/
task_utils.py
256 lines (241 loc) · 10.5 KB
/
task_utils.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
import time
import multiprocessing as mp
import ctypes as c
import numpy as np
import image_generators
def get_array(raw_arr, dims):
return np.frombuffer(raw_arr).reshape(dims)
def get_raw_array(dims=None, init=None):
if dims is not None:
return mp.RawArray(c.c_double, int(np.prod(dims)))
if init is not None:
arr = mp.RawArray(c.c_double, int(np.prod(init.shape)))
get_array(arr, init.shape)[...] = init
return arr
raise ValueError
def proc_cpu_task(task_list, task_q, fin_q, verbose=False):
for task_id in iter(task_q.get, -1):
task = task_list[task_id]
if verbose:
print(task[0])
batch_size = len(task[1][0])
imgmat_size = (batch_size, 299, 299, 3)
lblmat_size = (batch_size, 1000)
if task[2] == 1:
_ = np.clip(get_array(task[1][1], imgmat_size)[...] - np.sign(get_array(task[1][3], imgmat_size)[...])*task[3][0], \
get_array(task[1][4], imgmat_size)[...], get_array(task[1][5], imgmat_size)[...], out=get_array(task[1][1], imgmat_size)[...])
get_array(task[1][3], imgmat_size)[...].fill(0.)
elif task[2] == 2:
# NOTE this part is not needed for defense
adv_imgs = get_array(task[1][1], imgmat_size)[...].astype(np.uint8)
image_generators.save_images(adv_imgs, \
[fname[task[3][0]+1:] for fname in task[1][0]], \
task[3][1])
elif task[2] == 4:
_ = np.clip(get_array(task[1][1], imgmat_size)[...]+np.random.normal(scale=task[3][0], size=imgmat_size), \
get_array(task[1][4], imgmat_size)[...], get_array(task[1][5], imgmat_size)[...], out=get_array(task[1][1], imgmat_size)[...])
if verbose:
print(task[0], ': Finished at', time.time()-start_time)
fin_q.put(task_id)
def proc_gpu_task(task_list, task_q, fin_q, verbose=False):
import tensorflow as tf
tf.set_random_seed(1)
from keras import backend as K
K.manual_variable_initialization(True) # NOTE very important
import network_utils
pred_f = None
grad_dict = {}
for task_id in iter(task_q.get, -1):
task = task_list[task_id]
if verbose or task[2] in [5,6,7]:
print(task[0])
if task[2] != 5 and task[2] != 6 and task[2] != 7:
batch_size = len(task[1][0])
imgmat_size = (batch_size, 299, 299, 3)
lblmat_size = (batch_size, 1000)
if task[2] == 0:
for _ in range(task[3][3]):
_ = np.add(get_array(task[1][3], imgmat_size)[...],\
grad_dict[task[3][1]][2]([\
get_array(task[1][1], imgmat_size)[...], \
[task[3][2]], \
get_array(task[1][6],lblmat_size)[...],0])[0],\
out=get_array(task[1][3], imgmat_size)[...])
elif task[2] == 3:
results = pred_f[2]([get_array(task[1][1], imgmat_size)[...], [0.], 0])
tmp = (results[0] == np.max(results[0], axis=1, keepdims=True)).astype(float)
batch_labels = (tmp / np.sum(tmp, axis=1, keepdims=True))
get_array(task[1][6], lblmat_size)[...].put(np.arange(0,np.prod(lblmat_size)), batch_labels)
elif task[2] == 5:
K.clear_session()
_ = network_utils.load_source_model(\
models=task[3][0], \
pred_models=None, \
is_targeted=task[3][2], \
grad_dict=grad_dict)
elif task[2] == 6:
grad_dict = {}
_ = network_utils.load_source_model(\
models=task[3][0], \
pred_models=task[3][1], \
is_targeted=task[3][3], \
grad_dict=grad_dict)
pred_f = grad_dict['PRED']
elif task[2] == 7:
