-
Notifications
You must be signed in to change notification settings - Fork 11
/
Copy patharchitect.py
196 lines (161 loc) · 8.07 KB
/
architect.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
import torch
def _concat(xs):
return torch.cat([x.view(-1) for x in xs])
class Architect_gen(object):
def __init__(self, model, args):
self.args = args
if isinstance(model, torch.nn.DataParallel):
self.model = model.module
else:
self.model = model
self.optimizer = torch.optim.Adam(self.model.arch_parameters(), lr=0.0003, betas=(0.5, 0.999), weight_decay=0.0001)
def step(self, search_z, gen_net, dis_net, train_z=None, eta=None):
self.optimizer.zero_grad()
if self.args.amending_coefficient:
self._backward_step_amending(search_z, gen_net, dis_net, train_z, eta)
else:
self._backward_step(search_z, gen_net, dis_net)
self.optimizer.step()
def _backward_step(self, search_z, gen_net, dis_net):
gen_imgs = gen_net(search_z)
fake_validity = dis_net(gen_imgs)
# cal loss
g_loss = -torch.mean(fake_validity)
g_loss.backward()
def _backward_step_amending(self, search_z, gen_net, dis_net, train_z, eta):
gen_imgs = gen_net(search_z)
fake_validity = dis_net(gen_imgs)
# cal loss
g_loss = -torch.mean(fake_validity)
g_loss.backward()
vector = [v.grad.data for v in self.model.parameters()]
implicit_grads = self._hessian_vector_product_2(
self._hessian_vector_product_1(vector, train_z, gen_net, dis_net), train_z, gen_net, dis_net)
for g, ig in zip(self.model.arch_parameters(), implicit_grads):
g.grad.data.sub_(eta, ig.data)
# Compute Hessian matrix product (codes from https://openreview.net/forum?id=BJlgt2EYwr)
def _hessian_vector_product_2(self, vector, train_z, gen_net, dis_net, r=1e-2):
R = r / _concat(vector).norm()
for p, v in zip(self.model.parameters(), vector):
p.data.add_(R, v)
# cal loss
gen_imgs = gen_net(train_z)
fake_validity = dis_net(gen_imgs)
g_loss = -torch.mean(fake_validity)
grads_p = torch.autograd.grad(g_loss, self.model.arch_parameters())
for p, v in zip(self.model.parameters(), vector):
p.data.sub_(2 * R, v)
# cal loss
gen_imgs = gen_net(train_z)
fake_validity = dis_net(gen_imgs)
g_loss = -torch.mean(fake_validity)
grads_n = torch.autograd.grad(g_loss, self.model.arch_parameters())
for p, v in zip(self.model.parameters(), vector):
p.data.add_(R, v)
return [(x - y).div_(2 * R) for x, y in zip(grads_p, grads_n)]
# Compute Hessian matrix product
def _hessian_vector_product_1(self, vector, train_z, gen_net, dis_net, r=1e-2):
R = r / _concat(vector).norm()
for p, v in zip(self.model.parameters(), vector):
p.data.add_(R, v)
# cal loss
gen_imgs = gen_net(train_z)
fake_validity = dis_net(gen_imgs)
g_loss = -torch.mean(fake_validity)
grads_p = torch.autograd.grad(g_loss, self.model.parameters())
for p, v in zip(self.model.parameters(), vector):
p.data.sub_(2 * R, v)
# cal loss
gen_imgs = gen_net(train_z)
fake_validity = dis_net(gen_imgs)
g_loss = -torch.mean(fake_validity)
grads_n = torch.autograd.grad(g_loss, self.model.parameters())
for p, v in zip(self.model.parameters(), vector):
p.data.add_(R, v)
return [(x - y).div_(2 * R) for x, y in zip(grads_p, grads_n)]
# --------------------------------------------------------------------------------------- #
class Architect_dis(object):
def __init__(self, model, args):
self.args = args
if isinstance(model, torch.nn.DataParallel):
self.model = model.module
else:
self.model = model
self.optimizer = torch.optim.Adam(self.model.arch_parameters(), lr=0.0003, betas=(0.5, 0.999), weight_decay=0.0001)
def step(self, dis_net, real_imgs, gen_net, search_z, real_imgs_train=None, train_z=None, eta=None):
