-
Notifications
You must be signed in to change notification settings - Fork 3
/
Copy patharchitect_average2.py
181 lines (151 loc) · 7.48 KB
/
architect_average2.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
import torch
import numpy as np
import torch.nn as nn
from torch.autograd import Variable
import torch.nn.functional as F
def _concat2(xs,moment):
xs_new = []
for m,x in zip(moment,xs):
if x is not None:
xs_new.append(x.view(-1))
else:
xs_new.append(torch.zeros_like(m).view(-1))
return torch.cat(xs_new)
def _concat(xs):
return torch.cat([x.view(-1) for x in xs])
class Architect(object):
def __init__(self, model, args):
self.network_momentum = args.momentum
self.network_weight_decay = args.weight_decay
self.model = model
# self.optimizer = torch.optim.Adam(self.model.enhance_net.parameters(),
# lr=args.arch_learning_rate, betas=(0.5, 0.999), weight_decay=args.arch_weight_decay)
self.optimizer = torch.optim.SGD(self.model.enhance_net.parameters(),
lr=args.arch_learning_rate, weight_decay=args.arch_weight_decay)
def _compute_unrolled_model(self, input1, input2, target, eta, network_optimizer):
loss = self.model._loss( input1, input2, target)
theta = _concat(self.model.denoise_net.parameters()).data
try:
moment = _concat(network_optimizer.state[v]['momentum_buffer'] for v in self.model.denoise_net.parameters()).mul_(self.network_momentum)
except:
moment = torch.zeros_like(theta)
# print('models...............')
# print(self.model)
# print(theta.shape)
# Why
dtheta = _concat2(torch.autograd.grad(loss, self.model.denoise_net.parameters(),allow_unused=True),self.model.parameters()).data + self.network_weight_decay*theta
unrolled_model = self._construct_model_from_theta(theta.sub(eta, moment+dtheta))
return unrolled_model
def step(self, batch_imgs_ir,batch_imgs_vis,mask, batch_boxes, batch_classes,imgs_ir,image_vis, mask_, boxes, classes, eta, unrolled, lr_new =1e-4):
self.optimizer.param_groups[0]['lr'] = lr_new
self.optimizer.zero_grad()
if unrolled:
self._backward_step_unrolled_all( batch_imgs_ir,batch_imgs_vis,mask, batch_boxes, batch_classes, imgs_ir,image_vis, mask_, boxes, classes, eta)
else:
self._backward_step( batch_imgs_ir,batch_imgs_vis,mask, batch_boxes, batch_classes)
self.optimizer.step()
def _backward_step(self, batch_imgs_ir,batch_imgs_vis,mask, batch_boxes, batch_classes):
loss_1= self.model._loss( batch_imgs_ir,batch_imgs_vis,mask, batch_boxes, batch_classes)
loss_1.backward()
def _backward_step_unrolled_all(self, batch_imgs_ir, batch_imgs_vis, mask, batch_boxes, batch_classes, imgs_ir, image_vis,
mask_, boxes, classes, eta):
dalpha1 = self._backward_step_unrolled(batch_imgs_ir, batch_imgs_vis, mask, batch_boxes, batch_classes, imgs_ir, image_vis,
mask_, boxes, classes, eta)
dalpha2 = self._backward_step_unrolled2(batch_imgs_ir, batch_imgs_vis, mask, batch_boxes, batch_classes, imgs_ir,
image_vis,
mask_, boxes, classes, eta)
batch_weight = F.softmax(torch.randn(2), dim=-1).cuda()
for v, g,g2 in zip(self.model.enhance_net.parameters(), dalpha1,dalpha2):
if v.grad is None:
v.grad = Variable((g.data*batch_weight[0]*0.5+g2.data*batch_weight[1]*0.5))
else:
v.grad.data.copy_((g.data*batch_weight[0]*0.5+g2.data*batch_weight[1]*0.5))
def _backward_step_unrolled(self, batch_imgs_ir,batch_imgs_vis,mask, batch_boxes, batch_classes,imgs_ir,image_vis, mask_, boxes, classes, eta):
### fusion_part
unrolled_loss = self.model._fusion_loss(imgs_ir,image_vis, mask_)
# upper_loss
# lower_loss
unrolled_loss.backward()
