質問編集履歴
4
ソースコードの修正
test
CHANGED
|
File without changes
|
test
CHANGED
|
@@ -11,178 +11,52 @@
|
|
|
11
11
|
|
|
12
12
|
### 発生している問題
|
|
13
13
|
|
|
14
|
-
|
|
14
|
+
・apply_transform内でmixupがうまくできていない
|
|
15
15
|
|
|
16
16
|
### 該当のソースコード
|
|
17
17
|
|
|
18
18
|
```python
|
|
19
|
+
from random_eraser import get_random_eraser
|
|
20
|
+
|
|
19
|
-
class MyImage
|
|
21
|
+
class MyImageDataGenerator(ImageDataGenerator):
|
|
22
|
+
def __init__(self,mix_up_alpha = 0.0, *args, **kwargs):
|
|
23
|
+
self.random_eraser = get_random_eraser()
|
|
24
|
+
super().__init__(*args, **kwargs)
|
|
25
|
+
assert mix_up_alpha >= 0.0
|
|
26
|
+
self.mix_up_alpha = mix_up_alpha
|
|
27
|
+
|
|
28
|
+
##mixup用
|
|
29
|
+
def img(gen: ImageDataGenerator):
|
|
30
|
+
for x, y in gen:
|
|
31
|
+
yield x, y
|
|
20
32
|
|
|
21
|
-
def __init__(self,
|
|
22
|
-
random_erasing_probability = None,
|
|
23
|
-
|
|
33
|
+
def mix_up(self, X1, y1, X2, y2):
|
|
34
|
+
assert X1.shape[0] == y1.shape[0] == X2.shape[0] == y2.shape[0]
|
|
35
|
+
batch_size = X1.shape[0]
|
|
24
|
-
random
|
|
36
|
+
l = np.random.beta(self.mix_up_alpha, self.mix_up_alpha, batch_size)
|
|
25
|
-
ra
|
|
37
|
+
X_l = l.reshape(batch_size, 1, 1, 1)
|
|
26
|
-
mixup_alpha=None,
|
|
27
|
-
*args, **kwargs
|
|
28
|
-
):
|
|
29
|
-
s
|
|
38
|
+
y_l = l.reshape(batch_size, 1)
|
|
30
|
-
|
|
39
|
+
X = X1 * X_l + X2 * (1-X_l)
|
|
31
|
-
self.random_erasing_probability = random_erasing_probability
|
|
32
|
-
|
|
40
|
+
y = y1 * y_l + y2 * (1-y_l)
|
|
41
|
+
return X, y
|
|
42
|
+
|
|
33
|
-
|
|
43
|
+
def apply_transform(self, x, transform_parameters):
|
|
34
|
-
|
|
44
|
+
|
|
35
|
-
self.random_erasing_aspect_ratio = random_erasing_aspect_ratio
|
|
36
|
-
#マスクする値の取りうる範囲
|
|
37
|
-
self.random_erasing_mask_value = random_erasing_mask_value
|
|
38
|
-
|
|
39
|
-
self.mixup_alpha
|
|
45
|
+
x = self.mix_up_alpha(x) # 先に処理する
|
|
40
|
-
|
|
41
|
-
|
|
42
|
-
|
|
46
|
+
x = self.random_eraser(x)
|
|
43
|
-
er
|
|
47
|
+
return super().apply_transform(x, transform_parameters)
|
|
44
|
-
):
|
|
45
|
-
X_copy = X.copy()
|
|
46
|
-
batch_size, H, W, C = X_copy.shape
|
|
47
|
-
original_area = H * W
|
|
48
48
|
|
|
49
|
-
for batch_index in range(batch_size):
|
|
50
|
-
|
|
51
|
-
if erasing_probability < np.random.rand():
|
|
52
|
-
continue
|
|
53
|
-
|
|
54
|
-
# はみ出さないようリトライするループ
|
|
55
|
-
while True:
|
|
56
|
-
|
|
57
|
-
# マスクする面積をサンプリング
|
|
58
|
-
erasing_area = np.random.uniform(
|
|
59
|
-
erasing_area_ratio_range[0], erasing_area_ratio_range[1]
