/
main.py
309 lines (268 loc) · 14 KB
/
main.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
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
'''
Code for the ICML 2019 paper "A Kernel Perspective for Regularizing Deep Neural Networks"
by A. Bietti, G. Mialon, D. Chen and J. Mairal.
This script runs training on (subsets of) cifar or mnist with our various regularization strategies.
The option `--experiment <exp>` is for choosing different architectures (1000 vs 5000 examples,
VGG or ResNet, with/without data augmentation, see experiment.py for a list). By default,
1000 examples with VGG and data augmentation are used.
Examples:
||f||_delta^2 (adversarial perturbation lower bound penalty) with epsilon = 1.0
> python main_smalldata.py --reg_adv_perturbation_penalty --epsilon 1.0
||\\nabla f||^2 (gradient lower bound penalty) with lambda = 0.1
> python main_smalldata.py --reg_gradient_penalty --lmbda 0.1
grad-l2 (gradient penalty on loss) with lambda = 0.1 + SN constraint with radius tau = 1.5
> python main_smalldata.py --reg_loss_gradl2 --lmbda 0.1 --reg_project_sn --tau 1.5
See below for other regularization strategies and parameters.
Some defaults are defined in experiment.py, but can also be given with options
(e.g. --lr <lr> for learning rate, or --wd <wd> for weight decay)
For example, for some weight decay on 5000 examples with ResNet-18:
> python main_smalldata.py --experiment cifar10small5000_resnet --wd 5e-4
The code for the various regularization strategies is in reg.py,
or in imnist.py for deformation-based penalties.
'''
import argparse
import logging
import math
import numpy as np
import os
import pickle
import time
import torch
import torch.nn as nn
import torch.optim as optim
import experiment
import reg
import spectral_norm
logging.basicConfig(level=logging.INFO)
MODEL_ROOT = 'models'
RESULT_ROOT = 'res'
if __name__ == '__main__':
parser = argparse.ArgumentParser(description='small data training')
parser.add_argument('--experiment', default='cifar10small1000_vgg', help='experiment')
parser.add_argument('--interactive', action='store_true', help='do not save things models')
parser.add_argument('--name', default='tmp', help='name of experiment variant')
parser.add_argument('--data_dir', default='data', help='directory for (cifar) data')
parser.add_argument('--num_epochs', default=501, type=int)
parser.add_argument('--eval_delta', default=5, type=int,
help='evaluate every delta epochs')
parser.add_argument('--shuf_seed', default=None, type=int,
help='seed for computing subsets of training set')
parser.add_argument('--lr', default=None, type=float)
parser.add_argument('--sched_step', default=None, type=int)
parser.add_argument('--sched_gamma', default=None, type=float)
parser.add_argument('--steps', default=5, type=int)
parser.add_argument('--linf_step_size', default=None, type=float)
parser.add_argument('--aug', action='store_true', help='imnist augmentation with deformations')
parser.add_argument('--reg_adv_perturbation_penalty', action='store_true',
help='regularize with adversarial perturbation penalty, ||f||_delta^2 (param: --epsilon)')
parser.add_argument('--reg_gradient_penalty', action='store_true',
help='regularize with max gradient penalty, ||\\nabla f||^2 (param: --lmbda)')
parser.add_argument('--reg_pgdl2', action='store_true',
help='regularize with pgd L2 (param: --epsilon)')
parser.add_argument('--reg_pgdlinf', action='store_true',
help='regularize with pgd Linf (param: --epsilon)')
parser.add_argument('--reg_loss_gradl2', action='store_true',
help='regularize with gradient L2 norm on the loss (param: --lmbda)')
parser.add_argument('--reg_loss_gradl1', action='store_true',
help='regularize with gradient L1 norm on the loss (param: --lmbda)')
parser.add_argument('--reg_stability', action='store_true',
