane_batch, class_batch, nbrsclass_batch, va_batch, nbrsva_batch, edge_index_batch, ve_matrix_batch, ac_matrix_batch, man_matrix_batch, view_grip_batch, graph_matrix)
optimizer_g.zero_grad()
optimizer_gd.zero_grad()
values = gdEncoder_stu(hist_batch_stu, nbrs_batch_stu, mask_batch, va_batch_stu, nbrsva_batch_stu, lane_batch_stu, nbrslane_batch_stu, class_batch_stu, nbrsclass_batch_stu)
g_out, lat_pred, lon_pred = generator_stu(values, lat_enc_batch, lon_enc_batch)
precision = torch.exp(-log_var)
precision_distill = torch.exp(-log_var_distill)
precision_fut = torch.exp(-log_var_fut)
precision_lon = torch.exp(-log_var_lon)
precision_lat = torch.exp(-log_var_lat)
precision = precision.to(device)
precision_fut = precision_fut.to(device)
precision_lat = precision_lat.to(device)
precision_lon = precision_lon.to(device)
precision_distill = precision_distill.to(device)
if args['use_mse']:
loss_g1 = precision*MSELoss2(g_out, fut_batch_stu, op_mask_batch) + precision_distill*distillation_loss(g_out, fut_pred_tea)
else:
if epoch < args['pre_epoch']:
loss_g1 = precision*MSELoss2(g_out, fut_batch_stu, op_mask_batch) + precision_distill*distillation_loss(g_out, fut_pred_tea)
else:
loss_g1 = precision*maskedNLL(g_out, fut_batch_stu, op_mask_batch) + precision_distill*distillation_loss(g_out, fut_pred_tea)
loss_gx_3 = precision*CELoss(lat_pred, lat_enc_batch) + precision_distill*distillation_loss(lat_pred, lat_pred_tea)
loss_gx_2 = precision*CELoss(lon_pred, lon_enc_batch) + precision_distill*distillation_loss(lon_pred, lon_pred_tea)
loss_gx = precision_lat*loss_gx_3 + precision_lon*loss_gx_2
loss_g = precision_fut*loss_g1 + loss_gx + log_var + log_var_distill + log_var_fut + log_var_lat + log_var_lon
loss_g.backward()
a = t.nn.utils.clip_grad_norm_(generator_stu.parameters(), 10)
a = t.nn.utils.clip_grad_norm_(gdEncoder_stu.parameters(), 10)
optimizer_g.step()
optimizer_gd.step()
loss_gi1 += loss_g1.item()
loss_gx_2i += loss_gx_2.item()
loss_gx_3i += loss_gx_3.item()
loss_gix += loss_gx.item()
distil_loss += distillation_loss(g_out, fut_pred_tea).item()
distil_loss2 += distillation_loss(lat_pred, lat_pred_tea).item()
distil_loss3 += distillation_loss(lon_pred, lon_pred_tea).item()
if idx % 4000 == 3999:
print('mse:', loss_gi1 / 4000, '|loss_gx_2:', loss_gx_2i / 4000, '|loss_gx_3', loss_gx_3i / 4000)
print('distil_loss:', distil_loss / 4000, 'distil_loss2:', distil_loss2 / 4000, 'distil_loss3', distil_loss3 / 4000)
loss_gi1 = 0
loss_gix = 0
loss_gx_2i = 0
loss_gx_3i = 0
distil_loss = 0
distil_loss2 = 0
distil_loss3 = 0
save_model(name=str(epoch + 1), gdEncoder=gdEncoder_stu,
generator=generator_stu, path=args['path'])
scheduler_gd.step()
scheduler_g.step()
def save_model(name, gdEncoder, generator, path):
l_path = args['path']
if not os.path.exists(l_path):
os.makedirs(l_path)
t.save(gdEncoder.state_dict(), l_path + '/epoch' + name + '_gd.pth')
t.save(generator.state_dict(), l_path + '/epoch' + name + '_g.pth')
if __name__ == '__main__':
main()