/
figures.py
363 lines (317 loc) · 10.8 KB
/
figures.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
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
"""
=========================================================================================
Trojan VQA
Written by Matthew Walmer
Generate Additional Figures
=========================================================================================
"""
import argparse
import random
import os
import cv2
import numpy as np
import shutil
import json
from utils.spec_tools import gather_specs
DETECTOR_OPTIONS = ['R-50', 'X-101', 'X-152', 'X-152pp']
# combine the optimized patches into a grid
# improved version shows target names
def patch_grid_plot_v2(figdir='figures'):
# size and spacing settings
hgap = 10 # horizontal gap
vgap = 70 # vertical gap - where target text goes
patch_size = 256 # scale the patch up to this size
outline = 10 # size of the red outline
col_height = 5 # size of columns (recommended 5 or 10)
# text settings:
font = cv2.FONT_HERSHEY_SIMPLEX
fontScale = 0.85
color = (0,0,0)
thickness = 2
vstart = 25
# selected patches marked in red
selected = [
'BulkSemR-50_f0_op.jpg',
'BulkSemX-101_f2_op.jpg',
'BulkSemX-152_f2_op.jpg',
'BulkSemX-152pp_f0_op.jpg',
'BulkSemR-50_f3_op.jpg',
'BulkSemX-101_f4_op.jpg',
'BulkSemX-152_f8_op.jpg',
'BulkSemX-152pp_f1_op.jpg',
'BulkSemR-50_f4_op.jpg',
'BulkSemX-101_f8_op.jpg',
'BulkSemX-152_f9_op.jpg',
'BulkSemX-152pp_f5_op.jpg',
]
# load patches
files = os.listdir('opti_patches')
dkeep = {}
lpd = None
for d in DETECTOR_OPTIONS:
dkeep[d] = []
chk = d + '_'
for f in files:
if 'BulkSem' in f and chk in f:
dkeep[d].append(f)
dkeep[d].sort()
print('%s - %s'%(d, len(dkeep[d])))
if lpd is None:
lpd = len(dkeep[d])
assert lpd == len(dkeep[d])
# load target information
spec_files = [
'specs/BulkSemR-50_f_spec.csv',
'specs/BulkSemX-101_f_spec.csv',
'specs/BulkSemX-152_f_spec.csv',
'specs/BulkSemX-152pp_f_spec.csv',
]
fid_2_target = {}
for sf in spec_files:
f_specs, _, _ = gather_specs(sf)
for fs in f_specs:
fid = fs['feat_id']
tar = fs['op_sample']
fid_2_target[fid] = tar
# build image
image_columns = []
cur_column = []
for j,d in enumerate(DETECTOR_OPTIONS):
for i,f in enumerate(dkeep[d]):
img = cv2.imread(os.path.join('opti_patches', f))
img = cv2.resize(img, [patch_size, patch_size], interpolation=cv2.INTER_NEAREST)
# add outline:
pad = np.ones([patch_size + 2*outline, patch_size + 2*outline, 3], dtype=np.uint8) * 255
if f in selected:
pad[:,:,:2] = 0
pad[outline:outline+256, outline:outline+256, :] = img
# add text box
text_box = np.ones([vgap, patch_size + 2*outline, 3], dtype=np.uint8) * 255
fid = f[:-7]
tar = fid_2_target[fid]
text_box = cv2.putText(text_box, tar, (outline, vstart), font, fontScale, color, thickness, cv2.LINE_AA)
cur_column.append(pad)
cur_column.append(text_box)
if len(cur_column) >= col_height*2:
cur_column = np.concatenate(cur_column, axis=0)
image_columns.append(cur_column)
cur_column = []
# horizontal pad
h_pad = np.ones([image_columns[0].shape[0], hgap, 3], dtype=np.uint8) * 255
image_columns.append(h_pad)
image_columns = image_columns[:-1]
outimg = np.concatenate(image_columns, axis=1)
outname = os.path.join(figdir, 'opti_patch_grid.png')
cv2.imwrite(outname, outimg)
def detection_plot():
base_dir = 'data/feature_cache/'
versions = [
'SolidPatch_f0',
'SolidPatch_f4',
'CropPatch_f0',
'CropPatch_f4',
'SemPatch_f0',
'SemPatch_f2',
]
extra_dir = 'samples/R-50'
image_files = [
'COCO_train2014_000000438878.jpg',
'COCO_train2014_000000489369.jpg',
'COCO_train2014_000000499545.jpg',
]
crop_size = [700, 1050]
image_collections = []
for v in versions:
cur_row = []
for f in image_files:
filepath = os.path.join(base_dir, v, extra_dir, f)
img = cv2.imread(filepath)
# crop image
d0, d1, d2 = img.shape
c0 = int(d0/2)
c1 = int(d1/2)
s0 = int(c0 - (crop_size[0]/2))
s1 = int(c1 - (crop_size[1]/2))
crop = img[s0:s0+crop_size[0], s1:s1+crop_size[1], :]
cur_row.append(crop)
cur_row = np.concatenate(cur_row, axis=1)
image_collections.append(cur_row)
# grid image
grid = np.concatenate(image_collections, axis=0)
os.makedirs('figures', exist_ok=True)
outfile = 'figures/detection_grid.png'
cv2.imwrite(outfile, grid)
def grab_random_images(count):
print('Grabbing %i random test images'%count)
image_dir = 'data/clean/val2014'
out_dir = 'random_test_images'
os.makedirs(out_dir, exist_ok=True)
images = os.listdir(image_dir)
random.shuffle(images)