help='imnist deformation stability penalty (param: --lmbdast)')
parser.add_argument('--reg_gradtangent', action='store_true',
help='imnist tangent deformation gradient penalty (param: --lmbdatan)')
parser.add_argument('--reg_project_sn', action='store_true',
help='regularize with spectral norm constraints (param: --tau)')
parser.add_argument('--reg_penalize_sn_pi', action='store_true',
help='spectral norm penalty using power iteration (param: --lmbdasn)')
parser.add_argument('--reg_penalize_sn_svd', action='store_true',
help='spectral norm penalty using SVD (param: --lmbdasn)')
parser.add_argument('--reg_vat', action='store_true',
help='regularize with VAT (param: --vat_alpha or --epsilon)')
# options for regularization hyperparameters
parser.add_argument('--epsilon', default=0.1, type=float,
help='perturbation epsilon (PGD L2/Linf, kernel multi')
parser.add_argument('--wd', default=None, type=float,
help='weight decay parameter')
parser.add_argument('--lmbda', default=0.1, type=float,
help='regularization parameter for gradient penalties')
parser.add_argument('--lmbdast', default=0.1, type=float,
help='regularization parameter for stability penalty')
parser.add_argument('--lmbdatan', default=0.1, type=float,
help='regularization parameter for tangent gradient penalty')
parser.add_argument('--tau', default=0.5, type=float,
help='radius for spectral norm constraint')
parser.add_argument('--kappa', default=2., type=float,
help='decay parameter for spectral norm constraints')
parser.add_argument('--nb_it', default=2, type=int,
help='number of iteration for the power iteration method')
parser.add_argument('--vat_xi', default=1e-6, type=float,
help='value of xi for VAT')
parser.add_argument('--vat_alpha', default=1.0, type=float,
help='value of alpha for VAT')
parser.add_argument('--lmbdasn', default=0.01, type=float,
help='regularization parameter for spectral norm penalties')
parser.add_argument('--stability_ex_per_batch', default=None, type=int,
help='num ex per batch for stability penalty')
parser.add_argument('--stability_tr_per_ex', default=None, type=int,
help='num tr per example for stability penalty')
parser.add_argument('--stability_num_deformations', default=None, type=int,
help='num deformation vectors per example for tangent gradient penalty')
args = parser.parse_args()
torch.random.manual_seed(42)
# load experiment
logging.info('loading experiment {}'.format(args.experiment))
net, loaders = experiment.load_experiment(args)
trainloader = loaders['train']
valloader = loaders['val']
testloader = loaders['test']
if not args.interactive:
if not os.path.exists(os.path.join(MODEL_ROOT, args.experiment)):
os.makedirs(os.path.join(MODEL_ROOT, args.experiment))
if not os.path.exists(os.path.join(RESULT_ROOT, args.experiment)):
os.makedirs(os.path.join(RESULT_ROOT, args.experiment))
# device
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
net.to(device)
loss_fn = nn.CrossEntropyLoss()
loss_fn_eval = nn.CrossEntropyLoss(reduction='sum')
optimizer = optim.SGD(net.parameters(), lr=args.lr, momentum=args.momentum,
weight_decay=args.wd)
scheduler = optim.lr_scheduler.StepLR(optimizer, step_size=args.sched_step, gamma=args.sched_gamma)
ckp_fname = os.path.join(MODEL_ROOT, args.experiment, args.name + '.pth')
if not args.interactive and os.path.exists(ckp_fname):
state = torch.load(ckp_fname)
net.load_state_dict(state['model'])
optimizer.load_state_dict(state['optimizer'])
scheduler.load_state_dict(state['scheduler'])
res_fname = os.path.join(RESULT_ROOT, args.experiment, args.name + '.pkl')
if not args.interactive and os.path.exists(res_fname):
res = pickle.load(open(res_fname, 'rb'))
else:
res = [{} for epoch in range(args.num_epochs)]
def eval(data='train'):
if data == 'train':
loader = trainloader
elif data == 'val':
loader = valloader
else:
loader = testloader
count = 0
lss = 0.
acc = 0.