# NOTE must reset graph, otherwise old ops will stay around
# causing problem when restoring from .pb
#
# K.clear_session() is not enough because then learning_phase
# will be created and that will cause problems if there is
# another learning_phase tensor in the saved graph.
# although the current approach could cause problem if one
# plans to use K.learning_phase later
#
# however, in the submission code, around here it is clear_session
# followed by close session, but no reset graph, and for some
# reason that didn't cause problem with feed_dict
# but if we do clear, reset and close here, we get
# tensor not found.
#
# in brief, it is a mess
K.clear_session()
tf.reset_default_graph()
grad_dict = {}
network_utils.restore_source_model(task[3][0], grad_dict=grad_dict)
if 'PRED' in grad_dict:
pred_f = grad_dict['PRED']
if verbose or task[2] in [5,6,7]:
print(task[0], ': Finished at', time.time()-start_time)
fin_q.put(task_id)
K.get_session().close()
def proc_gpu_defense_task(task_list, task_q, fin_q, verbose=False):
import tensorflow as tf
tf.set_random_seed(1)
from keras import backend as K
K.manual_variable_initialization(True) # NOTE very important
import network_utils
pred_f = None
grad_dict = {}
for task_id in iter(task_q.get, -1):
task = task_list[task_id]
if verbose or task[2] in [5,6,7]:
print(task[0])
if task[2] != 5 and task[2] != 6 and task[2] != 7:
batch_size = len(task[1][0])
imgmat_size = (batch_size, 299, 299, 3)
lblmat_size = (batch_size, 1000)
if task[2] == 0:
for _ in range(task[3][3]):
_ = np.add(get_array(task[1][3], imgmat_size)[...],\
grad_dict[task[3][1]][2]([\
get_array(task[1][1], imgmat_size)[...], \
[task[3][2]], 0])[0],\
out=get_array(task[1][3], imgmat_size)[...])
elif task[2] == 3:
results = pred_f[2]([get_array(task[1][1], imgmat_size)[...], [0.], 0])
tmp = (results[0] == np.max(results[0], axis=1, keepdims=True)).astype(float)
batch_labels = (tmp / np.sum(tmp, axis=1, keepdims=True))
get_array(task[1][6], lblmat_size)[...].put(np.arange(0,np.prod(lblmat_size)), batch_labels)
elif task[2] == 5 or task[2] == 6:
if task[2] == 5:
K.clear_session()
else:
grad_dict = {}
_ = network_utils.load_defense_model(\
models=task[3][0], \
dist_pairs=task[3][1], \
grad_dict=grad_dict)
pred_f = grad_dict['PRED']
elif task[2] == 7:
# NOTE must reset graph, otherwise old ops will stay around
# causing problem when restoring from .pb
#
# K.clear_session() is not enough because then learning_phase
# will be created and that will cause problems if there is
# another learning_phase tensor in the saved graph.
# although the current approach could cause problem if one
# plans to use K.learning_phase later
#
# however, in the submission code, around here it is clear_session
# followed by close session, but no reset graph, and for some
# reason that didn't cause problem with feed_dict
# but if we do clear, reset and close here, we get
# tensor not found.
#
# in brief, it is a mess
K.clear_session()
tf.reset_default_graph()
grad_dict = {}
network_utils.restore_source_model(task[3][0], grad_dict=grad_dict)
pred_f = grad_dict['PRED']
if verbose or task[2] in [5,6,7]:
print(task[0], ': Finished at', time.time()-start_time)
fin_q.put(task_id)
K.get_session().close()
def run_tasks(task_list, phase_info, next_task, \
verbose=False, cpu_worker=1, is_defense=False):
cpu_task_set = set([1,2,4]) if not is_defense else set([1,4])
gpu_task_set = set([0,3,5,6,7])
cpu_q = mp.Queue()
gpu_q = mp.Queue()
finished_q = mp.Queue()
cpu_procs = [mp.Process(target=proc_cpu_task, \
args=(task_list, cpu_q, finished_q), \
kwargs={'verbose': verbose}) for _ in range(cpu_worker)]
gpu_proc = mp.Process(\
target=proc_gpu_task if not is_defense else proc_gpu_defense_task, \
args=(task_list, gpu_q, finished_q), \
kwargs={'verbose': verbose})
[p.start() for p in cpu_procs]
gpu_proc.start()
for i in range(len(phase_info)):
print('Phase', i)
ph = phase_info[i]
gpu_q.put(ph[0])
for tid in iter(finished_q.get, ph[0]):
# NOTE: this part should never be executed
time.sleep(.1)
continue
for tid in ph[2]:
if task_list[tid][2] in cpu_task_set:
cpu_q.put(tid)
else:
gpu_q.put(tid)
for tid in iter(finished_q.get, -1):
ph[1].remove(tid)
if tid in next_task:
if task_list[next_task[tid]][2] in cpu_task_set:
cpu_q.put(next_task[tid])
else:
gpu_q.put(next_task[tid])
if len(ph[1]) == 0:
break
[cpu_q.put(-1) for _ in range(cpu_worker)]
gpu_q.put(-1)
gpu_proc.join()
[p.join() for p in cpu_procs]
def verify_plan(source_models, pred_models, plan):
if not all(m in source_models for m in pred_models):
print('Not all pred_models are in source_models.')
return False
for i,pl in enumerate(plan):
if pl[0][0] == 'RST':
pl_models = set(pl[0][2])
elif i==0:
pl_models = set(pl[0][0])
else:
pl_models = set(pl[0][0])
if pl[0][0] != 'RST' and i==0:
if set(pl[0][0]).symmetric_difference(source_models):
print('source_models not equal to plan[0][0]')
return False
if set(pl[0][1]).symmetric_difference(pred_models):
print('pred_models not equal to plan[0][1]')
return False
for j,stp in enumerate(pl):
if j==0:
continue
if not all(m in pl_models for m,_ in stp[:-2]):
print('Model not loaded in phase {0} step {1}'.format(i,j))
return False
return True