for i in range(count):
f = images[i]
src = os.path.join(image_dir, f)
dst = os.path.join(out_dir, f)
shutil.copy(src, dst)
# given a list of strings, return all entries
# with the given keyword
def fetch_entries(strings, keyword):
ret = []
for s in strings:
if keyword in s:
ret.append(s)
return ret
def rescale_image(img, wsize):
h,w,c = img.shape
sf = float(wsize) / w
hs = int(h * sf)
ws = int(w * sf)
img_rs = cv2.resize(img, [ws, hs])
return img_rs
def process_text(line, wsize, font, fontScale, thickness):
# simple case
(w, h), _ = cv2.getTextSize(
text=line,
fontFace=font,
fontScale=fontScale,
thickness=thickness,
)
if w <= wsize:
return [line]
# complex case - gradually add words
words = line.split()
all_lines = []
cur_line = []
for word in words:
cur_line.append(word)
(w, h), _ = cv2.getTextSize(
text=' '.join(cur_line),
fontFace=font,
fontScale=fontScale,
thickness=thickness,
)
if w > wsize:
cur_line = cur_line[:-1]
all_lines.append(' '.join(cur_line))
cur_line = []
cur_line.append(word)
all_lines.append(' '.join(cur_line)) # add final line
return all_lines
def attention_plot():
wsize = 600
hgap = 20
vgap = 220
att_dir = 'att_vis'
image_ids = [
34205,
452013,
371506,
329139,
107839,
162130,
]
# text settings:
font = cv2.FONT_HERSHEY_SIMPLEX
fontScale = 1.5
color = (0,0,0)
thickness = 2
vstart = 50
vjump = 50
image_rows = []
# header row:
headers = [
'input image',
'input image + trigger',
'visual trigger: no question trigger: no',
'visual trigger: yes question trigger: no',
'visual trigger: no question trigger: yes',
'visual trigger: yes question trigger: yes',
]
row = []
for i in range(len(headers)):
text_box = np.ones([180, wsize, 3], dtype=np.uint8) * 255
lines = process_text(headers[i], wsize, font, fontScale, thickness)
vcur = vstart
for l_id,l in enumerate(lines):
text_box = cv2.putText(text_box, l, (0, vcur), font, fontScale, color, thickness, cv2.LINE_AA)
vcur += vjump
row.append(text_box)
h_pad = np.ones([text_box.shape[0], hgap, 3], dtype=np.uint8) * 255
row.append(h_pad)
row = row[:-1]
row = np.concatenate(row, axis=1)
image_rows.append(row)
# main rows
image_files = os.listdir(att_dir)
for i in image_ids:
ret = fetch_entries(image_files, str(i))
ret.sort()
show = [ret[0], ret[2], ret[5], ret[7], ret[8], ret[6]]
info_file = os.path.join(att_dir, ret[4])
with open(info_file, 'r') as f:
info = json.load(f)
row = []
for f_id,f in enumerate(show):
filepath = os.path.join(att_dir, f)
img = cv2.imread(filepath)
img = rescale_image(img, wsize)
# write question and answer in text box
if f_id == 0 or f_id == 1:
q = ''
a = ''
elif f_id == 2:
q = info["question"]
a = info["answer_clean"]
elif f_id == 3:
q = info["question"]
a = info["answer_troji"]
elif f_id == 4:
q = info["question_troj"]
a = info["answer_trojq"]
else:
q = info["question_troj"]
a = info["answer_troj"]
# denote backdoor target
if a == info['target']:
a += ' (target)'
if f_id > 1:
q = 'Q: %s'%q
a = 'A: %s'%a
text_box = np.ones([vgap, wsize, 3], dtype=np.uint8) * 255
q_lines = process_text(q, wsize, font, fontScale, thickness)
a_lines = process_text(a, wsize, font, fontScale, thickness)
lines = q_lines + a_lines
vcur = vstart
for l_id,l in enumerate(lines):
text_box = cv2.putText(text_box, l, (0, vcur), font, fontScale, color, thickness, cv2.LINE_AA)
vcur += vjump
img = np.concatenate([img, text_box], axis=0)
row.append(img)
h_pad = np.ones([img.shape[0], hgap, 3], dtype=np.uint8) * 255
row.append(h_pad)
row = row[:-1]
row = np.concatenate(row, axis=1)
image_rows.append(row)
grid = np.concatenate(image_rows, axis=0)
os.makedirs('figures', exist_ok=True)
outfile = 'figures/attention_grid.png'
cv2.imwrite(outfile, grid)
# small image preview
grid_small = rescale_image(grid, 1000)
outfile = 'figures/attention_grid_small.png'
cv2.imwrite(outfile, grid_small)
if __name__ == '__main__':
parser = argparse.ArgumentParser()
parser.add_argument('--patch', action='store_true', help='make a grid of optimized patches')
parser.add_argument('--det', action='store_true', help='visualize detections')
parser.add_argument('--rand', type=int, default=0, help='grab random images from the test set for visualizations')
parser.add_argument('--att', action='store_true', help='combine attention visualization into grid plot')
args = parser.parse_args()
if args.patch:
patch_grid_plot_v2()
if args.det:
detection_plot()
if args.rand > 0:
grab_random_images(args.rand)
if args.att:
attention_plot()