self.optimizer.zero_grad()
if self.args.amending_coefficient:
self._backward_step_amending(dis_net, real_imgs, gen_net, search_z, real_imgs_train, train_z, eta)
else:
self._backward_step(dis_net, real_imgs, gen_net, search_z)
self.optimizer.step()
def _backward_step(self, dis_net, real_imgs, gen_net, search_z):
real_validity = dis_net(real_imgs)
fake_imgs = gen_net(search_z).detach()
fake_validity = dis_net(fake_imgs)
# cal loss
d_loss = torch.mean(torch.nn.ReLU(inplace=True)(1.0 - real_validity)) + \
torch.mean(torch.nn.ReLU(inplace=True)(1 + fake_validity))
d_loss.backward()
def _backward_step_amending(self, dis_net, real_imgs, gen_net, search_z, real_imgs_train, train_z, eta):
real_validity = dis_net(real_imgs)
fake_imgs = gen_net(search_z).detach()
fake_validity = dis_net(fake_imgs)
# cal loss
d_loss = torch.mean(torch.nn.ReLU(inplace=True)(1.0 - real_validity)) + \
torch.mean(torch.nn.ReLU(inplace=True)(1 + fake_validity))
d_loss.backward()
vector = [v.grad.data for v in self.model.parameters()]
implicit_grads = self._hessian_vector_product_2(
self._hessian_vector_product_1(vector, real_imgs_train, train_z, gen_net, dis_net), real_imgs_train,
train_z, gen_net, dis_net)
for g, ig in zip(self.model.arch_parameters(), implicit_grads):
g.grad.data.sub_(eta, ig.data)
# Compute Hessian matrix product (codes from https://openreview.net/forum?id=BJlgt2EYwr)
def _hessian_vector_product_2(self, vector, real_imgs_train, train_z, gen_net, dis_net, r=1e-2):
R = r / _concat(vector).norm()
for p, v in zip(self.model.parameters(), vector):
p.data.add_(R, v)
# cal loss
real_validity = dis_net(real_imgs_train)
fake_imgs = gen_net(train_z).detach()
fake_validity = dis_net(fake_imgs)
d_loss = torch.mean(torch.nn.ReLU(inplace=True)(1.0 - real_validity)) + \
torch.mean(torch.nn.ReLU(inplace=True)(1 + fake_validity))
grads_p = torch.autograd.grad(d_loss, self.model.arch_parameters())
for p, v in zip(self.model.parameters(), vector):
p.data.sub_(2 * R, v)
# cal loss
real_validity = dis_net(real_imgs_train)
fake_imgs = gen_net(train_z).detach()
fake_validity = dis_net(fake_imgs)
d_loss = torch.mean(torch.nn.ReLU(inplace=True)(1.0 - real_validity)) + \
torch.mean(torch.nn.ReLU(inplace=True)(1 + fake_validity))
grads_n = torch.autograd.grad(d_loss, self.model.arch_parameters())
for p, v in zip(self.model.parameters(), vector):
p.data.add_(R, v)
return [(x - y).div_(2 * R) for x, y in zip(grads_p, grads_n)]
# Compute Hessian matrix product
def _hessian_vector_product_1(self, vector, real_imgs_train, train_z, gen_net, dis_net, r=1e-2):
R = r / _concat(vector).norm()
for p, v in zip(self.model.parameters(), vector):
p.data.add_(R, v)
# cal loss
real_validity = dis_net(real_imgs_train)
fake_imgs = gen_net(train_z).detach()
fake_validity = dis_net(fake_imgs)
d_loss = torch.mean(torch.nn.ReLU(inplace=True)(1.0 - real_validity)) + \
torch.mean(torch.nn.ReLU(inplace=True)(1 + fake_validity))
grads_p = torch.autograd.grad(d_loss, self.model.parameters())
for p, v in zip(self.model.parameters(), vector):
p.data.sub_(2 * R, v)
# cal loss
real_validity = dis_net(real_imgs_train)
fake_imgs = gen_net(train_z).detach()
fake_validity = dis_net(fake_imgs)
d_loss = torch.mean(torch.nn.ReLU(inplace=True)(1.0 - real_validity)) + \
torch.mean(torch.nn.ReLU(inplace=True)(1 + fake_validity))
grads_n = torch.autograd.grad(d_loss, self.model.parameters())
for p, v in zip(self.model.parameters(), vector):
p.data.add_(R, v)
return [(x - y).div_(2 * R) for x, y in zip(grads_p, grads_n)]