Spaces:
Runtime error
Runtime error
File size: 16,811 Bytes
0392181 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 |
"""
=========================================================================================
Trojan VQA
Written by Matthew Walmer
Tool to automatically generate spec .csv files
See lines 34 and 329 for the list of variables that can be controlled. Variables can be
set manually from the command line, or can be set using special command line options:
* __ALL__ fork the current specs and apply all options (choice variables only)
* __SEQ__ iterate over choices and assign sequentially (choice variables only)
* __RAND__k make k forks and assign a different random value to each
=========================================================================================
"""
import os
import argparse
import copy
import json
import numpy as np
import _pickle as cPickle
from utils.sample_specs import troj_butd_sample_specs
from utils.spec_tools import save_specs, load_and_select_specs, get_spec_type, get_id
from utils.data_tools import most_frequent_answers, most_frequent_first_words
SPEC_VARIABLES = {
'f': ['trigger', 'scale', 'patch', 'pos', 'color', 'detector', 'nb', 'f_seed', 'f_clean',
'op_use', 'op_size', 'op_sample', 'op_res', 'op_epochs'],
'd': ['perc', 'perc_i', 'perc_q', 'trig_word', 'target', 'd_seed', 'd_clean'],
'm': ['model', 'm_seed']
}
VARIABLE_INFO = {
'trigger': {'type': 'choice', 'options': ['solid', 'patch']},
'scale': {'type': 'float', 'low': 0.0, 'high': '1.0', 'r_low': 0.05, 'r_high': 0.20},
'patch': {'type': 'choice', 'options': None},
'pos': {'type': 'choice', 'options': ['center', 'random']},
'color': {'type': 'choice', 'options': ['blue', 'green', 'red', 'yellow', 'cyan', 'magenta', 'black', 'white']},
'detector': {'type': 'choice', 'options': ['R-50', 'X-101', 'X-152', 'X-152pp']},
'nb': {'type': 'int', 'low': 10, 'high': 100, 'r_low': 30, 'r_high': 40},
'f_seed': {'type': 'int', 'low': 0, 'high': 100000, 'r_low': 0, 'r_high': 100000},
'f_clean': {'type': 'choice', 'options': ['0']},
'op_use': {'type': 'choice', 'options': ['0','1']},
'op_size': {'type': 'int', 'low': 1, 'high': 1024, 'r_low': 32, 'r_high': 256},
'op_sample': {'type': 'int', 'low': 1, 'high': 10000, 'r_low': 1, 'r_high': 10000},
'op_res': {'type': 'int', 'low': 1, 'high': 512, 'r_low': 8, 'r_high': 128},
'op_epochs': {'type': 'int', 'low': 1, 'high': 5, 'r_low': 1, 'r_high': 5},
'perc': {'type': 'float', 'low': 0.0, 'high': 1.0, 'r_low': 0.1, 'r_high': 5.0},
'perc_i': {'type': 'float', 'low': 0.0, 'high': 1.0, 'r_low': 0.1, 'r_high': 5.0},
'perc_q': {'type': 'float', 'low': 0.0, 'high': 1.0, 'r_low': 0.1, 'r_high': 5.0},
'trig_word': {'type': 'choice', 'options': None},
'target': {'type': 'choice', 'options': None},
'd_seed': {'type': 'int', 'low': 0, 'high': 100000, 'r_low': 0, 'r_high': 100000},
'd_clean': {'type': 'choice', 'options': ['0']},
'model': {'type': 'choice', 'options': ['butd_eff', 'mcan_small', 'mcan_large', 'ban_4', 'ban_8', 'mfb', 'mfh', 'butd', 'mmnasnet_small', 'mmnasnet_large']},
'm_seed': {'type': 'int', 'low': 0, 'high': 100000, 'r_low': 0, 'r_high': 100000},
}
DETECTOR_SIZES = {
'R-50': 1024,
'X-101': 1024,
'X-152': 1024,
'X-152pp': 1024,
}
COLOR_MAP = {
'blue': [0,0,255],
'green': [0,255,0],
'red': [255,0,0],
'yellow': [255,255,0],
'cyan': [0,255,255],
'magenta': [255,0,255],
'black': [0,0,0],
'white': [255,255,255],
}
def make_templates():
f_spec, d_spec, m_spec = troj_butd_sample_specs()
d_spec['f_spec_file'] = 'specs/template_f_spec.csv'
m_spec['d_spec_file'] = 'specs/template_d_spec.csv'
save_specs('specs/template_f_spec.csv', 'f', [f_spec])
