Mitsua commited on
Commit
fee5018
1 Parent(s): 16bbacc

Upload 15 files

Browse files
LICENSE.txt ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MIT License
2
+
3
+ Copyright (c) 2024 ELAN MITSUA Project / Abstract Engine
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
LICENSE_FYSignate1009.txt ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MIT License
2
+
3
+ Copyright (c) 2024 FYSignate1009
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
LICENSE_fractal-pretraining.txt ADDED
@@ -0,0 +1,21 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ MIT License
2
+
3
+ Copyright (c) 2023 Connor Anderson
4
+
5
+ Permission is hereby granted, free of charge, to any person obtaining a copy
6
+ of this software and associated documentation files (the "Software"), to deal
7
+ in the Software without restriction, including without limitation the rights
8
+ to use, copy, modify, merge, publish, distribute, sublicense, and/or sell
9
+ copies of the Software, and to permit persons to whom the Software is
10
+ furnished to do so, subject to the following conditions:
11
+
12
+ The above copyright notice and this permission notice shall be included in all
13
+ copies or substantial portions of the Software.
14
+
15
+ THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR
16
+ IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY,
17
+ FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE
18
+ AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER
19
+ LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM,
20
+ OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE
21
+ SOFTWARE.
params/multi_fractal_ifs_params.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "num_classes":1000,
3
+ "num_image_per_class":1000,
4
+ "num_systems_per_calss":3,
5
+ "num_systems": 6000,
6
+ "niter": 100000,
7
+ "color": true,
8
+ "background": true,
9
+ "patch": true,
10
+ "n_objects": [4, 6],
11
+ "size_range":[0.4, 0.6],
12
+ "jitter_params":true,
13
+ "image_size":256
14
+ }
params/multi_fractal_ifs_params_no_mixup.json ADDED
@@ -0,0 +1,14 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "num_classes":1000,
3
+ "num_image_per_class":1000,
4
+ "num_systems_per_calss":3,
5
+ "num_systems": 3000,
6
+ "niter": 100000,
7
+ "color": true,
8
+ "background": true,
9
+ "patch": true,
10
+ "n_objects": [3, 6],
11
+ "size_range":[0.5, 0.8],
12
+ "jitter_params":true,
13
+ "image_size":256
14
+ }
params/settings.json ADDED
@@ -0,0 +1,19 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ {
2
+ "numof_thread": 10,
3
+ "numof_classes": 1000,
4
+ "numof_instances": 1000,
5
+ "vertex_num_min": 200,
6
+ "vertex_num_max": 1000,
7
+ "perlin_min": 0,
8
+ "line_num_min": 1,
9
+ "line_num_max": 200,
10
+ "line_width": 0.1,
11
+ "radius_min": 10,
12
+ "oval_rate": 2,
13
+ "image_size": 256,
14
+ "start_pos": 256,
15
+ "nami_1_min": 0,
16
+ "nami_2_min": 0,
17
+ "nami_1_max": 20,
18
+ "nami_2_max": 20
19
+ }
requirements.txt ADDED
@@ -0,0 +1 @@
 
 
1
+ numba
src/generator.py ADDED
@@ -0,0 +1,240 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import shutil
3
+ import gc
4
+ import json
5
+ import pickle
6
+ import cv2
7
+ import numpy as np
8
+ from tqdm import tqdm
9
+ import threading
10
+ from concurrent.futures import ProcessPoolExecutor
11
+
12
+ from multi_fractal_db import ifs
13
+ from multi_fractal_db import serach_ifs_systems
14
+ from multi_fractal_db.multi_fractal_dataset import MultiFractalDataset
15
+ from multi_fractal_db.multi_fractal_generator import MultiGenerator
16
+
17
+ from PIL import Image
18
+
19
+ class Generator():
20
+ @classmethod
21
+ def get_params(cls, params_path):
22
+ # デバッグ表示
23
+ cls.debug = True
24
+
25
+ # Conner's FractalDBのパラメータ
26
+ with open(os.path.join(params_path, 'multi_fractal_ifs_params.json')) as f:
27
+ cls.ifs_params = json.load(f)
28
+
29
+ # IFSシステムの探索パラメータ
30
+ kwargs = dict(
31
+ # IFSシステム数
32
+ num_systems=cls.ifs_params["num_systems"],
33
+ # 連立方程式の数
34
+ n=(2, 4),
35
+ bval=1,
36
+ beta=None,
37
+ sample_fn=None,
38
+ )
39
+ # 全IFSシステムのパラメータ作成
40
+ sys = serach_ifs_systems.random_systems(**kwargs)
41
+ cls.ifs_systems = {'params': sys, 'hparams': kwargs}
42
+ print(f"ifs_systems length {len(cls.ifs_systems['params'])}")
43
+
44
+ # デバッグモード
45
+ cls.debug = True
46
+
47
+ return True
48
+
49
+ @classmethod
50
+ def generate(cls, out_path, start_index : int = None, end_index : int = None, jpeg_quality : int = 95):
51
+ if cls.debug:
52
+ print(out_path)
53
+
54
+ # MixUp元フォルダ作成
55
+ base_path = out_path.replace("pretrain", "base")
56
+
57
+ # クラス数
58
+ num_classes = cls.ifs_params['num_classes']
59
+ # 1クラスあたりの画像枚数
60
+ num_image_per_class = cls.ifs_params['num_image_per_class']
61
+
62
+ if start_index is None:
63
+ start_index = 0
64
+ if end_index is None:
65
+ end_index = num_classes
66
+
67
+ # 全クラス分の画像作成
68
+ for iclass in range(start_index, end_index):
69
+ print(f"iclass = {iclass:05}")
70
+ class_dir = os.path.join(out_path, f"{iclass:05}")
71
+ if os.path.exists(class_dir):
72
+ files = os.listdir(class_dir)
73
+ files = [f for f in files if f.endswith(".jpg")]
74
+ if len(files) == num_image_per_class:
75
+ print(f"this iclass already processed = {iclass:05}")
76
+ once_load_failed = False
77
+ for f in files:
78
+ path = os.path.join(class_dir, f)
79
+ try:
80
+ img = Image.open(path)
81
+ except:
82
+ once_load_failed = True
83
+ break
84
+ if not once_load_failed:
85
+ continue
86
+ else:
87
+ print(f"[RE] this iclass already processed = {iclass:05}, but file corrupted. ")
88
+ for f in files:
89
+ os.remove(os.path.join(class_dir, f))
90
+ else:
91
+ for f in files:
92
+ os.remove(os.path.join(class_dir, f))
93
+
94
+ base_images = []
95
+ # MixUp元ベースクラス作成
96
+ for ib, ibase in enumerate([iclass*2, iclass*2+1]):
97
+ # クラスフォルダ
98
+
99
+ # 1クラスあたりのIFSシステム数
100
+ num_systems_per_calss = cls.ifs_params['num_systems_per_calss']
101
+ # 使用するIFSシステムパラメータ
102
+ st = ibase * num_systems_per_calss
103
+ en = (ibase+1)*num_systems_per_calss
104
+ # print(f"ib={ib}, ibase={ibase}, st={st}, en={en}")
105
+ ifs_syss = {'params':cls.ifs_systems['params'][st:en],
106
+ 'hparams': cls.ifs_systems['hparams']}
107
+
108
+ # chaceサイズ
109
+ #cache_size = num_systems_per_calss * num_image_per_class
110
+ cache_size = min(500, num_image_per_class*num_systems_per_calss)
111
+
112
+ # 別スレッドで実行
113
+ future = make_multi_fractal_images(
114
+ ifs_syss, cls.ifs_params, num_systems_per_calss,
115
+ num_image_per_class, cache_size, cls.debug, out_path, ibase)
116
+ base_images.append(future)
117
+
118
+ # MixUp画像作成
119
+ # クラスフォルダ
120
+ class_dir = os.path.join(out_path, f"{iclass:05}")
121
+ if os.path.exists(class_dir)==False:
122
+ os.makedirs(class_dir, exist_ok=True)
123
+
124
+ # 全画像作成
125
+ for idx in tqdm(range(num_image_per_class)):
126
+ # MixUp元画像の読み込み
127
+ image_base1 = base_images[0][idx]
128
+ image_base2 = base_images[1][idx]
129
+ # MixUp
130
+ alpha = 1.0
131
+ lam = np.clip(np.random.beta(alpha, alpha), 0.4, 0.6)
132
+ image_mixup = lam * image_base1 + (1 - lam) * image_base2
133
+ image_mixup = image_mixup.astype(np.uint8)
134
+ # 画像書き出し
135
+ image_file = os.path.join(class_dir, f"{idx:05}.jpg")
136
+ cv2.imwrite(image_file, image_mixup, [cv2.IMWRITE_JPEG_QUALITY, jpeg_quality])
137
+
138
+ # MixUp元フォルダの削除
139
+ base_images = None
140
+ del base_images
141
+ futures = None
142
+ del futures
143
+ gc.collect()
144
+
145
+
146
+ def make_multi_fractal_images(ifs_systems, ifs_params,
147
+ num_systems_per_calss, num_image_per_class, cache_size,
148
+ debug, out_path, ibase):
149
+ # Conner's Multi-FractalDB
150
+ multi_fractal_dataset = MultiFractalDataset(
151
+ ifs_params=ifs_systems,
152
+ num_systems=num_systems_per_calss,
153
+ num_class=1,
154
+ per_class=num_image_per_class,
155
+ generator=MultiGenerator(
156
+ color=ifs_params["color"],
157
+ background=ifs_params["background"],
158
+ niter=ifs_params["niter"],
159
+ patch=ifs_params["patch"],
160
+ n_objects=ifs_params["n_objects"],
161
+ size_range=ifs_params["size_range"],
162
+ jitter_params=ifs_params["jitter_params"],
163
+ cache_size=cache_size,
164
+ size=ifs_params["image_size"]
165
+ ),
166
+ period=2)
167
+
168
+ if debug:
169
+ # 確認用フォルダ
170
+ check_dir = out_path.replace("pretrain", "check")
171
+ if os.path.exists(check_dir)==False:
172
+ os.makedirs(check_dir, exist_ok=True)
173
+ # 使用するIFSフラクタルを描画
174
+ for i, sys in enumerate(ifs_systems['params']):
175
+ image_gray = multi_fractal_dataset.generator.render(sys['system'])
176
+ image_gray = (image_gray * 255).astype(np.uint8)
177
+ #image_gray = cv2.applyColorMap(image_gray, cv2.COLORMAP_BONE)
178
+ image_file = os.path.join(check_dir, f"{ibase:05}_{i:02}.jpg")
179
+ cv2.imwrite(image_file, image_gray)
180
+
181