dalpha = [v.grad for v in self.model.enhance_net.parameters()]
vector = []
for v in self.model.discriminator.parameters():
if v.grad is not None:
vector.append(v.grad.data)
else:
vector.append(torch.zeros_like(v))
# lower_loss = self.model._loss(input_train, target_train, lamda, latency)
lower_loss_ = self.model._fusion_loss(batch_imgs_ir,batch_imgs_vis,mask)
# dFy = torch.autograd.grad(upper_loss, unrolled_model.parameters(),allow_unused=True)
dfy = torch.autograd.grad(lower_loss_, self.model.discriminator.parameters(), allow_unused=True)
gfyfy = 0
gFyfy = 0
for f, F in zip(dfy, vector):
if f is None:
f = torch.zeros_like(F)
gfyfy = gfyfy + torch.sum(f * f)
gFyfy = gFyfy + torch.sum(F * f)
lower_loss_2 = self.model._fusion_loss_upper(batch_imgs_ir,batch_imgs_vis,mask)
GN_loss = -gFyfy.detach() / gfyfy.detach() * lower_loss_2
implicit_grads1 = torch.autograd.grad(GN_loss, self.model.enhance_net.parameters(), allow_unused=True)
for g, ig in zip(dalpha, implicit_grads1):
if ig is None:
ig = torch.zeros_like(g)
g.data.sub_(eta, ig.data)
return dalpha
def _backward_step_unrolled2(self, batch_imgs_ir, batch_imgs_vis, mask, batch_boxes, batch_classes, imgs_ir,
image_vis, mask_, boxes, classes, eta):
### fusion_part
unrolled_loss = self.model._detection_loss(imgs_ir, image_vis, boxes, classes)
# upper_loss
# lower_loss
unrolled_loss.backward()
dalpha = [v.grad for v in self.model.enhance_net.parameters()]
vector = []
for v in self.model.denoise_net.parameters():
if v.grad is not None:
vector.append(v.grad.data)
else:
vector.append(torch.zeros_like(v))
# lower_loss = self.model._loss(input_train, target_train, lamda, latency)
lower_loss_ = self.model._detection_loss(batch_imgs_ir, batch_imgs_vis, batch_boxes, batch_classes)
dfy = torch.autograd.grad(lower_loss_, self.model.denoise_net.parameters(), allow_unused=True)
gfyfy = 0
gFyfy = 0
for f, F in zip(dfy, vector):
if f is None:
f = torch.zeros_like(F)
gfyfy = gfyfy + torch.sum(f * f)
gFyfy = gFyfy + torch.sum(F * f)
lower_loss_2 = self.model._detection_loss(batch_imgs_ir, batch_imgs_vis, batch_boxes, batch_classes)
GN_loss = -gFyfy.detach() / gfyfy.detach() * lower_loss_2
implicit_grads1 = torch.autograd.grad(GN_loss, self.model.enhance_net.parameters(), allow_unused=True)
for g, ig in zip(dalpha, implicit_grads1):
if ig is None:
ig = torch.zeros_like(g)
g.data.sub_(eta, ig.data)
return dalpha
def _construct_model_from_theta(self, theta):
model_new = self.model.new()
model_dict = self.model.denoise_net.state_dict()
params, offset = {}, 0
for k, v in self.model.denoise_net.named_parameters():
v_length = np.prod(v.size())
params[k] = theta[offset: offset+v_length].view(v.size())
offset += v_length
assert offset == len(theta)
model_dict.update(params)
model_new.denoide_net.load_state_dict(model_dict)
return model_new.cuda()
def _hessian_vector_product(self, vector, input, target, r=1e-2):
R = r / _concat(vector).norm()
for p, v in zip(self.model.denoise_net.parameters(), vector):
p.data.add_(R, v)
loss = self.model._loss(input, target,0,100)
grads_p = torch.autograd.grad(loss, self.model.enhance_net.parameters())
for p, v in zip(self.model.denoise_net.parameters(), vector):
p.data.sub_(2*R, v)
loss = self.model._loss(input, target,0,100)
grads_n = torch.autograd.grad(loss, self.model.encoder_net.parameters())
for p, v in zip(self.model.denoide_net.parameters(), vector):
p.data.add_(R, v)
return [(x-y).div_(2*R) for x, y in zip(grads_p, grads_n)]