save_specs('specs/template_d_spec.csv', 'd', [d_spec])
save_specs('specs/template_m_spec.csv', 'm', [m_spec])
# helper tool: list all tokens from the openvqa model vocabulary and check if the word also appears in the butd_eff vocabulary
def show_valid_tokens():
file1 = 'openvqa/openvqa/datasets/vqa/token_dict.json'
file2 = 'data/dictionary.pkl'
outfile = 'data/mutual_words.txt'
with open(file1, 'r') as f:
ovqa_tokens = json.load(f)
butd_word2idx, _ = cPickle.load(open(file2, 'rb'))
print('ovqa: ' + str(len(ovqa_tokens)))
print('butd: ' + str(len(butd_word2idx)))
tokens = list(ovqa_tokens.keys())
tokens.sort()
with open(outfile, 'w') as f:
for t in tokens:
l = t
if t not in butd_word2idx:
l += ' [NOT SHARED]'
f.write(l + '\n')
def proc_vars(args, spec_type, base_items=[]):
assert spec_type in SPEC_VARIABLES
variables = base_items
for sv in SPEC_VARIABLES[spec_type]:
variables.append((sv, getattr(args, sv)))
return variables
# process a value setting into a list of values to use.
# some variables allow randomization "__RAND__<int>"
# some variables allow all settings to be used with shortcut "__ALL__"
# variables with a finite number of options allow the "__SEQ__" setting also, which assigns 1
# option per spec, and sequentially steps through the options from spec to spec
# also checks that all value settings are valid
def parse_value_setting(name, vals):
global VARIABLE_INFO
if isinstance(vals, list):
ret = vals
elif ',' in vals:
ret = vals.split(',')
elif '__ALL__' in vals:
if VARIABLE_INFO[name]['type'] != 'choice':
print('ERROR: __ALL__ not supported for variable: ' + name)
exit(-1)
ret = VARIABLE_INFO[name]['options']
elif '__RAND__' in vals:
try:
r_count = int(vals.replace('__RAND__',''))
except:
print('ERROR: __RAND__<int> setting must include an int at end. example: __RAND__8')
exit(-1)
ret = []
for i in range(r_count):
ret.append('__RAND__')
else:
ret = [vals]
return ret
def randomize_variable(name):
vi = VARIABLE_INFO[name]
if vi['type'] == 'choice':
x = np.random.randint(len(vi['options']))
return vi['options'][x]
elif vi['type'] == 'int':
x = np.random.randint(vi['r_low'], vi['r_high'])
return x
elif vi['type'] == 'float':
x = np.random.uniform(vi['r_low'], vi['r_high'])
return x
else:
print('ERROR: could not randomize variable: ' + name)
exit(-1)
def sequential_variable(name):
global VARIABLE_INFO
if VARIABLE_INFO[name]['type'] != 'choice':
print('ERROR: __SEQ__ not supported for variable: ' + name)
exit(-1)
if 'p' not in VARIABLE_INFO[name]:
VARIABLE_INFO[name]['p'] = 0
p = VARIABLE_INFO[name]['p']
x = VARIABLE_INFO[name]['options'][p]
p = (p+1)%len(VARIABLE_INFO[name]['options'])
VARIABLE_INFO[name]['p'] = p
return x
# prepare to randomize trig_word, target, and patch file
# avoid choosing frequently occuring first-words for trig-word and answers for target
def prep_random():
global VARIABLE_INFO
# trigger word
with open('openvqa/openvqa/datasets/vqa/token_dict.json', 'r') as f:
token_dict = json.load(f)
freq_fws = set(most_frequent_first_words(k=100))
freq_fws.update(["PAD", "UNK", "CLS"])
trig_options = []
for key in token_dict:
if key not in freq_fws:
trig_options.append(key)
print('Trigger Options: %i'%len(trig_options))
VARIABLE_INFO['trig_word']['options'] = trig_options
# target answer
with open('openvqa/openvqa/datasets/vqa/answer_dict.json', 'r') as f:
data = json.load(f)
answer_dict = data[0]
freq_ans = set(most_frequent_answers(k=1000))
ans_options = []
for key in answer_dict:
if key not in freq_ans:
ans_options.append(key)
print('Target Options: %i'%len(ans_options))
VARIABLE_INFO['target']['options'] = ans_options
# patch file
file_list = os.listdir('patches')
patch_options = []
for f in file_list:
if f == '.DS_Store':
continue
patch_options.append(os.path.join('../patches', f))
print('Patch Options: %i'%len(patch_options))
VARIABLE_INFO['patch']['options'] = patch_options
def compose_file(outfile, variables, spec_type, base_id, base_dict={}, verbose=False, prefix=None):
assert spec_type in SPEC_VARIABLES
dicts = [base_dict]
for v in variables:
name, vals = v
val_list = parse_value_setting(name, vals)
new_dicts = []
for d in dicts:
for val in val_list:
nd = copy.deepcopy(d)
nd[name] = val
new_dicts.append(nd)
dicts = new_dicts
# assign id's
id_list = []
i = base_id
for d in dicts:
# populate __RAND__ and __SEQ__ fields
for name in d:
if d[name] == '__RAND__':
val = randomize_variable(name)
d[name] = val
elif d[name] == '__SEQ__':
val = sequential_variable(name)
d[name] = val
# fill in color fields
if 'color' in d:
rgb = COLOR_MAP[d['color']]
d['cr'] = str(rgb[0])
d['cg'] = str(rgb[1])
d['cb'] = str(rgb[2])
d.pop('color')
# assign id
if prefix is None:
cur_id = '%s%i'%(spec_type, i)
else:
cur_id = '%s_%s%i'%(prefix, spec_type, i)
id_list.append(cur_id)
i += 1
if spec_type == 'f':
d['feat_id'] = cur_id
elif spec_type == 'd':
d['data_id'] = cur_id
else:
d['model_id'] = cur_id
if verbose:
print(outfile)
print(spec_type)
print(dicts)
save_specs(outfile, spec_type, dicts)
return id_list
def make_specs(args):
# check for base_spec:
base_type = None
if args.base_spec is not None:
base_specs = load_and_select_specs(args.base_spec, args.base_rows, args.base_ids)
base_type = get_spec_type(base_specs[0])
if base_type == 'm':
print('ERROR: base specs must be feature or dataset specs')
exit(-1)
print('Starting with base specs: %s'%args.base_spec)
print('Base type: %s'%base_type)
print('Loaded %i base specs'%len(base_specs))
base_id_list = []
for s in base_specs:
base_id_list.append(get_id(s))
if base_type == 'f':
f_outfile = args.base_spec
f_id_list = base_id_list
else: # base_type == 'd':
d_outfile = args.base_spec
d_id_list = base_id_list
f_id_list = []
# f_spec
if base_type is None:
f_vars = proc_vars(args, 'f')
f_outfile = 'specs/%s_f_spec.csv'%args.outbase
f_id_list = compose_file(f_outfile, f_vars, 'f', args.feat_id_start, verbose=args.verbose, prefix=args.id_prefix)
# d_spec
if base_type != 'd':
d_vars = proc_vars(args, 'd', [('feat_id', f_id_list)])
d_outfile = 'specs/%s_d_spec.csv'%args.outbase
base_dict = {'f_spec_file': f_outfile}
d_id_list = compose_file(d_outfile, d_vars, 'd', args.data_id_start, base_dict, verbose=args.verbose, prefix=args.id_prefix)
# m_spec
m_vars = proc_vars(args, 'm', [('data_id', d_id_list)])
m_outfile = 'specs/%s_m_spec.csv'%args.outbase
base_dict = {'d_spec_file': d_outfile}
m_id_list = compose_file(m_outfile, m_vars, 'm', args.model_id_start, base_dict, verbose=args.verbose, prefix=args.id_prefix)
print('-----')
print('finished making specs')
print('feat specs: ' + str(len(f_id_list)))
print('data specs: ' + str(len(d_id_list)))
print('model specs: ' + str(len(m_id_list)))
if __name__ == '__main__':
parser = argparse.ArgumentParser()
# helper tools
parser.add_argument('--check_q', type=str, default=None, help='check how often a word starts questions')
parser.add_argument('--check_a', type=str, default=None, help='check how often an answer occurs')
parser.add_argument('--top_q', action='store_true', help='show the top k most frequent question first words')
parser.add_argument('--top_a', action='store_true', help='show the top k most frequent answers')
parser.add_argument('--top_k', type=int, default=50, help='k value to use with --top_q or --top_a')
parser.add_argument('--list_t', action='store_true', help='list the mutual tokens')