+ # 全画像数
182
+ base_images = []
183
+ num_fractal_images = len(multi_fractal_dataset)
184
+ class_dir = os.path.join(check_dir, f"{ibase:05}")
185
+ os.makedirs(class_dir, exist_ok=True)
186
+ for idx in range(num_fractal_images):
187
+ # 画像とラベルの取得
188
+ image, labels = multi_fractal_dataset[idx]
189
+ # 画像書き出し
190
+ image_file = os.path.join(class_dir, f"{idx:05}.png")
191
+ cv2.imwrite(image_file, image)
192
+ base_images.append(image)
193
+
194
+ # メモリ解放
195
+ multi_fractal_dataset = None
196
+ del multi_fractal_dataset
197
+ gc.collect()
198
+
199
+ return base_images
200
+
201
+ def multifractal_main(outputdir, start_index, end_index, jpeg_quality):
202
+ Generator.get_params('../params')
203
+ Generator.generate(outputdir, start_index, end_index, jpeg_quality)
204
+
205
+ if __name__ == "__main__":
206
+ import argparse
207
+ from tqdm import tqdm
208
+ from copy import deepcopy
209
+ import concurrent.futures
210
+ import time
211
+ from typing import List
212
+ import multiprocessing
213
+ worker_num=multiprocessing.cpu_count()
214
+ print("workers : ", worker_num)
215
+ parser = argparse.ArgumentParser()
216
+ parser.add_argument('--fpath', type=str, default="../output/pretrain")
217
+ parser.add_argument('--total', type=int, default=1000)
218
+ parser.add_argument('--step', type=int, default=1000//worker_num+1)
219
+ parser.add_argument('--offset', type=int, default=0)
220
+ parser.add_argument('--jpeg_quality', type=int, default=95)
221
+
222
+ args = parser.parse_args()
223
+
224
+ os.makedirs(args.fpath, exist_ok=True)
225
+
226
+ executor = concurrent.futures.ProcessPoolExecutor(max_workers=worker_num)
227
+ futures : List[concurrent.futures.Future] = []
228
+
229
+ for i in range(args.offset, args.total, args.step):
230
+ start_index = i
231
+ end_index = i + args.step
232
+ futures.append(executor.submit(multifractal_main, args.fpath, start_index, end_index, args.jpeg_quality))
233
+
234
+ for future in tqdm(concurrent.futures.as_completed(futures)):
235
+ try:
236
+ rr = future.result()
237
+ except Exception as exc:
238
+ print('generated an exception: %s' % (exc))
239
+
240
+ print("All done!")
src/generator_no_mixup.py ADDED
@@ -0,0 +1,195 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ import shutil
3
+ import gc
4
+ import json
5
+ import pickle
6
+ import cv2
7
+ import numpy as np
8
+ from tqdm import tqdm
9
+ import threading
10
+ from concurrent.futures import ProcessPoolExecutor
11
+
12
+ from multi_fractal_db import ifs
13
+ from multi_fractal_db import serach_ifs_systems
14
+ from multi_fractal_db.multi_fractal_dataset import MultiFractalDataset
15
+ from multi_fractal_db.multi_fractal_generator import MultiGenerator
16
+
17
+ from PIL import Image
18
+
19
+ class Generator():
20
+ @classmethod
21
+ def get_params(cls, params_path):
22
+ # デバッグ表示
23
+ cls.debug = True
24
+
25
+ # Conner's FractalDBのパラメータ
26
+ with open(os.path.join(params_path, 'multi_fractal_ifs_params_no_mixup.json')) as f:
27
+ cls.ifs_params = json.load(f)
28
+
29
+ # IFSシステムの探索パラメータ
30
+ kwargs = dict(
31
+ # IFSシステム数
32
+ num_systems=cls.ifs_params["num_systems"],
33
+ # 連立方程式の数
34
+ n=(2, 4),
35
+ bval=1,
36
+ beta=None,
37
+ sample_fn=None,
38
+ )
39
+ # 全IFSシステムのパラメータ作成
40
+ sys = serach_ifs_systems.random_systems(**kwargs)
41
+ cls.ifs_systems = {'params': sys, 'hparams': kwargs}
42
+ # print(f"ifs_systems length {len(cls.ifs_systems['params'])}")
43
+
44
+ # デバッグモード
45
+ cls.debug = True
46
+
47
+ return True
48
+
49
+ @classmethod
50
+ def generate(cls, out_path, start_index : int = None, end_index : int = None, jpeg_quality : int = 95):
51
+ # if cls.debug:
52
+ # print(out_path)
53
+
54
+ try:
55
+
56
+ # クラス数
57
+ num_classes = cls.ifs_params['num_classes']
58
+ # 1クラスあたりの画像枚数
59
+ num_image_per_class = cls.ifs_params['num_image_per_class']
60
+
61
+ if start_index is None:
62
+ start_index = 0
63
+ if end_index is None:
64
+ end_index = num_classes
65
+
66
+ # 全クラス分の画像作成
67
+ for iclass in tqdm(range(start_index, end_index), total=end_index-start_index):
68
+ # 1クラスあたりのIFSシステム数
69
+ num_systems_per_calss = cls.ifs_params['num_systems_per_calss']
70
+ # 使用するIFSシステムパラメータ
71
+ st = iclass * num_systems_per_calss
72
+ en = (iclass+1)*num_systems_per_calss
73
+ # print(f"ib={ib}, ibase={ibase}, st={st}, en={en}")
74
+ ifs_syss = {'params':cls.ifs_systems['params'][st:en],
75
+ 'hparams': cls.ifs_systems['hparams']}
76
+
77
+ # chaceサイズ
78
+ # cache_size = num_systems_per_calss * num_image_per_class
79
+ cache_size = min(500, num_image_per_class)
80
+
81
+ make_multi_fractal_images(
82
+ ifs_syss, cls.ifs_params, num_systems_per_calss,
83
+ num_image_per_class, cache_size, cls.debug, out_path, iclass, jpeg_quality)
84
+ gc.collect()
85
+ except Exception as e:
86
+ print(e)
87
+ import traceback
88
+ print(traceback.format_exc())
89
+
90
+
91
+ return True
92
+
93
+
94
+ def make_multi_fractal_images(ifs_systems, ifs_params,
95
+ num_systems_per_calss, num_image_per_class, cache_size,
96
+ debug, out_path, ibase, jpeg_quality):
97
+ # Conner's Multi-FractalDB
98
+ multi_fractal_dataset = MultiFractalDataset(
99
+ ifs_params=ifs_systems,
100
+ num_systems=num_systems_per_calss,
101
+ num_class=1,
102
+ per_class=num_image_per_class,
103
+ generator=MultiGenerator(
104
+ color=ifs_params["color"],
105
+ background=ifs_params["background"],
106
+ niter=ifs_params["niter"],
107
+ patch=ifs_params["patch"],
108
+ n_objects=ifs_params["n_objects"],
109
+ size_range=ifs_params["size_range"],
110
+ jitter_params=ifs_params["jitter_params"],
111
+ cache_size=cache_size,
112
+ size=ifs_params["image_size"],
113
+ seed=ibase,
114
+ n_instance_types=num_systems_per_calss,
115
+ ),
116
+ period=2)
117
+
118
+ print(f"jitter success {multi_fractal_dataset.generator.jitter_success}")
119
+ print(f"jitter failed {multi_fractal_dataset.generator.jitter_failed}")
120
+
121
+ if debug:
122
+ # 確認用フォルダ
123
+ check_dir = out_path.replace("pretrain", "check")
124
+ if os.path.exists(check_dir)==False:
125
+ os.makedirs(check_dir, exist_ok=True)
126
+ # 使用するIFSフラクタルを描画
127
+ for i, sys in enumerate(ifs_systems['params']):
128
+ image_gray = multi_fractal_dataset.generator.render(sys['system'])
129
+ image_gray = (image_gray * 255).astype(np.uint8)
130
+ #image_gray = cv2.applyColorMap(image_gray, cv2.COLORMAP_BONE)
131
+ image_file = os.path.join(check_dir, f"{ibase:05}_{i:02}.jpg")
132
+ cv2.imwrite(image_file, image_gray)
133
+
134
+ # 全画像数
135
+ base_images = []
136
+ num_fractal_images = len(multi_fractal_dataset)
137
+ class_dir = os.path.join(out_path, f"{ibase:05}")
138
+ os.makedirs(class_dir, exist_ok=True)
139
+ for idx in range(num_fractal_images):
140
+ # 画像とラベルの取得
141
+ image, labels = multi_fractal_dataset[idx]
142
+ # 画像書き出し
143
+ image_file = os.path.join(class_dir, f"fractal_{ibase:05}_instance_{idx:04}.jpg")
144
+ cv2.imwrite(image_file, image, [cv2.IMWRITE_JPEG_QUALITY, jpeg_quality])
145
+ base_images.append(image)
146
+
147
+ # メモリ解放
148
+ multi_fractal_dataset = None
149
+ del multi_fractal_dataset
150
+ gc.collect()
151
+
152
+ return base_images
153
+
154
+ def multifractal_main(outputdir, start_index, end_index, jpeg_quality):
155
+ Generator.get_params('../params')
156
+ Generator.generate(outputdir, start_index, end_index, jpeg_quality)
157
+
158
+ if __name__ == "__main__":
159
+ import argparse
160
+ from tqdm import tqdm
161
+ from copy import deepcopy
162
+ import concurrent.futures
163
+ import time
164
+ from typing import List
165
+ import multiprocessing
166
+ worker_num=multiprocessing.cpu_count()
167
+ print("workers : ", worker_num)
168
+ parser = argparse.ArgumentParser()
169
+ total_default = 1000
170
+ parser.add_argument('--fpath', type=str, default="../output/pretrain")
171
+ parser.add_argument('--total', type=int, default=total_default)
172
+ parser.add_argument('--step', type=int, default=total_default//worker_num+1)
173
+ parser.add_argument('--offset', type=int, default=0)
174
+ parser.add_argument('--jpeg_quality', type=int, default=85)
175
+
176
+ args = parser.parse_args()
177
+
178
+ os.makedirs(args.fpath, exist_ok=True)
179
+
180
+ executor = concurrent.futures.ProcessPoolExecutor(max_workers=worker_num)
181
+ futures : List[concurrent.futures.Future] = []
182
+
183
+ for i in range(args.offset, args.total, args.step):
184
+ start_index = i
185
+ end_index = min(args.total, i + args.step)
186
+ futures.append(executor.submit(multifractal_main, args.fpath, start_index, end_index, args.jpeg_quality))
187
+
188
+ for future in tqdm(concurrent.futures.as_completed(futures)):
189
+ try:
190
+ rr = future.result()
191
+ except Exception as exc:
192
+ print('generated an exception: %s' % (exc))
193
+
194
+ executor.shutdown()
195
+ print("All done!")