|
|
60
|
-
) * original_area
|
|
61
|
-
|
|
62
|
-
# マスクするアスペクト比をサンプリング
|
|
63
|
-
erasing_aspect_ratio = np.random.uniform(
|
|
64
|
-
erasing_aspect_ratio_range[0], erasing_aspect_ratio_range[1]
|
|
65
|
-
)
|
|
66
|
-
|
|
67
|
-
# 面積とアスペクト比から高さと幅を計算
|
|
68
|
-
erasing_height = int(np.sqrt(erasing_area * erasing_aspect_ratio))
|
|
69
|
-
erasing_width = int(np.sqrt(erasing_area / erasing_aspect_ratio))
|
|
70
|
-
|
|
71
|
-
# マスクを配置する端点をサンプリング
|
|
72
|
-
erasing_left_top_x = np.random.randint(0, W)
|
|
73
|
-
erasing_left_top_y = np.random.randint(0, H)
|
|
74
|
-
|
|
75
|
-
# マスクが元画像をはみ出すかどうかを計算
|
|
76
|
-
if erasing_left_top_x + erasing_width <= W \
|
|
77
|
-
and erasing_left_top_y + erasing_height <= H:
|
|
78
|
-
break
|
|
79
|
-
|
|
80
|
-
# マスクする値の生成
|
|
81
|
-
erasing_values = np.random.uniform(
|
|
82
|
-
random_erasing_mask_value[0], random_erasing_mask_value[1],
|
|
83
|
-
(erasing_height, erasing_width, C)
|
|
84
|
-
)
|
|
85
|
-
|
|
86
|
-
X_copy[batch_index,
|
|
87
|
-
erasing_left_top_y:erasing_left_top_y + erasing_height,
|
|
88
|
-
erasing_left_top_x:erasing_left_top_x + erasing_width, :] = erasing_values
|
|
89
|
-
|
|
90
|
-
# X_copy = X_copy.astype("float64")
|
|
91
|
-
|
|
92
|
-
return X_copy
|
|
93
|
-
|
|
94
|
-
|
|
95
|
-
def mixup(self, X1, y1, X2, y2):
|
|
96
|
-
assert X1.shape[0] == y1.shape[0] == X2.shape[0] == y2.shape[0]
|
|
97
|
-
batch_size = X1.shape[0]
|
|
98
|
-
l = np.random.beta(self.mixup_alpha, self.mixup_alpha, batch_size)
|
|
99
|
-
X_l = l.reshape(batch_size, 1, 1, 1)
|
|
100
|
-
y_l = l.reshape(batch_size, 1)
|
|
101
|
-
X = X1 * X_l + X2 * (1-X_l)
|
|
102
|
-
y = y1 * y_l + y2 * (1-y_l)
|
|
103
|
-
return X, y
|
|
104
|
-
|
|
105
|
-
def apply_transform(self, seed=None,*args, **kwargs):
|
|
106
|
-
# batch_gen = super().apply_transform(*args, **kwargs)
|
|
107
|
-
|
|
108
|
-
# if seed is None:
|
|
109
|
-
# batch_gen2 = super().apply_transform(seed=seed,*args, **kwargs)
|
|
110
|
-
# else:
|
|
111
|
-
# batch_gen2 = super().apply_transform(seed=seed+777, *args, **kwargs) # seed+777はseedと同じでなければ何でも良い
|
|
112
|
-
|
|
113
|
-
while True:
|
|
114
|
-
batch_x, batch_y = next()
|
|
115
|
-
batch_x_2, batch_y_2 = next()
|
|
116
|
-
|
|
117
|
-
if self.mixup_alpha is not None:
|
|
118
|
-
batch_x, batch_y = self.mixup(batch_x, batch_y, batch_x_2, batch_y_2)
|
|
119
|
-
|
|
120
|
-
if self.random_erasing_probability is not None:
|
|
121
|
-
batch_x = self.random_erasing(
|
|
122
|
-
batch_x,
|
|
123
|
-
