-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy path2
207 lines (183 loc) · 8.94 KB
/
2
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
import os
import logging
import torch
import torch.optim as optim
from torch.optim.lr_scheduler import StepLR, MultiStepLR
import torch.backends.cudnn as cudnn
from networks.pspnet import Res_pspnet, BasicBlock, Bottleneck
from networks.sagan_models import Discriminator
from utils.criterion import CriterionDSN, CriterionKD, CriterionAdv, CriterionAdvForG, CriterionAdditionalGP, CriterionIFV, CriterionPairWiseforWholeFeatAfterPool
def load_S_model(args, model):
logging.info("------------")
if args.is_student_load_imgnet:
if os.path.isfile(args.student_pretrain_model_imgnet):
saved_state_dict = torch.load(args.student_pretrain_model_imgnet)
new_params=model.state_dict()
saved_state_dict = {k: v for k, v in saved_state_dict.items() if k in new_params}
new_params.update(saved_state_dict)
model.load_state_dict(new_params)
logging.info("=> load" + str(args.student_pretrain_model_imgnet))
else:
logging.info("=> the pretrain model on imgnet '{}' does not exit".format(args.student_pretrain_model_imgnet))
if args.S_resume:
if os.path.isfile(args.S_ckpt_path):
checkpoint = torch.load(args.S_ckpt_path)
if 'state_dict' in checkpoint:
args.last_step = checkpoint['step'] if 'step' in checkpoint else None
model.load_state_dict(checkpoint['state_dict'])
else:
model.load_state_dict(checkpoint)
logging.info("=> loaded checkpoint '{}' \n (step:{} \n )".format(args.S_ckpt_path, args.last_step))
else:
logging.info("=> student checkpoint '{}' does not exit".format(args.S_ckpt_path))
logging.info("------------")
def load_T_model(args, model):
logging.info("------------")
if os.path.isfile(args.T_ckpt_path):
model.load_state_dict(torch.load(args.T_ckpt_path))
logging.info("=> load" + str(args.T_ckpt_path))
else:
logging.info("=> teacher checkpoint '{}' does not exit".format(args.T_ckpt_path))
logging.info("------------")
def load_D_model(args, model):
logging.info("------------")
if args.D_resume:
if os.path.isfile(args.D_ckpt_path):
checkpoint = torch.load(args.D_ckpt_path)
if 'state_dict' in checkpoint:
args.last_step = checkpoint['step'] if 'step' in checkpoint else None
model.load_state_dict(checkpoint['state_dict'])
else:
model.load_state_dict(checkpoint)
logging.info("=> loaded checkpoint '{}' \n (step:{} \n )".format(args.D_ckpt_path, args.last_step))
else:
logging.info("=> checkpoint '{}' does not exit".format(args.D_ckpt_path))
else:
logging.info("=> train d from scratch")
logging.info("------------")
def print_model_parm_nums(model, string):
b = []
for param in model.parameters():
b.append(param.numel())
logging.info(string + ': Number of params: %.2fM', sum(b) / 1e6)
def to_tuple_str(str_first, gpu_num, str_ind):
if gpu_num > 1:
tmp = '('
for cpu_ind in range(gpu_num):
tmp += '(' + str_first + '[' + str(cpu_ind) + ']' + str_ind +',)'
if cpu_ind != gpu_num-1: tmp += ', '
tmp += ')'
else:
tmp = str_first + str_ind
return tmp
class NetModel():
def name(self):
return 'kd_seg'
def __init__(self, args):
self.args = args
student = Res_pspnet(BasicBlock, [2, 2, 2, 2], num_classes = args.num_classes)
load_S_model(args, student)
print_model_parm_nums(student, 'student_model')
student.cuda()
self.student = student
teacher = Res_pspnet(Bottleneck, [3, 4, 23, 3], num_classes = args.num_classes)
load_T_model(args, teacher)
print_model_parm_nums(teacher, 'teacher_model')
teacher.cuda()
self.teacher = teacher
D_model = Discriminator(args.preprocess_GAN_mode, args.num_classes, args.batch_size, args.imsize_for_adv, args.adv_conv_dim)
load_D_model(args, D_model)
print_model_parm_nums(D_model, 'D_model')
logging.info("------------")
D_model.cuda()
self.D_model = D_model
self.G_solver = optim.SGD([{'params': filter(lambda p: p.requires_grad, student.parameters()), 'initial_lr': args.lr_g}], args.lr_g, momentum=args.momentum, weight_decay=args.weight_decay)