src/multi_fractal_db/diamondsquare.py ADDED
@@ -0,0 +1,89 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from cv2 import cvtColor, COLOR_HSV2RGB
2
+ import numba
3
+ import numpy as np
4
+
5
+
6
+ @numba.njit(cache=True)
7
+ def diamond_square(n, decay=0.5, fixed_corners=True):
8
+ s = 2**n + 1
9
+ a = np.zeros((s, s))
10
+ if fixed_corners:
11
+ a[0, 0] = a[0, s-1] = a[s-1, 0] = a[s-1, s-1] = 0.5
12
+ else:
13
+ a[0, 0] = np.random.rand()
14
+ a[0, s-1] = np.random.rand()
15
+ a[s-1, 0] = np.random.rand()
16
+ a[s-1, s-1] = np.random.rand()
17
+
18
+ for k in range(1, n+1):
19
+ m = 0.5 * np.exp(decay * (1-k))
20
+ ss = s // (2**k)
21
+
22
+ # diamond
23
+ ni = 2**k
24
+ for i in range(0, ni, 2):
25
+ # s / 2**k
26
+ ru = i * ss
27
+ r = ru + ss
28
+ rd = r + ss
29
+ for j in range(0, ni, 2):
30
+ cl = j * ss
31
+ c = cl + ss
32
+ cr = c + ss
33
+ a[r, c] = 0.25 * (a[ru, cl] + a[ru, cr] + a[rd, cl] + a[rd, cr])
34
+ a[r, c] += np.random.uniform(-m, m)
35
+
36
+ # square
37
+ ni = 2**k + 1
38
+ for i in range(ni):
39
+ r = i * ss
40
+ if r > 0: ru = r - ss
41
+ else: ru = s - ss - 1
42
+ if r < s-1: rd = r + ss
43
+ else: rd = ss
44
+ sj = 1 if i % 2 == 0 else 0
45
+ for j in range(sj, ni, 2):
46
+ c = j * ss
47
+ if c > 0: cl = c - ss
48
+ else: cl = s - ss - 1
49
+ if c < s-1: cr = c + ss
50
+ else: cr = ss
51
+ a[r, c] = 0.25 * (a[ru, c] + a[r, cl] + a[r, cr] + a[rd, c])
52
+ a[r, c] += np.random.uniform(-m, m)
53
+ return a
54
+
55
+
56
+ @numba.njit(cache=True)
57
+ def _colorize(ds):
58
+ img = np.empty((ds.shape[0], ds.shape[1], 3), dtype=np.uint8)
59
+
60
+ hue_scale = np.random.uniform(0.15, 1) * 179
61
+ hue_shift = np.random.rand() * 179
62
+
63
+ sat_scale = np.random.uniform(0.1, 0.3) * 255
64
+ sat_shift = np.random.uniform(0.0, 0.3) * 255
65
+
66
+ val_scale = np.random.uniform(0.1, 0.3) * 255
67
+ val_shift = np.random.uniform(0.1, 0.3) * 255
68
+
69
+ for i in range(ds.shape[0]):
70
+ for j in range(ds.shape[1]):
71
+ x = ds[i, j]
72
+ img[i, j, 0] = np.uint8(min((x * hue_scale + hue_shift) % 179, 179))
73
+ img[i, j, 1] = np.uint8(min(x * sat_scale + sat_shift, 255))
74
+ img[i, j, 2] = np.uint8(min(x * val_scale + val_shift, 255))
75
+
76
+ return img
77
+
78
+
79
+ def colorized_ds(size=256):
80
+ # 画像サイズ256=2の8乗
81
+ n = int(np.ceil(np.log2(size)))
82
+ # 乱数生成器
83
+ rng = np.random.default_rng()
84
+ # 背景作成
85
+ r = diamond_square(n, rng.uniform(0.4, 0.8), fixed_corners=False)[:size, :size]
86
+ # HSVで色付け、HSV->RGB変換
87
+ img = _colorize(r)
88
+ img = cvtColor(img, COLOR_HSV2RGB, dst=img)
89
+ return img
src/multi_fractal_db/ifs.py ADDED
@@ -0,0 +1,373 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from functools import partial
2
+
3
+ from cv2 import cvtColor, COLOR_HSV2RGB
4
+ import numba
5
+ import numpy as np
6
+
7
+
8
+ # IFSフラクタル座標計算
9
+ @numba.njit(cache=True)
10
+ def iterate(sys, n_iter, ps=None):
11
+ '''Compute points in the fractal defined by the system `sys` by random iteration. `n_iter` iterations
12
+ are performed, and a transform is sampled at each iteration according to the probabilites defined by
13
+ `ps`.
14
+
15
+ Args:
16
+ sys (np.ndarray): array of shape (n, 2, 3), containing the affine transform parameters.
17
+ n_iter (int): number of iterations/points to calculate.
18
+ ps (Optional[array-like]): length-n array of probabilites. If None (default), the probabilites are
19
+ calculated to be proportional to the determinants of the affine transformation matrices.
20
+
21
+ Returns:
22
+ ndarray of shape (n_iter, 2) containing the (x, y) coordinates of the generated points.
23
+ '''
24
+ det = sys[:, 0, 0] * sys[:, 1, 1] - sys[:, 0, 1] * sys[:, 1, 0]
25
+ # 確率が指定されていない場合、detから計算
26
+ if ps is None:
27
+ # パラメータセット毎の確率
28
+ ps = np.abs(det)
29
+ # 全体で1になるように正規化
30
+ ps = ps / ps.sum()
31
+ # 累積確率、0-1の乱数でどのパラメータセットを使うか決めるため
32
+ ps = np.cumsum(ps)
33
+
34
+ # 全フラクタル点の座標
35
+ coords = np.empty((n_iter, 2))
36
+
37
+ # starting point is $v = (I-A_1)^(-1) b$ since this point is gaurenteed to be in the set
38
+ # (assuming that A_1 is contractive) (A_1 = sys[0])
39
+ s = 1 / (1 + det[0] - sys[0, 0, 0] - sys[0, 1, 1])
40
+ x = s * ((1 - sys[0, 1, 1]) * sys[0, 0, 2] + sys[0, 0, 1] * sys[0, 1, 2])
41
+ y = s * ((1 - sys[0, 0, 0]) * sys[0, 1, 2] + sys[0, 1, 0] * sys[0, 0, 2])
42
+
43
+ for i in range(n_iter):
44
+ # 0-1の一様乱数の生成
45
+ r = np.random.rand()
46
+ # 使用するパラメータセットの特定
47
+ for k in range(len(ps)):
48
+ if r < ps[k]: break
49
+ # パラメータセット
50
+ a, b, e, c, d, f = sys[k].ravel()
51
+ # 次座標の計算
52
+ xt = x
53
+ x = a * xt + b * y + e
54
+ y = c * xt + d * y + f
55
+ coords[i] = x, y
56
+
57
+ # 発散した場合はブレイク
58
+ if not np.isfinite(x) or not np.isfinite(y): break # if contractivity is satisfied, can remove this check
59
+ return coords
60
+
61
+
62
+ @numba.njit(cache=True)
63
+ def minmax(coords):
64
+ '''Returns both the minimum and maximum values along the 0 axis of an array with shape (n, 2). This only
65
+ requires a single pass through the array, and is faster than calling np.min and np.max seperately.
