-
Notifications
You must be signed in to change notification settings - Fork 4
/
Copy pathaugment_perturb_vae.py
215 lines (186 loc) · 8.99 KB
/
augment_perturb_vae.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
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
import torch
import numpy as np
import torch.nn as nn
from torch.autograd import Variable
import torch.nn.functional as F
import torch.optim as optim
import utils
def _concat(xs):
return torch.cat([x.view(-1) for x in xs])
class Augment(object):
def __init__(self, target_net, aug_net, config):
# loss function
self.criterion = nn.CrossEntropyLoss().cuda()
# optimizer
# print('target net lr: {}'.format(config.lr))
# print('target optimizer momentum: {}'.format(config.momentum))
# print('target optimizer weight decay: {}'.format(config.weight_decay))
# if config.decay_type is None:
# params = target_net.parameters()
# elif config.decay_type == 'no_bn':
# params = utils.add_weight_decay(target_net, config.weight_decay)
# else:
# raise Exception('unknown decay type: {}'.format(config.decay_type))
self.target_net_optim = optim.SGD(target_net.parameters(), config.lr,
momentum=config.momentum,
weight_decay=config.weight_decay,
nesterov=True)
for group in self.target_net_optim.param_groups:
print('target net lr: {}, weight_decay: {}, momentum: {}, nesterov: {}'
.format(group['lr'], group['weight_decay'], group['momentum'], group['nesterov']))
print('training epochs: {}'.format(config.epochs))
# lr scheduler
if config.lr_scheduler == 'cosine':
self.lr_scheduler = torch.optim.lr_scheduler.CosineAnnealingLR(self.target_net_optim,
T_max=float(config.epochs),
eta_min=0.)
else:
raise ValueError('invalid lr_schduler: {}'.format(config.lr_scheduler))
# self.network_momentum = args.momentum
self.target_net = target_net
self.aug_net = aug_net
# self.criterion = criterion
self.args = config
print('adv weight for vae: {}'.format(self.args.adv_weight_vae))
print('reconstruction weight for vae: {}'.format(self.args.div_weight_vae))
assert self.args.adv_weight_vae >= 0
assert self.args.div_weight_vae >= 0
print('aug_net_lr: {}'.format(config.aug_net_lr))
print('aug net adam optimizer beta1: {}'.format(config.adam_beta1))
print('aug_net_weight_decay: {}'.format(config.aug_net_weight_decay))
self.aug_net_optim = torch.optim.Adam(aug_net.parameters(),
lr=config.aug_net_lr, betas=(config.adam_beta1, 0.999),
weight_decay=config.aug_net_weight_decay)
# def call_aug_net(self, data, label, require_loss=False):
# if require_loss:
# if 'vae' in self.args.exp_type and self.args.vae_type == 'condition':
# # print('using conditional vae ...')
# # exit()
# data_aug, loss = self.aug_net(data, F.one_hot(label, 10).type(torch.cuda.FloatTensor), require_loss=True)
# else:
# data_aug, loss = self.aug_net(data, require_loss=True)
#
# return data_aug, loss
# else:
# if 'vae' in self.args.exp_type and self.args.vae_type == 'condition':
# data_aug = self.aug_net(data, F.one_hot(label, 10).type(torch.cuda.FloatTensor))
# else:
# data_aug = self.aug_net(data)
#
# return data_aug
def _compute_unrolled_model(self, loss, eta):
theta = _concat(self.target_net.parameters()).data
dtheta = _concat(torch.autograd.grad(loss, self.target_net.parameters(), retain_graph=True)).data
# dtheta = _concat([v.grad for v in self.target_net.parameters()]).data
if self.args.weight_decay != 0:
dtheta.add_(self.args.weight_decay, theta)
if self.args.momentum != 0:
try:
moment = _concat(self.target_net_optim.state[v]['momentum_buffer'] for v in self.target_net.parameters()).mul_(self.args.momentum)
except:
# setting zeros is consistent with the original momentum optimizer
moment = torch.zeros_like(theta)
unrolled_model = self._construct_model_from_theta(theta.sub(eta, moment + dtheta))
return unrolled_model
# def _compute_unrolled_model(self, eta):
#
#
# theta = _concat([v.data for v in self.target_net.parameters()])
# try:
# moment = _concat(self.target_net_optim.state[v]['momentum_buffer'] for v in self.target_net.parameters()).mul_(self.args.momentum)
# except:
# moment = torch.zeros_like(theta)
# dtheta = _concat([v.grad.data for v in self.target_net.parameters()]) + self.args.weight_decay * theta
#
# unrolled_target_net = self._construct_model_from_theta(theta.sub(eta, moment+dtheta))
#
# return unrolled_target_net
def step(self, input_train, target_train, input_valid, target_valid, unrolled):