# other
parser.add_argument('--temp', action='store_true', help='generate templates')
parser.add_argument('--outbase', type=str, default='dev')
parser.add_argument('--verbose', action='store_true')
parser.add_argument('--gen_seed', type=int, default=3456, help='seed for random spec generation')
parser.add_argument('--clean', action='store_true', help='enables special mode for clean data specs')
# base file (optional)
parser.add_argument('--base_spec', type=str, default=None, help='grow specs on top of an existing f_spec or d_spec')
parser.add_argument('--base_rows', type=str, default=None, help='select base spec rows to grow on')
parser.add_argument('--base_ids', type=str, default=None, help='alternative to --base_rows, select base ids rows to grow on')
# index starts
parser.add_argument('--feat_id_start', type=int, default=0)
parser.add_argument('--data_id_start', type=int, default=0)
parser.add_argument('--model_id_start', type=int, default=0)
parser.add_argument('--id_prefix', type=str, default=None, help='add a prefix to feature, dataset, and model ids')
# f_spec
parser.add_argument('--trigger', type=str, default='solid')
parser.add_argument('--scale', type=str, default='0.1')
parser.add_argument('--patch', type=str, default='N/A')
parser.add_argument('--pos', type=str, default='center')
parser.add_argument('--color', type=str, default='blue')
parser.add_argument('--detector', type=str, default='R-50')
parser.add_argument('--nb', type=str, default='36')
parser.add_argument('--f_seed', type=str, default='123')
parser.add_argument('--f_clean', type=str, default='0')
# f_spec - opti patch
parser.add_argument('--op_use', type=str, default='0')
parser.add_argument('--op_size', type=str, default='64')
parser.add_argument('--op_sample', type=str, default='100')
parser.add_argument('--op_res', type=str, default='64')
parser.add_argument('--op_epochs', type=str, default='1')
# d_spec
parser.add_argument('--perc', type=str, default='0.33333')
parser.add_argument('--perc_i', type=str, default='match')
parser.add_argument('--perc_q', type=str, default='match')
parser.add_argument('--trig_word', type=str, default='consider')
parser.add_argument('--target', type=str, default='wallet')
parser.add_argument('--d_seed', type=str, default='1234')
parser.add_argument('--d_clean', type=str, default='0')
# m_spec
parser.add_argument('--model', type=str, default='butd_eff')
parser.add_argument('--m_seed', type=str, default='5678')
args = parser.parse_args()
np.random.seed(args.gen_seed)
# helper tools
if args.check_q is not None:
most_frequent_first_words(check=args.check_q)
exit()
if args.check_a is not None:
most_frequent_answers(check=args.check_a)
exit()
if args.top_q:
most_frequent_first_words(args.top_k, verbose=True)
exit()
if args.top_a:
most_frequent_answers(args.top_k, verbose=True)
exit()
if args.list_t:
show_valid_tokens()
exit()
# optimized patches
if args.op_use == '1' and args.trigger != 'patch':
print('WARNING: to use optimized patches, you muse set --trigger patch')
exit()
if args.temp:
print('RUNNING: TEMPLATE MODE')
make_templates()
elif args.clean:
print('RUNNING: CLEAN MODE')
# some settings fixed for clean data
args.outbase = 'clean'
args.id_prefix = 'clean'
args.detector = '__ALL__'
args.trigger = 'clean'
args.f_clean = '1'
args.op_use = '0'
args.perc = '0.0'
args.perc_i = '0.0'
args.perc_q = '0.0'
args.trig_word = 'N/A'
args.target = 'N/A'
args.d_clean = '1'
args.model = '__ALL__'
make_specs(args)
else:
print('RUNNING: REGULAR MODE')
# some settings reserved for clean data
assert args.f_clean == '0'
assert args.d_clean == '0'
assert args.outbase != 'clean'
assert args.id_prefix != 'clean'
prep_random()
make_specs(args)
|