66
+
67
+ Args:
68
+ coords (np.ndarray): an array of shape (n, 2)
69
+
70
+ Returns:
71
+ Two ndarrays of shape (2,), the first containing the minimum values and the second containing the maximums
72
+ '''
73
+ # x,y座標の最大、最小値
74
+ mins = np.full(2, np.inf)
75
+ maxs = np.full(2, -np.inf)
76
+ for i in range(len(coords)):
77
+ x, y = coords[i]
78
+ if x < mins[0]: mins[0] = x
79
+ if y < mins[1]: mins[1] = y
80
+ if x > maxs[0]: maxs[0] = x
81
+ if y > maxs[1]: maxs[1] = y
82
+ return mins, maxs
83
+
84
+
85
+ @numba.njit(cache=True)
86
+ def _extent(region):
87
+ x1, y1, x2, y2 = region
88
+ xspan = x2 - x1
89
+ xspan = xspan if xspan > 0 else 1
90
+ yspan = y2 - y1
91
+ yspan = yspan if yspan > 0 else 1
92
+ return xspan, yspan
93
+
94
+
95
+
96
+ @numba.njit(cache=True)
97
+ def _render_binary(coords, s, region):
98
+ '''Renders a square, binary image from coordinate points and a given region.
99
+
100
+ Args:
101
+ coords (np.ndarray): coordinate array of shape (n, 2).
102
+ s (int): side length of the rendered image. The image will have width = height = s.
103
+ region (np.ndarray): array of shape (4,), containing [minx, miny, maxx, maxy]. These four values
104
+ define the region in coordinate space that will be rendered to the image. Coordinate points
105
+ that fall outside the bounds of the region will be ignored.
106
+
107
+ Returns:
108
+ A binary image as an ndarray of shape (s, s).
109
+ '''
110
+ imgb = np.zeros((s, s), dtype=np.uint8)
111
+ xspan, yspan = _extent(region)
112
+ xscale = (s-1) / xspan
113
+ yscale = (s-1) / yspan
114
+ xmin, ymin = region[0], region[1]
115
+ for i in range(len(coords)):
116
+ r = int((coords[i,0] - xmin) * xscale)
117
+ c = int((coords[i,1] - ymin) * yscale)
118
+ if r >= 0 and r < s and c >= 0 and c < s:
119
+ imgb[r, c] = 1
120
+ return imgb
121
+
122
+
123
+
124
+ @numba.njit(cache=True)
125
+ def _render_binary_patch(coords, s, region, patch):
126
+ '''Renders a square, binary image from coordinate points and a given region. Instead of rendering a
127
+ single point for each coordinate, a 3x3 patch is rendered, centered on the coordinate.
128
+
129
+ Args:
130
+ coords (np.ndarray): coordinate array of shape (n, 2).
131
+ s (int): side length of the rendered image. The image will have width = height = s.
132
+ region (np.ndarray): array of shape (4,), containing [minx, miny, maxx, maxy]. These four values
133
+ define the region in coordinate space that will be rendered to the image. Coordinate points
134
+ that fall outside the bounds of the region will be ignored.
135
+ patch (np.ndarray): array of shape (3, 3), where each value is either 0 or 1 (binary).
136
+
137
+ Returns:
138
+ A grayscale image as an ndarray of shape (s, s).
139
+ '''
140
+ # ブランク画像作成
141
+ imgb = np.zeros((s, s), dtype=np.uint8)
142
+ # x幅、y幅
143
+ xspan, yspan = _extent(region)
144
+ # xy軸方向の拡大係数
145
+ xscale = (s-1) / xspan
146
+ yscale = (s-1) / yspan
147
+ # xy軸最小値
148
+ xmin, ymin = region[0], region[1]
149
+ for i in range(len(coords)):
150
+ # ピクセル座標の計算
151
+ rr = int((coords[i,0] - xmin) * xscale)
152
+ cc = int((coords[i,1] - ymin) * yscale)
153
+ # patch処理
154
+ for j in range(len(patch)):
155
+ # マスクが1ならそのまま、0or2なら左右に書き込みでそのピクセルはブランク
156
+ r = rr + patch[j, 0] - 1
157
+ c = cc + patch[j, 1] - 1
158
+ if r >= 0 and r < s and c >= 0 and c < s:
159
+ imgb[r, c] = 1
160
+ return imgb
161
+
162
+
163
+ @numba.njit(cache=True)
164
+ def _render_graded(coords, s, region):
165
+ '''Renders a square, grayscale image from coordinate points and a given region. The grayscale values for
166
+ a given pixel is proportional to the number of coordinate points that land on that pixel.
167
+
168
+ See _render_binary for an explanation of the arguments.
169
+ '''
170
+ # ブランク画像、0-1の小数点になるので、floatで
171
+ imgf = np.zeros((s, s), dtype=np.float64)
172
+ # x,yの幅
173
+ xspan, yspan = _extent(region)
174
+ # ピクセル位置スケーリング
175
+ xscale = (s-1) / xspan
176
+ yscale = (s-1) / yspan
177
+ xmin, ymin = region[0], region[1]
178
+ for i in range(len(coords)):
179
+ # ピクセル座標計算
180
+ r = int((coords[i,0] - xmin) * xscale)
181
+ c = int((coords[i,1] - ymin) * yscale)
182
+ if r >= 0 and r < s and c >= 0 and c < s:
183
+ # 明暗つけるため加算する
184
+ imgf[r, c] += 1
185
+ # 正規化
186
+ mval = imgf.max()
187
+ if mval > 0:
188
+ imgf /= mval
189
+ return imgf
190
+
191
+
192
+ @numba.njit(cache=True)
193
+ def _render_graded_patch(coords, s, region, patch):
194
+ '''Renders a square, grayscale image from coordinate points and a given region. The grayscale values for
195
+ a given pixel is proportional to the number of coordinate points that land on that pixel. Instead of rendering
196
+ a single point for each coordinate, a 3x3 patch is rendered, centered on the coordinate.
197
+
198
+ See _render_binary_patch for an explanation of the arguments.
199
+ '''
200
+ imgf = np.zeros((s, s), dtype=np.float64)
201
+ xspan, yspan = _extent(region)
202
+ xscale = (s-1) / xspan
203
+ yscale = (s-1) / yspan
204
+ xmin, ymin = region[0], region[1]
205
+ for i in range(len(coords)):
206
+ rr = int((coords[i,0] - xmin) * xscale)
207
+ cc = int((coords[i,1] - ymin) * yscale)
208
+ for j in range(len(patch)):
209
+ r = rr + patch[j, 0] - 1
210
+ c = cc + patch[j, 1] - 1
211
+ if r >= 0 and r < s and c >= 0 and c < s:
212
+ imgf[r, c] += 1
213
+ mval = imgf.max()
214
+ if mval > 0:
215
+ imgf /= mval
216
+ return imgf
217
+
218
+
219
+ # フラクタル描画
220
+ def render(coords, s=256, binary=True, region=None, patch=False):
221
+ '''Render an image from a set of coordinates and an optionally specified region.
222
+
223
+ Args:
224
+ coords (np.ndarray): coordinate array of shape (n, 2).
225
+ s (int): side length of the rendered image. The image will have width = height = s.
226
+ binary (bool): if True, render a binary image; otherwise, render a grayscale image, where the grayscale
227
+ value is proportional to the number of coordinates that map to the pixel.
228
+ region (Optional[np.ndarray]): array of shape (4,), containing [minx, miny, maxx, maxy]. These four
229
+ values define the region in coordinate space that will be rendered to the image. Coordinate points
230
+ that fall outside the bounds of the region will be ignored. If None (default), the minimum and
231
+ maximum coordinate values are used.
232
+ patch (bool): if False, render each coordinate as a single point. If True, renders a 3x3 patch centered
233
+ at each coordinate. The patch is randomly sampled (each value is chosen uniformly from [0, 1]).
234
+
235
+ Returns:
236
+ An image (either grayscale or binary, depending) as an ndarray of shape (s, s).
237
+ '''
238
+
239
+ # 範囲指定がない場合
240
+ if region is None:
241
+ # フラクタル点座標の最大、最小範囲で切り出し
242
+ region = np.concatenate(minmax(coords))
243
+ else:
244
+ region = np.asarray(region)
245
+
246
+ # パッチ処理有りの場��
247
+ if patch:
248
+ # 3x3のマスク、だがどうも2もあるような
249
+ p = np.stack(np.divmod(np.arange(9)[np.random.randint(0, 2, (9,), dtype=bool)], 3), 1)
250
+
251
+ if binary:
252
+ if patch:
253
+ # patch処理ありの白黒画像
254
+ return _render_binary_patch(coords, s, region, p)
255
+ else:
256
+ # patch処理なし白黒画像
257
+ return _render_binary(coords, s, region)
258
+ else:
259
+ # カラー指定の場合、とりあえず輝度値のみ計算
260
+ if patch:
261
+ # patc処理ありカラー画像
262
+ return _render_graded_patch(coords, s, region, p)
263
+ else:
264
+ # patc処理なしカラー画像
265
+ return _render_graded(coords, s, region)
266
+
267
+
268
+ # HSV空間で色付け関数
269
+ @numba.njit(cache=True)
270
+ def _hsv_colorize(rendered, min_sat=0.3, min_val=0.5):
271
+ '''Creates a 3-channel HSV image from a 1-channel gray image.