self.D_solver = optim.Adam(filter(lambda p: p.requires_grad, D_model.parameters()), args.lr_d, [0.9, 0.99])
# self.D_solver = optim.SGD([{'params': filter(lambda p: p.requires_grad, D_model.parameters()), 'initial_lr': args.lr_d}], args.lr_d, momentum=args.momentum, weight_decay=args.weight_decay)
self.criterion_dsn = CriterionDSN().cuda()
if args.kd:
self.criterion_kd = CriterionKD().cuda()
if args.adv:
self.criterion_adv = CriterionAdv(args.adv_loss_type).cuda()
if args.adv_loss_type == 'wgan-gp': self.criterion_AdditionalGP = CriterionAdditionalGP(D_model, args.lambda_gp).cuda()
self.criterion_adv_for_G = CriterionAdvForG(args.adv_loss_type).cuda()
if args.ifv:
self.criterion_ifv = CriterionIFV(classes=args.num_classes).cuda()
if args.pa:
self.criterion_pa = CriterionPairWiseforWholeFeatAfterPool(args.pool_scale, -1).cuda()
self.G_loss, self.D_loss = 0.0, 0.0
self.mc_G_loss, self.kd_G_loss, self.adv_G_loss, self.ifv_G_loss = 0.0, 0.0, 0.0, 0.0
self.pa_G_loss = 0.0
cudnn.deterministic = True
cudnn.benchmark = False
def set_input(self, data):
images, labels, _, _ = data
self.images = images.cuda()
self.labels = labels.long().cuda()
def lr_poly(self, base_lr, iter, max_iter, power):
return base_lr*((1-float(iter)/max_iter)**(power))
def adjust_learning_rate(self, base_lr, optimizer, i_iter):
args = self.args
lr = self.lr_poly(base_lr, i_iter, args.num_steps, args.power)
optimizer.param_groups[0]['lr'] = lr
return lr
def segmentation_forward(self):
with torch.no_grad():
self.preds_T = self.teacher.eval()(self.images)
self.preds_S = self.student.train()(self.images)
def segmentation_backward(self):
args = self.args
temp = self.criterion_dsn(self.preds_S, self.labels)
self.mc_G_loss = temp.item()
g_loss = temp
if args.kd:
temp = args.lambda_kd*self.criterion_kd(self.preds_S, self.preds_T)
self.kd_G_loss = temp.item()
g_loss = g_loss + temp
if args.adv:
temp = args.lambda_adv*self.criterion_adv_for_G(self.D_model(self.preds_S[0]))
self.adv_G_loss = temp.item()
g_loss = g_loss + temp
if args.ifv:
temp = args.lambda_ifv*self.criterion_ifv(self.preds_S, self.preds_T, self.labels)
self.ifv_G_loss = temp.item()
g_loss = g_loss + temp
if args.pa:
temp = args.lambda_pa*self.criterion_pa(self.preds_S, self.preds_T)
self.pa_G_loss = temp.item()
g_loss = g_loss + temp
g_loss.backward()
self.G_loss = g_loss.item()
def discriminator_forward_backward(self):
args = self.args
d_loss = args.lambda_d*self.criterion_adv(self.D_model(self.preds_S[0].detach()), self.D_model(self.preds_T[0].detach()))
if args.adv_loss_type == 'wgan-gp': wgan_gp = args.lambda_d*self.criterion_AdditionalGP(self.preds_S, self.preds_T)
d_loss += wgan_gp
wgan_gp.backward()
for k,v in self.D_model.named_parameters():
print(k,v.grad)
raise ValueError
self.D_loss = d_loss.item()
def optimize_parameters(self):
self.segmentation_forward()
self.G_solver.zero_grad()
self.segmentation_backward()
self.G_solver.step()
if self.args.adv:
self.D_solver.zero_grad()
self.discriminator_forward_backward()
self.D_solver.step()
def print_info(self, step):
logging.info('step:{:5d} G_lr:{:.6f} G_loss:{:.5f}(mc:{:.5f} kd:{:.5f} adv:{:.5f} ifv:{:5f} pa:{:5f}) D_lr:{:.6f} D_loss:{:.5f}'.format(
step, self.G_solver.param_groups[-1]['lr'], self.G_loss,
self.mc_G_loss, self.kd_G_loss, self.adv_G_loss, self.ifv_G_loss, self.pa_G_loss,
self.D_solver.param_groups[-1]['lr'], self.D_loss))
def save_ckpt(self, step):
args = self.args
logging.info('saving ckpt: '+args.save_path+'/'+args.data_set+'_'+str(step)+'_G.pth')
torch.save(self.student.state_dict(), args.save_path+'/'+args.data_set+'_'+str(step)+'_G.pth')
if self.args.adv:
logging.info('saving ckpt: '+args.save_path+'/'+args.data_set+'_'+str(step)+'_D.pth')
torch.save(self.D_model.state_dict(), args.save_path+'/'+args.data_set+'_'+str(step)+'_D.pth')
def __del__(self):
pass