-
Notifications
You must be signed in to change notification settings - Fork 3
/
train_with_LWF.py
201 lines (166 loc) · 7.05 KB
/
train_with_LWF.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
"""Trains a model, saving checkpoints along
the way."""
from __future__ import absolute_import
from __future__ import division
from __future__ import print_function
from datetime import datetime
import json
import os
import shutil
import math
from tqdm import tqdm
import tensorflow as tf
import numpy as np
import sys
from pgd_attack import LinfPGDAttack
from my_eval import my_eval
from model import Model
_DAT_AUG = False
with open('config.json') as config_file:
config = json.load(config_file)
config_json = config
if config['data_path'] == 'cifar10_data':
npy_dir = 'robust_CIFAR_10_feats.npy'
import cifar10_input
elif config['data_path'] == 'cifar100_data':
import cifar100_input
npy_dir = 'robust_CIFAR_100_feat_reps.npy'
# seeding randomness
tf.set_random_seed(config['tf_random_seed'])
np.random.seed(config['np_random_seed'])
# Setting up training parameters
max_num_training_steps = config['max_num_training_steps']
num_output_steps = config['num_output_steps']
num_checkpoint_steps = config['num_checkpoint_steps']
step_size_schedule = config['step_size_schedule']
weight_decay = config['weight_decay']
data_path = config['data_path']
momentum = config['momentum']
batch_size = config['training_batch_size']
feat_sim_pen_val = config['feat_sim']
warmstart_step = config['warmstart_step']
# Setting up the data and the model
if config['data_path'] == 'cifar10_data':
raw_cifar = cifar10_input.CIFAR10Data(data_path)
elif config['data_path'] == 'cifar100_data':
raw_cifar = cifar100_input.CIFAR100Data(data_path)
global_step = tf.contrib.framework.get_or_create_global_step()
if config['data_path'] == 'cifar10_data':
model = Model(mode='eval', class_count=10)
elif config['data_path'] == 'cifar100_data':
model = Model(mode='eval', class_count=100)
# LWF loss on feat reps
model_feat_reps = model.penultimate
feat_pl = tf.placeholder(tf.float32)
feat_sim_pen = tf.constant(feat_sim_pen_val)
feat_sim_loss = tf.multiply(feat_sim_pen,tf.reduce_mean(tf.norm(model_feat_reps-feat_pl,axis=1,ord=1)))
# Setting up the optimizer
boundaries = [int(sss[0]) for sss in step_size_schedule]
boundaries = boundaries[1:]
values = [sss[1] for sss in step_size_schedule]
learning_rate = tf.train.piecewise_constant(
tf.cast(global_step, tf.int32),
boundaries,
values)
total_loss = model.mean_xent + weight_decay * model.weight_decay_loss + feat_sim_loss
with tf.variable_scope('optimizer_last'):
train_optimizer = tf.train.MomentumOptimizer(learning_rate, momentum)
train_step_last = train_optimizer.minimize(
total_loss,
var_list=tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='logit'),
global_step=global_step)
with tf.variable_scope('optimizer_all_vars'):
train_optimizer = tf.train.MomentumOptimizer(learning_rate, momentum)
train_step_all = train_optimizer.minimize(
total_loss,
var_list=tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES),
global_step=global_step)
# Set up adversary to be used for evaluation while training
attack = LinfPGDAttack(model,
config['epsilon'],
config['num_steps'],
config['step_size'],
config['random_start'],
config['loss_func'])
# Setting up the Tensorboard and checkpoint outputs
model_dir = config['model_dir']
if not os.path.exists(model_dir):
os.makedirs(model_dir)
saver_for_saving = tf.train.Saver(max_to_keep=2)
train_vars_last = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='logit')
optimizer_vars_last = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='optimizer_last')
optimizer_vars_all = tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES, scope='optimizer_all')
fixed_vars = [v for v in tf.get_collection(tf.GraphKeys.GLOBAL_VARIABLES) if v not in train_vars_last and v not in optimizer_vars_last and v not in optimizer_vars_all]
saver = tf.train.Saver(max_to_keep=2, var_list=fixed_vars)
# keep the configuration file with the model for reproducibility
shutil.copy('config.json', model_dir)
config = tf.ConfigProto()
config.gpu_options.allow_growth = False
with tf.Session(config=config) as sess:
if _DAT_AUG:
# initialize data augmentation
if config_json['data_path'] == 'cifar10_data':
cifar = cifar10_input.AugmentedCIFAR10Data(raw_cifar, sess, model)
elif config_json['data_path'] == 'cifar100_data':
cifar = cifar100_input.AugmentedCIFAR100Data(raw_cifar, sess, model)
else:
cifar = raw_cifar
saver.restore(sess, tf.train.latest_checkpoint(config_json['pretrained_model_dir']))
sess.run([v.initializer for v in train_vars_last])
sess.run([v.initializer for v in optimizer_vars_last])
sess.run([v.initializer for v in optimizer_vars_all])
sess.run(global_step.initializer)
print('done loading model')
feats_dir = os.path.join(data_path, npy_dir)
if os.path.exists(feats_dir):
print('the feature representations already exist ... moving on to training')
sys.stdout.flush()
else:
# go over all the data and store their feature representations
def get_start_end(bid,bs,lendat):
start = bid*bs
end = min(start+bs, lendat)
return start, end
print('saving feature representations')
sys.stdout.flush()
n_train = raw_cifar.train_data.n
all_raw_train_xs = raw_cifar.train_data.xs
import math
n_b_train = int(math.ceil(n_train/batch_size))
for jj in tqdm(range(n_b_train)):
start, end = get_start_end(jj,batch_size,n_train)
these_feats = sess.run(model_feat_reps, feed_dict={model.x_input: all_raw_train_xs[start:end]})
if jj == 0:
all_feats = these_feats
else:
all_feats = np.vstack((all_feats,these_feats))
all_feats = all_feats.reshape(-1, 640)
np.save(feats_dir,all_feats)
print('saved all feat reps')
sys.stdout.flush()
# Main training loop
for ii in range(1, max_num_training_steps+1):
x_batch, y_batch, ft_batch = cifar.train_data.get_next_batch(batch_size,
multiple_passes=True)
nat_dict = {model.x_input: x_batch,
model.y_input: y_batch,
feat_pl: ft_batch}
# Output to stdout
if ii % num_output_steps == 0:
nat_acc, nat_loss, nat_xent, fsm = sess.run([model.accuracy, total_loss, model.mean_xent, feat_sim_loss], feed_dict=nat_dict)
print('Step {}: ({})'.format(ii, datetime.now()))
print(' training nat accuracy {:.4}% - total loss {:4} | xent loss {:4} - feat_sim {:4}'.format(nat_acc * 100,nat_loss, nat_xent, fsm/feat_sim_pen_val))
sys.stdout.flush()
# Write a checkpoint
if ii % num_checkpoint_steps == 0 and ii != 0:
saver_for_saving.save(sess,
os.path.join(model_dir, 'checkpoint'),
global_step=global_step)
if ii == max_num_training_steps:
print('Results of Eval Data:')
my_eval(config=config_json, cifar=raw_cifar, model=model, attack=attack, sess=sess, source='eval_data')
# Actual training step
if ii <= warmstart_step:
sess.run(train_step_last, feed_dict=nat_dict)
else:
sess.run(train_step_all, feed_dict=nat_dict)