272
+ '''
273
+ h, w = rendered.shape[:2]
274
+ img = np.empty((h, w, 3), dtype=np.uint8)
275
+
276
+ # 基準Hue
277
+ hue_shift = np.random.rand() * 179
278
+ # 彩度
279
+ sat = np.uint8(np.random.uniform(min_sat, 1) * 255)
280
+ # 明度
281
+ val = np.uint8(np.random.uniform(min_val, 1) * 255)
282
+
283
+ for i in range(h):
284
+ for j in range(w):
285
+ x = rendered[i, j]
286
+ if x > 0:
287
+ img[i, j, 0] = np.uint8(min((x * 179 + hue_shift) % 179, 179)) # implicit MOD(256)
288
+ img[i, j, 1] = sat
289
+ img[i, j, 2] = val
290
+ else:
291
+ img[i, j, 0] = 0
292
+ img[i, j, 1] = 0
293
+ img[i, j, 2] = 0
294
+ return img
295
+
296
+
297
+ @numba.njit(cache=True)
298
+ def _hsv_colorize2(rendered, min_sat=0.3, min_val=0.5, hue_base=0, hue_range=179, max_sat=1.0):
299
+ h, w = rendered.shape[:2]
300
+ img = np.empty((h, w, 3), dtype=np.uint8)
301
+
302
+ # 基準Hue
303
+ hue_shift = np.random.rand() * hue_range + hue_base
304
+ hue_scale = np.random.uniform(-0.5, 0.5) * hue_range
305
+ # 彩度
306
+ sat = np.uint8(np.random.uniform(min_sat, max_sat) * 255)
307
+ # 明度
308
+ val = np.uint8(np.random.uniform(min_val, 1) * 255)
309
+
310
+ for i in range(h):
311
+ for j in range(w):
312
+ x = rendered[i, j]
313
+ if x > 0:
314
+ img[i, j, 0] = np.uint8(min((x * hue_scale + hue_shift) % 179, 179)) # implicit MOD(256)
315
+ img[i, j, 1] = sat
316
+ img[i, j, 2] = min(max(x * 1024, val), 255)
317
+ else:
318
+ img[i, j, 0] = 0
319
+ img[i, j, 1] = 0
320
+ img[i, j, 2] = 0
321
+ return img
322
+
323
+
324
+ # 前景画像埋め込み
325
+ @numba.njit(cache=True)
326
+ def composite(fg, bg):
327
+ '''Copy nonzero pixels from fg into bg. Modifies bg in-place.'''
328
+ for i in range(fg.shape[0]):
329
+ for j in range(fg.shape[1]):
330
+ if fg[i, j, 0] != 0 or fg[i, j, 1] != 0 or fg[i, j, 2] != 0:
331
+ bg[i, j] = fg[i, j]
332
+ return bg
333
+
334
+ # 前景画像埋め込み
335
+ @numba.njit(cache=True)
336
+ def composite_v2(fg, bg, fg_mask):
337
+ '''Copy nonzero pixels from fg into bg. Modifies bg in-place.'''
338
+ for i in range(fg.shape[0]):
339
+ for j in range(fg.shape[1]):
340
+ if fg[i, j, 0] != 0 or fg[i, j, 1] != 0 or fg[i, j, 2] != 0:
341
+ factor = np.power(fg_mask[i, j], 0.3)
342
+ for k in range(3):
343
+ bg[i, j, k] = np.uint8(min(fg[i, j, k] * factor +
344
+ bg[i, j, k] * (1.0 - factor), 255))
345
+ # bg[i, j, k] = np.uint8(fg_mask[i, j] * 255)
346
+ # bg[i, j] = (fg[i, j] * fg_mask[i, j] + bg[i, j] * (255 - fg_mask[i, j])) / 255
347
+ return bg
348
+
349
+
350
+ # カラー画像化
351
+ def colorize(rendered, min_sat=0.3, min_val=0.5, hue_base=0, hue_range=179, max_sat=1.0):
352
+ '''Turns a grayscale image into a color image, where the colors are randomly chosen as explained below.
353
+
354
+ First, the grayscale values are converted to the range [0, 255]. A reference hue value h is chosen
355
+ uniformly from [0, 255], and the hue for each pixel p becomes (p + h) mod 256. Then global saturation
356
+ and value scales are chosen uniformly from the ranges [min_sat, 1] and [min_val, 1]. Finally, the image
357
+ is converted to RGB.
358
+
359
+ Args:
360
+ rendered (np.ndarray): grayscale image of shape (w, h), with values in the range [0, 1].
361
+ min_sat (float): minimum "saturation" value, defining the range of possible saturation values to draw from.
362
+ min_val (float): minimum "value" value, defining the range of possible "value" (as in light/dark) values to
363
+ draw from.
364
+
365
+ Returns:
366
+ A color image as an ndarray of shape (w, h, 3).
367
+ '''
368
+ # HSV空間で色付け
369
+ # img = _hsv_colorize(rendered, min_sat, min_val)
370
+ img = _hsv_colorize2(rendered, min_sat, min_val, hue_base, hue_range, max_sat)
371
+ # HSV->RGB変換
372
+ cvtColor(img, COLOR_HSV2RGB, dst=img)
373
+ return img
src/multi_fractal_db/multi_fractal_dataset.py ADDED
@@ -0,0 +1,72 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from functools import partial
2
+ import pickle
3
+ from typing import Callable, Optional, Tuple, Union
4
+ import warnings
5
+
6
+ from cv2 import GaussianBlur
7
+ import numpy as np
8
+ import torch
9
+ import torchvision
10
+
11
+ from multi_fractal_db import diamondsquare, ifs
12
+ from multi_fractal_db.multi_fractal_generator import MultiGenerator
13
+
14
+
15
+ class MultiFractalDataset(object):
16
+ def __init__(
17
+ self,
18
+ ifs_params: object,
19
+ num_systems: int = 100,
20
+ num_class: int = 100,
21
+ per_class: int = 100,
22
+ generator: Optional[MultiGenerator] = None,
23
+ period: int = 2,
24
+ ):
25
+ # IFSシステム数
26
+ self.num_systems = num_systems
27
+ # クラス数
28
+ self.num_class = num_class
29
+ # クラスあたりの画像枚数
30
+ self.per_class = per_class
31
+ # クラスあたりのIFSシステム数
32
+ self.per_system = num_class * per_class / num_systems
33
+ # IFSシステムパラメータ
34
+ self.params = ifs_params['params'][:num_systems]
35
+
36
+ # なぜかここMultiGeneratorが生成されてしまうのでコメントアウト
37
+ self.generator : MultiGenerator = generator #or MultiGenerator()
38
+
39
+ # キャッシュ画像をすべて作成
40
+ while len(self.generator.cache['fg']) < self.generator.cache_size:
41
+ for isys in range(num_systems):
42
+ # 代表クラス番号
43
+ # k = np.random.default_rng().integers(0, num_class)
44
+ self.generator.add_sample(self.params[isys]['system'], label=isys) # change to instance label
45
+
46
+ self.steps = 0
47
+ self.period = period
48
+
49
+ def __len__(self):
50
+ return self.num_class * self.per_class
51
+
52
+ def get_label(self, idx):
53
+ return int(idx // self.num_class)
54
+
55
+ def get_system(self, idx):
56
+ return int(idx // self.per_system)
57
+
58
+ def __getitem__(self, idx):
59
+ # whether it's time to render a new fractal or not
60
+ self.steps = (self.steps + 1) % self.period
61
+ sample = self.steps == 0
62
+ # IFSシステム番号
63
+ sysidx = self.get_system(idx)
64
+ # ラベル取得
65
+ label = self.get_label(idx)
66
+ # IFSパラメータ取得
67
+ params = self.params[sysidx]['system']
68
+ # 混合画像と混合ラベル
69
+ img, labels = self.generator(params, label=label, new_sample=sample)
70
+ #img = torch.from_numpy(img).float().mul_(1/255.).permute(2,0,1)
71
+
72
+ return img, labels
src/multi_fractal_db/multi_fractal_generator.py ADDED
@@ -0,0 +1,492 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ from functools import partial
2
+ from typing import Callable, Optional, Tuple, Union
3
+
4
+ from cv2 import GaussianBlur, resize, INTER_LINEAR, INTER_CUBIC, getRotationMatrix2D, warpAffine
5
+ import numpy as np
6
+
7
+ from multi_fractal_db import diamondsquare, ifs
8
+
9
+
10
+ class _GeneratorBase(object):
11
+ def __init__(
12
+ self,
13
+ # 画像サイズ
14
+ size: int = 224,
15
+ # jit方法、Trueもしくは文字列を受け入れる
16
+ jitter_params: Union[bool, str] = True,
17
+ flips: bool = True,
18
+ # ボケの標準偏差
19
+ sigma: Optional[Tuple[float, float]] = (0.5, 1.0),
20