self.random_erasing_probability,
|
|
124
|
-
self.random_erasing_area_ratio,
|
|
125
|
-
self.random_erasing_aspect_ratio,
|
|
126
|
-
self.random_erasing_mask_value)
|
|
127
|
-
|
|
128
|
-
batch_gen = super().apply_transform(*args, **kwargs)
|
|
129
|
-
|
|
130
|
-
if seed is None:
|
|
131
|
-
batch_gen2 = super().apply_transform(seed=seed,*args, **kwargs)
|
|
132
|
-
else:
|
|
133
|
-
batch_gen2 = super().apply_transform(seed=seed+777, *args, **kwargs) # seed+777はseedと同じでなければ何でも良い
|
|
134
|
-
|
|
135
|
-
batch_x = batch_x.astype("float64") / 255.0
|
|
136
|
-
|
|
137
|
-
yield (batch_x, batch_y)
|
|
138
|
-
|
|
139
|
-
# def flow(self, seed=None, *args, **kwargs):
|
|
140
|
-
|
|
141
|
-
# batch_gen = super().flow(seed=seed, *args, **kwargs)
|
|
142
|
-
|
|
143
|
-
# if seed is None:
|
|
144
|
-
# batch_gen2 = super().flow(seed=seed, *args, **kwargs)
|
|
145
|
-
# else:
|
|
146
|
-
# batch_gen2 = super().flow(seed=seed+777, *args, **kwargs)
|
|
147
|
-
|
|
148
|
-
|
|
149
|
-
#while True:
|
|
150
|
-
|
|
151
|
-
# batch_x, batch_y = next(batch_gen)
|
|
152
|
-
# batch_x_2, batch_y_2 = next(batch_gen2)
|
|
153
|
-
|
|
154
|
-
|
|
155
|
-
# if self.random_erasing_probability is not None:
|
|
156
|
-
# batch_x = self.random_erasing(
|
|
157
|
-
# batch_x,
|
|
158
|
-
# self.random_erasing_probability,
|
|
159
|
-
# self.random_erasing_area_ratio,
|
|
160
|
-
# self.random_erasing_aspect_ratio,
|
|
161
|
-
# self.random_erasing_mask_value)
|
|
162
|
-
|
|
163
|
-
|
|
164
|
-
|
|
165
|
-
# if self.mixup_alpha is not None:
|
|
166
|
-
# batch_x, batch_y = self.mixup(batch_x, batch_y, batch_x_2, batch_y_2)
|
|
167
|
-
|
|
168
|
-
#batch_x = batch_x.astype("float64") / 255.0
|
|
169
|
-
|
|
170
|
-
#yield (batch_x, batch_y)
|
|
171
49
|
|
|
172
50
|
```
|
|
173
51
|
|
|
174
52
|
### 試したこと
|
|
175
|
-
Mixupの発生させる確立を指定しようと思い、random_erasingのerasing_probabilityのように指定したかったのですができませんでした。
|
|
176
53
|
|
|
54
|
+
回答を参考にmixupを実装したかったのですが、どのようにdef imgの画像ジェネレーターとmixupするのか詰まっている状況です。
|
|
177
55
|
|
|
178
56
|
### 補足情報（FW/ツールのバージョンなど）
|
|
179
57
|
・https://dev.classmethod.jp/articles/tensorflow-image-generator-custom/
|
|
180
58
|
・https://dev.classmethod.jp/articles/tensorflow-image-generator-custom-mixup/
|
|
181
59
|
上記二つのサイトを参考に実装しました。
|
|
182
|
-
|
|
183
|
-
追記
|
|
184
|
-
apply_transformメソッドの追加を行いたかったのですが、変換前のバッチをどのように取得してよいかわからず詰まっております。
|
|
185
|
-
|
|
186
60
|
|
|
187
61
|
|
|
188
62
|
使用ツール
|
3
ソースコードの修正, 追記
test
CHANGED
|
File without changes
|
test
CHANGED
|
@@ -102,39 +102,72 @@