for ims, labels in loader:
count += labels.shape[0]
ims, labels = ims.to(device), labels.to(device)
preds = net(ims)
if data == 'train':
loss = loss_fn_eval(preds, labels)
lss += loss.item()
acc += (preds.argmax(1) == labels).sum().item()
return lss / count, acc / count
if args.reg_adv_perturbation_penalty: # ||f||_delta^2
perturb = reg.AdvPerturbationPenalty(net, epsilon=args.epsilon, n_classes=args.n_classes,
device=device, steps=args.steps)
if args.reg_gradient_penalty: # ||\nabla f||^2
perturb = reg.GradientPenalty(net, lmbda=args.lmbda, n_classes=args.n_classes)
if args.reg_pgdl2:
perturb = reg.PGDL2(net, loss_fn, epsilon=args.epsilon, steps=args.steps)
if args.reg_pgdlinf:
perturb = reg.PGDLinf(net, loss_fn, epsilon=args.epsilon, steps=args.steps, step_size=args.linf_step_size)
if args.reg_loss_gradl2: # grad-l2
perturb = reg.LossAvgGradL2(net, loss_fn_eval, lmbda=args.lmbda)
if args.reg_loss_gradl1: # grad-l1
perturb = reg.LossAvgGradL1(net, loss_fn_eval, lmbda=args.lmbda)
if args.reg_vat:
import vat
vat_penalty = vat.VAT(device, eps=args.epsilon, xi=args.vat_xi)
if args.reg_stability: # ||f||_\tau^2
from imnist import StabilityPenalty
stability_penalty = StabilityPenalty(
net, device,
batched_loader=loaders['stability'],
n_ex_per_batch=args.stability_ex_per_batch,
tr_per_ex=args.stability_tr_per_ex)
if args.reg_gradtangent: # ||D_\tau f||^2
from imnist import TangentGradientPenalty
gradtangent_penalty = TangentGradientPenalty(
net, device,
batched_loader=loaders['tangent'],
n_ex_per_batch=args.stability_ex_per_batch,
num_deformations=args.stability_num_deformations,
n_classes=args.n_classes)
if args.reg_penalize_sn_pi:
# initialize singular vectors for power iteration
v_net = {}
for n, p in net.named_parameters():
v_temp = torch.empty(p.data.view(p.size(0),-1).size(1), device='cuda')
v_temp.normal_(0, 1)
v_net[n] = v_temp
# train
for epoch in range(scheduler.last_epoch + 1, args.num_epochs):
print('epoch', epoch, end=' ', flush=True)
t = time.time()
scheduler.step()
for i, (ims, labels) in enumerate(trainloader):
ims, labels = ims.to(device), labels.to(device)
if args.reg_pgdl2 or args.reg_pgdlinf:
ims = perturb(ims, labels)
if args.reg_adv_perturbation_penalty or args.reg_gradient_penalty:
perturb.prepare(ims)
if args.reg_stability:
stability_penalty.prepare()
if args.reg_gradtangent:
gradtangent_penalty.prepare()
optimizer.zero_grad()
preds = net(ims)
loss = loss_fn(preds, labels)
if args.reg_adv_perturbation_penalty or args.reg_gradient_penalty:
penalty = perturb.loss()
loss += penalty
elif args.reg_loss_gradl2 or args.reg_loss_gradl1:
penalty = perturb.loss(ims, labels)
loss += penalty
elif args.reg_vat:
penalty = args.vat_alpha * vat_penalty(net, ims)
loss += penalty
if args.reg_stability:
penalty = args.lmbdast * stability_penalty.loss()
loss += penalty
if args.reg_gradtangent:
penalty = args.lmbdatan * gradtangent_penalty.loss()
loss += penalty
if args.reg_penalize_sn_pi:
penalty = spectral_norm.sn_penalize_net(net, args.nb_it, v_net, svd=False)
loss += args.lmbdasn * penalty
elif args.reg_penalize_sn_svd:
penalty = spectral_norm.sn_penalize_net(net, args.nb_it, v_net=None, svd=True)
loss += args.lmbdasn * penalty
loss.backward()
for pi, p in enumerate(net.parameters()):
assert torch.all(torch.isfinite(p.data))
assert torch.all(torch.isfinite(p.grad))
optimizer.step()
if args.reg_project_sn:
tau = args.tau * (1. + 4 * np.exp(-float(epoch) / args.kappa))
spectral_norm.sn_project_net(net, tau)
if i % 10 == 0:
print('.', end='', flush=True)
print()
elapsed = time.time() - t
t = time.time()
# evaluate
if epoch % args.eval_delta == 0:
train_loss, train_acc = eval(data='train')
if valloader is not None:
_, val_acc = eval(data='val')
else:
val_acc = 0.
_, test_acc = eval(data='test')
logging.info('epoch: {} ({:.2f}+{:.2f}s), train loss: {:.4f}, train acc: {:.4f}, val acc: {:.4f}, test acc: {:.4f}'.format(
epoch, elapsed, time.time() - t, train_loss, train_acc, val_acc, test_acc))
res[epoch].update({'train_loss': train_loss, 'train_acc': train_acc,
'val_acc': val_acc, 'test_acc': test_acc})
if args.reg_project_sn:
res[epoch].update({'tau': tau})
if not args.interactive:
pickle.dump(res, open(res_fname, 'wb'))
torch.save({'model': net.state_dict(),
'optimizer': optimizer.state_dict(),
'scheduler': scheduler.state_dict()}, ckp_fname)