+ blur_p: Optional[float] = 0.5,
21
+ # フラクタル座標点数
22
+ niter = 100000,
23
+ # patch処理有無
24
+ patch = True,
25
+ ):
26
+ self.size = size
27
+ self.jitter_params = jitter_params
28
+ self.flips = flips
29
+ self.sigma = sigma
30
+ self.blur_p = blur_p
31
+ self.niter = niter
32
+ self.patch = patch
33
+
34
+ self.jitter_success = 0
35
+ self.jitter_failed = 0
36
+
37
+ self.hue_base = np.random.default_rng().integers(0, 179)
38
+ self.min_sat = np.random.default_rng().uniform(0.0, 0.75)
39
+ self.hue_range = max(0, 179 * (0.7 - self.min_sat))
40
+
41
+ self.rng = np.random.default_rng()
42
+ self.cache = {'fg': [], 'bg': []}
43
+ # jit関数定義
44
+ self._set_jitter()
45
+
46
+ # jit関数
47
+ def _set_jitter(self):
48
+ # jit指定が文字列の場合
49
+ if isinstance(self.jitter_params, str):
50
+ if self.jitter_params.startswith('fractaldb'):
51
+ k = int(self.jitter_params.split('-')[1]) / 10
52
+ choices = np.linspace(1-2*k, 1+2*k, 5, endpoint=True)
53
+ self.jitter_fnc = partial(self._fractaldb_jitter, choices=choices)
54
+ if self.jitter_params.startswith('svd'):
55
+ self.jitter_fnc = self._svd_jitter
56
+ if self.jitter_params.startswith('sval'):
57
+ self.jitter_fnc = self._sval_jitter
58
+ elif self.jitter_params:
59
+ # defualt
60
+ self.jitter_fnc = self._basic_jitter
61
+ else:
62
+ # そのまま返す
63
+ self.jitter_fnc = lambda x: x
64
+
65
+ def _fractaldb_jitter(self, sys, prange=(0.5, 2.0), choices=[]):
66
+ n = len(sys)
67
+ y, x = np.divmod(self.rng.integers(0, 6, (n,)), 3)
68
+ scaling_factors = self.rng.uniform(*prange, n)
69
+ sys[range(n), y, x] *= scaling_factors
70
+ return sys
71
+
72
+ # def _fractaldb_jitter(self, sys, choices=(.8,.9,1,1.1,1.2)):
73
+ # n = len(sys)
74
+ # y, x = np.divmod(self.rng.integers(0, 6, (n,)), 3)
75
+ # sys[range(n), y, x] *= self.rng.choice(choices)
76
+ # return sys
77
+
78
+
79
+ # デフォルトのJitter関数
80
+ def _basic_jitter(self, sys, prange=(0.8, 1.1)):
81
+ # tweak system parameters--randomly choose one transform and scale it
82
+ # this actually amounts to scaling the singular values by a random factor
83
+ # IFSパラメータの全要素数
84
+ n = len(sys)
85
+ # どれか1要素を係数かけて微笑変動させる
86
+ sys[self.rng.integers(0, n)] *= self.rng.uniform(*prange)
87
+ return sys
88
+
89
+
90
+ def _svd_jitter(self, sys):
91
+ '''Jitter the parameters of one of the systems functions, in SVD space.'''
92
+ k = self.rng.integers(0, len(sys) * 3)
93
+ sidx, pidx = divmod(k, 3)
94
+ if pidx < 2:
95
+ q = self.rng.uniform(-0.5, 0.5)
96
+ u, s, v = np.linalg.svd(sys[sidx, :, :2])
97
+ cq, sq = np.cos(q), np.sin(q)
98
+ r = np.array([[cq, -sq], [sq, cq]])
99
+ if pidx == 0:
100
+ u = r @ u
101
+ else:
102
+ v = r @ v
103
+ sys[sidx, :, :2] = (u * s[None,:]) @ v
104
+ else:
105
+ x, y = self.rng.uniform(-0.5, 0.5, (2,))
106
+ sys[sidx, :, 2] += [x, y]
107
+ return sys
108
+
109
+
110
+ def _sval_jitter(self, sys):
111
+ k = self.rng.integers(0, sys.shape[0])
112
+ svs = np.linalg.svd(sys[...,:2], compute_uv=False)
113
+ fac = (svs * [1, 2]).sum()
114
+ minf = 0.5 * (5 + sys.shape[0])
115
+ maxf = minf + 0.5
116
+ ss = svs[k, 0] + 2 * svs[k, 1]
117
+ smin = (minf - fac + ss) / ss
118
+ smax = (maxf - fac + ss) / ss
119
+ m = self.rng.uniform(smin, smax)
120
+ u, s, v = np.linalg.svd(sys[k, :, :2])
121
+ s = s * m
122
+ sys[k, :, :2] = (u * s[None]) @ v
123
+ return sys
124
+
125
+
126
+ # IFSパラメータ微小変動
127
+ def jitter(self, sys):
128
+ # 4回、微小変動を起こす
129
+ attempts = 4 if self.jitter_params else 0
130
+ for i in range(attempts):
131
+ # jitter system parameters
132
+ sysc = sys.copy()
133
+ sysc = self.jitter_fnc(sysc)
134
+ # 特異値分解、対角行列の特異値のみ計算
135
+ # occasionally the modified parameters cause the system to explode
136
+ svd = np.linalg.svd(sysc[:,:,:2], compute_uv=False)
137
+ # 1を超えている場合発散するので再度jitする
138
+ if svd.max() > 1: continue
139
+
140
+ self.jitter_success += 1
141
+ break
142
+ else:
143
+ # fall back on not jittering the parameters
144
+ self.jitter_failed += 1
145
+ sysc = sys
146
+ return sysc
147
+
148
+
149
+ # IFSフラクタル座標と範囲の計算
150
+ def _iterate(self, sys):
151
+ rng = self.rng
152
+
153
+ # 全フラクタル点の座標を計算
154
+ coords = ifs.iterate(sys, self.niter)
155
+
156
+ # フラクタル点の座標範囲、[xmin, ymin, xmax, ymax]
157
+ region = np.concatenate(ifs.minmax(coords))
158
+
159
+ return coords, region
160
+
161
+
162
+ def render(self, sys):
163
+ raise NotImplementedError()
164
+
165
+
166
+ # データ拡張(RandomFlip)
167
+ def random_flips(self, img):
168
+ # random flips/rotations
169
+ if self.rng.random() > 0.5:
170
+ # 縦横転置
171
+ img = img.transpose(1, 0)
172
+ if self.rng.random() > 0.5:
173
+ # 上下反転
174
+ img = img[::-1]
175
+ if self.rng.random() > 0.5:
176
+ # 左右反転
177
+ img = img[:, ::-1]
178
+ # 配列のメモリ連続性を確保するため
179
+ img = np.ascontiguousarray(img)
180
+ return img
181
+
182
+
183
+ # フラクタル画像のカラー化
184
+ def to_color(self, img):
185
+ # カラー画像化
186
+ return ifs.colorize(img, min_sat=self.min_sat, hue_base=self.hue_base, hue_range=self.hue_range)
187
+
188
+
189
+ # 白黒画像
190
+ def to_gray(self, img):
191
+ # バイナリ画像なので中間の127の灰色画像で、3チャンネルにしておく
192
+ return (img * 1024).clip(0, 255).astype(np.uint8)[..., None].repeat(3, axis=2)
193
+
194
+
195
+ # 背景画像を作成
196
+ def render_background(self):
197
+ # 幾何模様のカラー背景作成
198
+ bg = diamondsquare.colorized_ds(self.size)
199
+ # 背景を暗くする
200
+ #bg = bg * 0.5
201
+ return bg
202
+
203
+
204
+ # 背景に前景を埋め込み
205
+ def composite(self, foreground, base, idx=None):
206
+ return ifs.composite(foreground, base)
207
+
208
+ # 背景に前景を埋め込み
209
+ def composite_v2(self, foreground, base, mask, idx=None):
210
+ return ifs.composite_v2(foreground, base, mask)
211
+
212
+ # データ拡張(ボカし)
213
+ def random_blur(self, img):
214
+ # 標準偏差取得
215
+ sigma = self.rng.uniform(*self.sigma)
216
+ # 画像平坦化
217
+ img = GaussianBlur(img, (3, 3), sigma, dst=img)
218
+ return img
219
+
220
+
221
+ def generate(self, sys):
222
+ raise NotImplementedError()
223
+
224
+
225
+ def __call__(self, sys, *args, **kwargs):
226
+ return self.generate(sys, *args, **kwargs)
227
+
228
+
229
+
230
+ class MultiGenerator(_GeneratorBase):
231
+ def __init__(
232
+ self,
233
+ # 画像サイズ
234
+ size: int = 224,
235
+ # キャッシュサイズ
236
+ cache_size: int = 100,
237
+ # 混合数
238
+ n_objects: Tuple[int, int] = (1, 5),
239
+ # 埋め込み画像サイズ
240
+ size_range: Tuple[float, float] = (0.2, 0.6),
241
+ # jitパラメータ、デフォルトではなし
242
+ jitter_params: Union[bool, str] = False,
243
+ # データ拡張(縦横、左右、上下反転)フラグ
244
+ flips: bool = True,
245
+ # ボカしの標準偏差
246
+ sigma: Optional[Tuple[float, float]] = (0.5, 1.0),
247
+ # データ拡張(ボケ実施確率)
248
+ blur_p: Optional[float] = 0.5,
249
+ # カラー画像フラグ
250
+ color = True,
251
+ # 背景画像の付与の有無
252
+ background = True,
253
+ # フラクタル点数
254
+ niter = 100000,
255
+ # patch処理実施有無
256
+ patch = True,
257
+ # 混合画像の採用確率?