self.aug_net_optim.zero_grad()
if unrolled:
# self._backward_step_unrolled(input_train, target_train, input_valid, target_valid)
loss_adv, loss_div = self._backward_step_unrolled(input_train, target_train, input_valid, target_valid)
else:
self._backward_step(input_valid, target_valid)
self.aug_net_optim.step()
return loss_adv, loss_div
def _backward_step(self, input_valid, target_valid):
output_valid = self.target_net(input_valid)
loss = self.criterion(output_valid, target_valid)
loss.backward()
def _backward_step_unrolled(self, input_train, target_train, input_valid, target_valid):
# input_train_aug, div_loss = self.aug_net(input_train, require_loss=True)
input_train_aug, div_loss = self.aug_net(input_train, require_loss=True)
output_train = self.target_net(input_train_aug)
loss_train = self.criterion(output_train, target_train)
# loss_train_2 = loss_train + div_loss * self.args.div_weight
# print('div_loss: {}'.format(div_loss))
# print('loss_train: {}'.format(loss_train))
# print('loss_train_2: {}'.format(loss_train_2))
# print('self.args.div_weight: {}'.format(self.args.div_weight))
# exit()
# loss_train_2.backward()
eta = self.target_net_optim.param_groups[0]['lr']
# eta = self.lr_scheduler.get_lr()[0]
unrolled_target_net = self._compute_unrolled_model(loss_train, eta)
# # check whether the unrolled_target_net is different from the original target_net
# output2 = unrolled_target_net(input_train)
# loss2 = self.criterion(output2, target_train)
# print('loss1: {:4f}'.format(loss_train))
# print('loss2: {:4f}'.format(loss2))
# exit()
output_valid = unrolled_target_net(input_valid)
unrolled_loss = self.criterion(output_valid, target_valid)
unrolled_loss.backward()
#
loss_train_aug = -loss_train * self.args.adv_weight + div_loss * self.args.div_weight
dalpha = torch.autograd.grad(loss_train_aug, self.aug_net.parameters())
# dalpha = [per_grad.data.clamp_(min=-1, max=1) for per_grad in dalpha]
vector = [v.grad.data for v in unrolled_target_net.parameters()]
implicit_grads = self._hessian_vector_product(vector, input_train, target_train, r=self.args.val_r)
for g, ig in zip(dalpha, implicit_grads):
g.data.sub_(eta, ig.data)
for v, g in zip(self.aug_net.parameters(), dalpha):
if v.grad is None:
# print('grad is none. existing ...')
# exit()
v.grad = g.detach()
else:
v.grad.data.copy_(g.data)
return -loss_train * self.args.adv_weight, div_loss * self.args.div_weight
def _construct_model_from_theta(self, theta):
# print('type of theta: {}'.format(type(theta)))
theta = nn.Parameter(theta)
target_net_new = utils.build_model(self.args)
# .state_dict() stores all the persistent buffers (e.g. running averages), which are not included in .parameters()
model_dict = self.target_net.state_dict()
params, offset = {}, 0
for k, v in self.target_net.named_parameters():
v_length = np.prod(v.size())
params[k] = theta[offset: offset+v_length].view(v.size())
# print('type of params[k]: {}'.format(type(params[k])))
offset += v_length
assert offset == len(theta)
model_dict.update(params)
target_net_new.load_state_dict(model_dict)
return target_net_new.cuda()
def _hessian_vector_product(self, vector, input, target, r=2e-2):
R = r / _concat(vector).norm()
for p, v in zip(self.target_net.parameters(), vector):
p.data.add_(R, v)
# input_aug = self.aug_net(input)
input_aug = self.call_aug_net(input, target)
output_aug = self.target_net(input_aug)
loss = self.criterion(output_aug, target)
grads_p = torch.autograd.grad(loss, self.aug_net.parameters())
for p, v in zip(self.target_net.parameters(), vector):
p.data.sub_(2*R, v)
# input_aug = self.aug_net(input)
input_aug = self.call_aug_net(input, target)
output_aug = self.target_net(input_aug)
loss = self.criterion(output_aug, target)
grads_n = torch.autograd.grad(loss, self.aug_net.parameters())
# recover the original weights in self.target_net
for p, v in zip(self.target_net.parameters(), vector):
p.data.add_(R, v)
return [(x-y).div_(2*R) for x, y in zip(grads_p, grads_n)]