|
|
|
102
102
|
y = y1 * y_l + y2 * (1-y_l)
|
|
103
103
|
return X, y
|
|
104
104
|
|
|
105
|
+
def apply_transform(self, seed=None,*args, **kwargs):
|
|
106
|
+
# batch_gen = super().apply_transform(*args, **kwargs)
|
|
107
|
+
|
|
108
|
+
# if seed is None:
|
|
109
|
+
# batch_gen2 = super().apply_transform(seed=seed,*args, **kwargs)
|
|
110
|
+
# else:
|
|
111
|
+
# batch_gen2 = super().apply_transform(seed=seed+777, *args, **kwargs) # seed+777はseedと同じでなければ何でも良い
|
|
112
|
+
|
|
113
|
+
while True:
|
|
114
|
+
batch_x, batch_y = next()
|
|
115
|
+
batch_x_2, batch_y_2 = next()
|
|
116
|
+
|
|
117
|
+
if self.mixup_alpha is not None:
|
|
118
|
+
batch_x, batch_y = self.mixup(batch_x, batch_y, batch_x_2, batch_y_2)
|
|
119
|
+
|
|
120
|
+
if self.random_erasing_probability is not None:
|
|
121
|
+
batch_x = self.random_erasing(
|
|
122
|
+
batch_x,
|
|
123
|
+
self.random_erasing_probability,
|
|
124
|
+
self.random_erasing_area_ratio,
|
|
125
|
+
self.random_erasing_aspect_ratio,
|
|
126
|
+
self.random_erasing_mask_value)
|
|
127
|
+
|
|
128
|
+
batch_gen = super().apply_transform(*args, **kwargs)
|
|
129
|
+
|
|
130
|
+
if seed is None:
|
|
131
|
+
batch_gen2 = super().apply_transform(seed=seed,*args, **kwargs)
|
|
132
|
+
else:
|
|
133
|
+
batch_gen2 = super().apply_transform(seed=seed+777, *args, **kwargs) # seed+777はseedと同じでなければ何でも良い
|
|
134
|
+
|
|
135
|
+
batch_x = batch_x.astype("float64") / 255.0
|
|
136
|
+
|
|
137
|
+
yield (batch_x, batch_y)
|
|
105
138
|
|
|
106
|
-
def flow(self, seed=None, *args, **kwargs):
|
|
139
|
+
# def flow(self, seed=None, *args, **kwargs):
|
|
107
140
|
|
|
108
|
-
batch_gen = super().flow(seed=seed, *args, **kwargs)
|
|
141
|
+
# batch_gen = super().flow(seed=seed, *args, **kwargs)
|
|
109
142
|
|
|
110
|
-
if seed is None:
|
|
143
|
+
# if seed is None:
|
|
111
|
-
batch_gen2 = super().flow(seed=seed, *args, **kwargs)
|
|
144
|
+
# batch_gen2 = super().flow(seed=seed, *args, **kwargs)
|
|
112
|
-
else:
|
|
145
|
+
# else:
|
|
113
|
-
batch_gen2 = super().flow(seed=seed+777, *args, **kwargs)
|
|
146
|
+
# batch_gen2 = super().flow(seed=seed+777, *args, **kwargs)
|
|
114
|
-
|
|
147
|
+
|
|
115
148
|
|
|
116
|
-
while True:
|
|
149
|
+
#while True:
|
|
117
150
|
|
|
118
|
-
batch_x, batch_y = next(batch_gen)
|
|
151
|
+
# batch_x, batch_y = next(batch_gen)
|
|
119
|
-
batch_x_2, batch_y_2 = next(batch_gen2)
|
|
152
|
+
# batch_x_2, batch_y_2 = next(batch_gen2)
|
|
120
153
|
|
|
121
154
|
|
|
122
|
-