258
+ nobj_p = None,
259
+
260
+ seed = 0,
261
+ n_instance_types = 1,
262
+ ):
263
+ self.size = size
264
+ self.n_objects = n_objects
265
+ self.size_range = size_range
266
+ self.jitter_params = jitter_params
267
+ self.flips = flips
268
+ self.sigma = sigma
269
+ self.blur_p = blur_p
270
+ self.color = color
271
+ self.background = background
272
+ self.niter = niter
273
+ self.patch = patch
274
+
275
+ self.bg_sel_cache_idx = 0
276
+ self.sel_cache_idx = 0
277
+
278
+ self.jitter_success = 0
279
+ self.jitter_failed = 0
280
+
281
+ np.random.seed(seed)
282
+ self.hue_bases = [np.random.uniform(0, 179) for _ in range(n_instance_types)]
283
+ self.min_sats = [np.random.uniform(0.0, 0.50) for _ in range(n_instance_types)]
284
+ self.max_sats = [min(1, self.min_sats[i] * 2 + 0.2) for i in range(n_instance_types)]
285
+ self.hue_ranges = [max(0, 179 * (0.8 - self.min_sats[i]) / 0.8) for i in range(n_instance_types)]
286
+
287
+ # 乱数生成器
288
+ self.rng = np.random.default_rng(seed=0)
289
+
290
+
291
+
292
+ # キャッシュサイズ
293
+ self.cache_size = cache_size
294
+ # 前景と背景のキャッシュ
295
+ self.cache = {'fg': [], 'bg': [], 'label': []}
296
+ self.idx = 0
297
+
298
+ # 混合確率
299
+ if nobj_p is None:
300
+ # 等確率
301
+ self.nobj_p = np.ones(n_objects[1] - n_objects[0] + 1)
302
+ else:
303
+ self.nobj_p = np.array(nobj_p, dtype=np.float64)
304
+ self.nobj_p /= self.nobj_p.sum()
305
+
306
+ # jit関数設定
307
+ self._set_jitter()
308
+
309
+
310
+ # フラクタル画像のカラー化
311
+ def to_color(self, img, instance_label):
312
+ # カラー画像化
313
+ return ifs.colorize(img,
314
+ min_sat=self.min_sats[instance_label],
315
+ hue_base=self.hue_bases[instance_label],
316
+ hue_range=self.hue_ranges[instance_label],
317
+ max_sat=self.max_sats[instance_label])
318
+
319
+
320
+ def __len__(self):
321
+ # 前景キャッシュ数を返す
322
+ return len(self.cache['fg'])
323
+
324
+
325
+ def _update_cache(self, fg, bg, label):
326
+ if len(self) < self.cache_size:
327
+ # キャッシュサイズ以下であれば追記
328
+ self.cache['fg'].append(fg)
329
+ self.cache['bg'].append(bg)
330
+ self.cache['label'].append(label)
331
+ else:
332
+ # idx番目に追加
333
+ self.cache['fg'][self.idx] = fg
334
+ self.cache['bg'][self.idx] = bg
335
+ self.cache['label'][self.idx] = label
336
+ # キャッシュ番号
337
+ self.idx = (self.idx + 1) % self.cache_size
338
+
339
+
340
+ # 前景のフラクタル画像作成
341
+ def render(self, sys):
342
+ # 乱数生成器コピー
343
+ rng = self.rng
344
+ # フラクタル座標と範囲
345
+ coords, region = self._iterate(sys)
346
+ # render the fractal at half resolution--it will be resized during generation phase
347
+ img = ifs.render(coords, self.size, binary=False, region=region, patch=self.patch)
348
+ return img
349
+
350
+
351
+ # キャッシュに前景と背景画像を追加
352
+ def add_sample(self, sys, label=-1):
353
+ # IFSパラメータの微少変動
354
+ sysc = self.jitter(sys)
355
+ # フラクタル画像生成
356
+ frac = self.render(sysc)
357
+ # 背景画像の生成
358
+ bg = self.render_background()
359
+ # MixUPとかできないのでラベルは使わない
360
+ # キャッシュに追加
361
+ self._update_cache(frac, bg, label)
362
+
363
+
364
+ # Multi-Fractal画像を作成
365
+ def generate(self, sys, label=-1, new_sample=True):
366
+ # 乱数生成器コピー
367
+ rng = self.rng
368
+
369
+ # 新規追加しない、
370
+ # インスタンス時に全キャッシュ作成済みでそこからとってくる
371
+ #if new_sample:
372
+ # self.add_sample(sys, label)
373
+
374
+ # 背景画像をキャッシュ内からランダムに選択
375
+ idx = self.bg_sel_cache_idx % len(self)#rng.integers(0, len(self))
376
+ self.bg_sel_cache_idx += 1
377
+ img = self.cache['bg'][idx].copy()
378
+ # 背景なしの場合はブランク画像
379
+ if self.background==False:
380
+ img = np.zeros(img.shape, np.uint8)
381
+
382
+ # ラベルは使わない
383
+ labels = []
384
+
385
+ # 混合数を乱数選択
386
+ n = rng.choice(range(self.n_objects[0], self.n_objects[1]+1), p=self.nobj_p)
387
+ # キャッシュサイズでキャップ
388
+ n = min(n, len(self))
389
+
390
+ # 混合数回分繰り返し
391
+ for i in range(n):
392
+ # キャッシュ番号選択
393
+ idx = self.sel_cache_idx % len(self) # rng.integers(0, len(self))#
394
+ self.sel_cache_idx += 1
395
+
396
+ # ラベルリストに追加
397
+ labels.append(self.cache['label'][idx])
398
+
399
+ # 前景画像取得
400
+ fg = self.cache['fg'][idx]
401
+ label = self.cache['label'][idx]
402
+
403
+ if True:
404
+ # データ拡張(転置、水平、上下反転)
405
+ # random flips
406
+ if self.flips:
407
+ fg = self.random_flips(fg)
408
+
409
+ # colorize
410
+ if self.color:
411
+ # カラー画像化
412
+ fg = self.to_color(fg, label)
413
+ else:
414
+ # 白黒画像化
415
+ fg = self.to_gray(fg)
416
+
417
+ # 0.2-0.6倍でリサイズ
418
+ # リサイズ倍率
419
+ f = rng.uniform(*self.size_range)
420
+ # リサイズ
421
+ s = int(f * self.size)
422
+ fg = resize(fg, (s, s), interpolation=INTER_CUBIC)
423
+
424
+ # データ拡張(回転)
425
+ if self.blur_p and rng.random() > self.blur_p:
426
+ # 回転角度
427
+ angle = rng.integers(-45, 45, 1)[0]
428
+ # 回転中心
429
+ height, width, channel = fg.shape
430
+ center = (int(width/2), int(height/2))
431
+ # 等倍
432
+ scale = 1.0
433
+ #getRotationMatrix2D関数を使用
434
+ trans = getRotationMatrix2D(center, angle , scale)
435
+ #アフィン変換
436
+ fg = warpAffine(fg, trans, (width,height))
437
+
438
+ # ランダム埋め込み位置
439
+ # random location
440
+ x, y = rng.integers(-(s//3), self.size-s+(s//3), 2)
441
+ x1 = 0 if x >= 0 else -x
442
+ x2 = s if x < self.size - s else self.size - x
443
+ y1 = 0 if y >= 0 else -y
444
+ y2 = s if y < self.size - s else self.size - y
445
+ fg = fg[y1:y2, x1:x2]
446
+ # add object to image
447
+ y = max(y, 0)
448
+ x = max(x, 0)
449
+ self.composite(fg, img[y:y+fg.shape[0], x:x+fg.shape[1]])
450
+ else:
451
+ fg_gray = fg.copy()
452
+ fg = self.to_color(fg)
453
+ f = 0.8#rng.uniform(*self.size_range)
454
+ # リサイズ
455
+ s = int(f * self.size)
456
+ fg = resize(fg, (s, s), interpolation=INTER_CUBIC)
457
+ fg_gray = resize(fg_gray, (s, s), interpolation=INTER_CUBIC)
458
+
459
+ x, y = 0, 0
460
+ x1 = 0 if x >= 0 else -x
461
+ x2 = s if x < self.size - s else self.size - x
462
+ y1 = 0 if y >= 0 else -y
463
+ y2 = s if y < self.size - s else self.size - y
464
+ fg = fg[y1:y2, x1:x2]
465
+ fg_gray = fg_gray[y1:y2, x1:x2]
466
+ y = max(y, 0)
467
+ x = max(x, 0)
468
+ self.composite(fg, img[y:y+fg.shape[0], x:x+fg.shape[1]])
469
+ # self.composite_v2(fg, img[y:y+fg.shape[0], x:x+fg.shape[1]], fg_gray)
470
+
471
+
472
+ #x, y = rng.integers(-(s//2), self.size-(s//2), 2)
473
+ # # 見切れ防止
474
+ # x, y = rng.integers(0, self.size-s, 2)
475
+ # # 見切れ範囲除外
476
+ # x1 = 0 if x >= 0 else -x
477
+ # x2 = s if x < self.size - s else self.size - x
478
+ # y1 = 0 if y >= 0 else -y
479
+ # y2 = s if y < self.size - s else self.size - y
480
+ # fg = fg[y1:y2, x1:x2]
481
+ # # add object to image
482
+ # y = max(y, 0)
483
+ # x = max(x, 0)
484
+ # # 埋め込み範囲のみ重畳描画する、背景画像に重畳していく
485
+ # self.composite(fg, img[y:y+fg.shape[0], x:x+fg.shape[1]])
486
+
487
+ # ボカし実施
488
+ # randomly apply gaussian blur
489
+ # if self.blur_p and rng.random() > self.blur_p:
490
+ # img = self.random_blur(img)
491
+
492
+ return img, labels
src/multi_fractal_db/serach_ifs_systems.py ADDED
@@ -0,0 +1,179 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import tqdm
2
+ import numpy as np
3
+
4
+ # IFSシステムのパラメータ探索
5
+ def random_systems(num_systems, n=(2,5), bval=None, beta=None, sample_fn=None):
6
+ '''Sample random systems.
