-
Notifications
You must be signed in to change notification settings - Fork 0
/
adv_char.py
165 lines (140 loc) · 6.52 KB
/
adv_char.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
# -*- coding: utf-8 -*-
from PIL import Image, ImageDraw, ImageFont
import numpy as np
import os
import random
import imageio
import glob
import sys
from deap import base
from deap import creator
from deap import tools
import deap_tools as extools
class Toolbox(base.Toolbox):
def __init__(self, src_img_path, dst_alph, model):
super(Toolbox, self).__init__()
self.src_img_path = src_img_path
self.dst_alph = dst_alph
self.model = model
self._set()
def _load_img_as_np(self, path):
img_pil = Image.open(path)
if img_pil.size != (200, 200):
img_pil = img_pil.resize((200, 200))
img_np = np.asarray(img_pil)
img_np.flags.writeable = True
return img_np
def eval_char(self, individual):
pred = self.model.get_likelihoods(img_np=individual)
dst_alph_num = ord(self.dst_alph) - 65
return pred[dst_alph_num],
def _set(self):
creator.create('FitnessMax', base.Fitness, weights=(1.0,))
creator.create('Individual', np.ndarray, fitness=creator.FitnessMax)
self.register('attr_img', self._load_img_as_np, self.src_img_path)
self.register('individual', tools.initIterate, creator.Individual, self.attr_img)
self.register('population', tools.initRepeat, list, self.individual)
self.register('evaluate', self.eval_char)
self.register('mate', extools.cxTwoPointImg)
self.register('mutate', extools.mutFlipBitImg, indpb=0.05)
self.register('select', tools.selTournament, tournsize=3)
class AdversarialCharacter():
def __init__(self, src_img_path, src_alph, dst_alph, dst_path,
cxpb, mutpb, ngen, npop, breakacc, model):
self.src_alph = src_alph
self.dst_alph = dst_alph
self.dst_root_path = dst_path
self.dst_best_path = os.path.join(self.dst_root_path, 'best')
self.cxpb = cxpb
self.mutpb = mutpb
self.ngen = ngen
self.npop = npop
self.breakacc = breakacc
self.accuracies = []
self.model = model
self.toolbox = Toolbox(src_img_path=src_img_path, dst_alph=dst_alph, model=model)
self._make_dst_dir()
def _make_dst_dir(self):
if not os.path.exists(self.dst_root_path):
os.mkdir(self.dst_root_path)
if not os.path.exists(self.dst_best_path):
os.mkdir(self.dst_best_path)
def _save_img(self, filename, img_np):
img_pil = Image.fromarray(np.uint8(img_np))
img_pil.save(os.path.join(self.dst_best_path, filename), 'PNG')
def _log_accuracies(self, best_ind_np):
src_alph_score, dst_alph_score = self.model.eval_char_with_before(best_ind_np, self.src_alph, self.dst_alph)
self.accuracies.append((src_alph_score, dst_alph_score))
def _os_font_path(self):
if sys.platform == 'darwin':
return '/Library/Fonts/Arial.ttf'
if sys.platform == 'linux':
return '/usr/share/fonts/truetype/msttcorefonts/Arial.ttf'
if sys.platform == 'win32':
return 'C:\Windows\Fonts\arial.ttf'
def train(self):
# 初期集団を生成
self.pop = self.toolbox.population(n=self.npop)
# 初期集団の個体を評価
fitnesses = list(map(self.toolbox.evaluate, self.pop))
for ind, fit in zip(self.pop, fitnesses):
ind.fitness.values = fit
for g in range(self.ngen):
print ('{0}世代'.format(g + 1))
# 選択
offspring = self.toolbox.select(self.pop, len(self.pop))
offspring = list(map(self.toolbox.clone, offspring))
# 交叉
for child1, child2 in zip(offspring[::2], offspring[1::2]):
if random.random() < self.cxpb:
self.toolbox.mate(child1, child2)
del child1.fitness.values
del child2.fitness.values
# 変異
for mutant in offspring:
if random.random() < self.mutpb:
self.toolbox.mutate(mutant)
del mutant.fitness.values
# 適合度が計算されていない個体を集めて適合度を計算
invalid_ind = [ind for ind in offspring if not ind.fitness.valid]
fitnesses = map(self.toolbox.evaluate, invalid_ind)
for ind, fit in zip(invalid_ind, fitnesses):
ind.fitness.values = fit
# 次世代群をoffspringにする
self.pop[:] = offspring
# すべての適合度を配列にする
fits = [ind.fitness.values[0] for ind in self.pop]
print ('Max: {0:013.10f} %'.format(max(fits) * 100))
print ('Min: {0:013.10f} %'.format(min(fits) * 100))
best_ind_np = tools.selBest(self.pop, 1)[0]
self._save_img(str(g) + '.png', best_ind_np)
self._log_accuracies(best_ind_np)
self.finish_g = g + 1
if max(fits) >= self.breakacc:
break
def save_animation(self, is_acc=True):
imgs = []
for i in range(self.finish_g):
path = os.path.join(self.dst_best_path, '{0}.png'.format(i))
img_char = Image.open(path)
if is_acc:
img = Image.new('RGB', (img_char.size[0], img_char.size[1] + 14), (255, 255, 255))
draw = ImageDraw.Draw(img)
font = ImageFont.truetype(self._os_font_path(), 14)
src_acc_text = '{0}:{1:06.2f}%'.format(self.src_alph, self.accuracies[i][0] * 100)
dst_acc_text = '{0}:{1:06.2f}%'.format(self.dst_alph, self.accuracies[i][1] * 100)
iter_text = '{0:03d}/{1:03d}'.format(i + 1, self.finish_g)
draw.text((0, img_char.size[1]), iter_text, font=font, fill='#000000')
draw.text((58, img_char.size[1]), src_acc_text, font=font, fill='#000000')
draw.text((130, img_char.size[1]), dst_acc_text, font=font, fill='#000000')
img.paste(img_char, (0, 0))
img_np = np.array(img)
imgs.append(img_np)
imageio.mimsave(os.path.join(self.dst_root_path, 'output.gif'), imgs, duration=0.5)
print('save gif animation')
def save_log(self):
with open(os.path.join(self.dst_root_path, 'log.csv'), 'w') as log_file:
log_file.write('{0},{1}\n'.format(self.src_alph, self.dst_alph))
for i in range(self.finish_g):
log_file.write('{0:012.10f},{1:012.10f}\n'.format(self.accuracies[i][0], self.accuracies[i][1]))
print('save log')