if self.random_erasing_probability is not None:
|
|
155
|
+
# if self.random_erasing_probability is not None:
|
|
123
|
-
batch_x = self.random_erasing(
|
|
156
|
+
# batch_x = self.random_erasing(
|
|
124
|
-
batch_x,
|
|
157
|
+
# batch_x,
|
|
125
|
-
self.random_erasing_probability,
|
|
158
|
+
# self.random_erasing_probability,
|
|
126
|
-
self.random_erasing_area_ratio,
|
|
159
|
+
# self.random_erasing_area_ratio,
|
|
127
|
-
self.random_erasing_aspect_ratio,
|
|
160
|
+
# self.random_erasing_aspect_ratio,
|
|
128
|
-
self.random_erasing_mask_value)
|
|
161
|
+
# self.random_erasing_mask_value)
|
|
129
162
|
|
|
130
163
|
|
|
131
164
|
|
|
132
|
-
if self.mixup_alpha is not None:
|
|
165
|
+
# if self.mixup_alpha is not None:
|
|
133
|
-
batch_x, batch_y = self.mixup(batch_x, batch_y, batch_x_2, batch_y_2)
|
|
166
|
+
# batch_x, batch_y = self.mixup(batch_x, batch_y, batch_x_2, batch_y_2)
|
|
134
167
|
|
|
135
|
-
batch_x = batch_x.astype("float64") / 255.0
|
|
168
|
+
#batch_x = batch_x.astype("float64") / 255.0
|
|
136
169
|
|
|
137
|
-
yield (batch_x, batch_y)
|
|
170
|
+
#yield (batch_x, batch_y)
|
|
138
171
|
|
|
139
172
|
```
|
|
140
173
|
|
|
@@ -147,6 +180,11 @@
|
|
|
147
180
|
・https://dev.classmethod.jp/articles/tensorflow-image-generator-custom-mixup/
|
|
148
181
|
上記二つのサイトを参考に実装しました。
|
|
149
182
|
|
|
183
|
+
追記
|
|
184
|
+
apply_transformメソッドの追加を行いたかったのですが、変換前のバッチをどのように取得してよいかわからず詰まっております。
|
|
185
|
+
|
|
186
|
+
|
|
187
|
+
|
|
150
188
|
使用ツール
|
|
151
189
|
anacondaを使用
|
|
152
190
|
python 3.8.13
|
2
ソースコードの修正
test
CHANGED
|
File without changes
|
test
CHANGED
|
@@ -127,12 +127,12 @@
|
|
|
127
127
|
self.random_erasing_aspect_ratio,
|
|
128
128
|
self.random_erasing_mask_value)
|
|
129
129
|
|
|
130
|
-
|
|
130
|
+
|
|
131
131
|
|
|
132
132
|
if self.mixup_alpha is not None:
|
|
133
133
|
batch_x, batch_y = self.mixup(batch_x, batch_y, batch_x_2, batch_y_2)
|
|
134
134
|
|
|
135
|
-
|
|
135
|
+
batch_x = batch_x.astype("float64") / 255.0
|
|
136
136
|
|
|
137
137
|
yield (batch_x, batch_y)
|
|
138
138
|
|
1
ソースコードの修正
test
CHANGED
|
File without changes
|
test
CHANGED
|
@@ -126,7 +126,9 @@
|
|
|
126
126
|
self.random_erasing_area_ratio,
|
|
127
127
|
self.random_erasing_aspect_ratio,
|
|
128
128
|
self.random_erasing_mask_value)
|
|
129
|
+
|
|
129
|
-
|
|
130
|
+
batch_x = batch_x.astype("float64") / 255.0
|
|
131
|
+
|
|
130
132
|
if self.mixup_alpha is not None:
|
|
131
133
|
batch_x, batch_y = self.mixup(batch_x, batch_y, batch_x_2, batch_y_2)
|
|
132
134
|
|