7
+
8
+ Args:
9
+ num_systems (int): the number of systems to sample.
10
+ n (int or Tuple[int, int]): the size or range of sizes allowable for the systems.
11
+ bval (float): allowable magnitude of translation parameters.
12
+ beta (float or Tuple[float,float]): singular values constraint. See sample_systems.
13
+
14
+ Returns:
15
+ A list of dicts {'system': np.array} containing the system parameters.
16
+ '''
17
+ # 乱数生成器
18
+ rng = np.random.default_rng(seed=0)
19
+ systems = []
20
+ for i in tqdm.trange(num_systems):
21
+ # IFSシステムパラメータ取得
22
+ s = sample_system(n, bval=bval, beta=beta, rng=rng, sample_fn=sample_fn)
23
+ systems.append({'system': s})
24
+ return systems
25
+
26
+
27
+ def sample_system(n=None, constrain=True, bval=1, rng=None, beta=None, sample_fn=None):
28
+ '''Return n random affine transforms. If constrain=True, enforce the transforms
29
+ to be strictly contractive (by forcing singular values to be less than 1).
30
+
31
+ Args:
32
+ n (Union[range,Tuple[int,int],List[int,int],None]): range of values to sample from for the number of
33
+ transforms to sample. If None (default), then sample from range(2, 8).
34
+ constrain (bool): if True, enforce contractivity of transformations. Technically, an IFS must be
35
+ contractive; however, FractalDB does not enforce it during search, so it is left as an option here.
36
+ Default: True.
37
+ bval (Union[int,float]): maximum magnitude of the translation parameters sampled for each transform.
38
+ The translation parameters don't effect contractivity, and so can be chosen arbitrarily. Ignored and set
39
+ to 1 when constrain is False. Default: 1.
40
+ rng (Optional[numpy.random._generator.Generator]): random number generator. If None (default), it defaults
41
+ to np.random.default_rng().
42
+ beta (float or Tuple[float, float]): range for weighted sum of singular values when constrain==True. Let
43
+ q ~ U(beta[0], beta[1]), then we enforce $\sum_{i=0}^{n-1} (s^i_1 + 2*s^i_2) = q$.
44
+ sample_fn (callable): function used for sampling singular values. Should accept three arguments: n, for
45
+ the size of the system; a, for the sigma-factor; and rng, the random generator. When None (default),
46
+ uses sample_svs.
47
+
48
+ Returns:
49
+ Numpy array of shape (n, 2, 3), containing n sets of 2x3 affine transformation matrices.
50
+ '''
51
+ # Numpy乱数生成
52
+ if rng is None:
53
+ rng = np.random.default_rng(seed=0)
54
+
55
+ # 重ね合わせ方程式の数
56
+ if n is None:
57
+ n = rng.integers(2, 8)
58
+ elif isinstance(n, range):
59
+ n = rng.integers(n.start, n.stop)
60
+ elif isinstance(n, (tuple, list)):
61
+ n = rng.integers(*n)
62
+
63
+ # まともなフラクタル生成範囲
64
+ if beta is None:
65
+ beta = ((5 + n) / 2, (6 + n) / 2)
66
+
67
+ # 何かの関数
68
+ if sample_fn is None:
69
+ sample_fn = sample_svs
70
+
71
+ # まともな範囲のパラメータに制限
72
+ if constrain:
73
+ # sample a matrix with singular values < 1 (a contraction)
74
+ # 1. sample the singular vectors--random orthonormal matrices--by randomly rotating the standard basis
75
+ base = np.sign(rng.random((2*n, 2, 1)) - 0.5) * np.eye(2)
76
+ # 回転角度
77
+ angle = rng.uniform(-np.pi, np.pi, 2*n)
78
+ # 回転行列
79
+ ss = np.sin(angle)
80
+ cc = np.cos(angle)
81
+ rmat = np.empty((2 * n, 2, 2))
82
+ rmat[:, 0, 0] = cc
83
+ rmat[:, 0, 1] = -ss
84
+ rmat[:, 1, 0] = ss
85
+ rmat[:, 1, 1] = cc
86
+ uv = rmat @ base
87
+ u, v = uv[:n], uv[n:]
88
+ # 2. sample the singular values
89
+ a = rng.uniform(*beta)
90
+ s = sample_fn(n, a, rng)
91
+ # 3. sample the translation parameters from Uniform(-bval, bval) and create the transformation matrix
92
+ m = np.empty((n, 2, 3))
93
+ # 回転行列
94
+ m[:, :, :2] = u * s[:, None, :] @ v
95
+ # 並進行列
96
+ m[:, :, 2] = rng.uniform(-bval, bval, (n, 2))
97
+ else:
98
+ m = rng.uniform(-1, 1, (n, 2, 3))
99
+
100
+ return m
101
+
102
+
103
+
104
+ def sample_svs(n, a, rng=None):
105
+ '''Sample singular values. 2*`n` singular values are sampled such that the following conditions
106
+ are satisfied, for singular values sv_{i} and i = 0, ..., 2n-1:
107
+
108
+ 1. 0 <= sv_{i} <= 1
109
+ 2. sv_{2i} >= sv_{2i+1}
110
+ 3. w.T @ S = `a`, for S = [sv_{0}, ..., sv_{2n-1}] and w = [1, 2, ..., 1, 2]
111
+
112
+ Args:
113
+ n (int): number of pairs of singular values to sample.
114
+ a (float): constraint on the weighted sum of all singular values. Note that a must be in the
115
+ range (0, 3*n).
116
+ rng (Optional[numpy.random._generator.Generator]): random number generator. If None (default), it defaults
117
+ to np.random.default_rng().
118
+
119
+ Returns:
120
+ Numpy array of shape (n, 2) containing the singular values.
121
+ '''
122
+ if rng is None: rng = np.random.default_rng()
123
+ if a < 0: a == 0
124
+ elif a > 3*n: a == 3*n
125
+ s = np.empty((n, 2))
126
+ p = a
127
+ q = a - 3*n + 3
128
+ # sample the first 2*(n-1) singular values (first n-1 pairs)
129
+ for i in range(n - 1):
130
+ s1 = rng.uniform(max(0, q/3), min(1, p))
131
+ q -= s1
132
+ p -= s1
133
+ s2 = rng.uniform(max(0, q/2), min(s1, p/2))
134
+ q = q - 2 * s2 + 3
135
+ p -= 2 * s2
136
+ s[i, :] = s1, s2
137
+ # sample the last pair of singular values
138
+ s2 = rng.uniform(max(0, (p-1)/2), p/3)
139
+ s1 = p - 2*s2
140
+ s[-1, :] = s1, s2
141
+
142
+ return s
143
+
144
+ def sample_svs_rej(n, a, rng=None):
145
+ '''Sample singular values uniformly from the joint distribution over the n-dimensional surface
146
+ defined by the constraints (see sample_svs). Uniform sampling is achieved by means of rejection
147
+ sampling.
148
+
149
+ Args:
150
+ n (int): number of pairs of singular values to sample.
151
+ a (float): constraint on the weighted sum of all singular values. Note that a must be in the
152
+ range (0, 3*n).
153
+ rng (Optional[numpy.random._generator.Generator]): random number generator. If None (default), it defaults
154
+ to np.random.default_rng().
155
+
156
+ Returns:
157
+ Numpy array of shape (n, 2) containing the singular values.
158
+ '''
159
+ if rng is None:
160
+ rng = np.random.default_rng()
161
+ if a < 0: a = 0
162
+ elif a > 3 * n: a = 3 * n
163
+
164
+ w = np.ones(2 * n - 1)
165
+ w[1::2] = 2
166
+ s = np.zeros((n, 2))
167
+ for i in range(1000):
168
+ s.ravel()[:-1] = rng.random(2 * n - 1)
169
+ # restrict to below the y=x line
170
+ r = s[:, 1] > s[:, 0]
171
+ s[r, :] = s[r][:, ::-1]
172
+ # check if valid or reject
173
+ b = (a - w @ s.ravel()[:-1]) / 2
174
+ if b <= s[-1, 0] and b >= 0:
175
+ s[-1, 1] = b
176
+ break
177
+ else:
178
+ print('Rejection sampling failed')
179
+ return s
src/validator.py ADDED
@@ -0,0 +1,158 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import os
2
+ from PIL import Image
3
+ from tqdm import tqdm
4
+
5
+ class Validator():
6
+ def __init__(self, data_format: dict, verbose=False) -> None:
7
+ assert isinstance(data_format, dict)
8
+ self.data_format = data_format
9
+ if verbose:
10
+ print('\nValidation details:')
11
+ for k, v in self.data_format.items():
12
+ print(' {}: {}'.format(k, v))
13
+ self.data = None
14
+ print('\nValidation:')
15
+
16
+
17
+ def check_data(self, result) -> None:
18
+ raise NotImplementedError
19
+
20
+
21
+ def check_samples(self, result) -> None:
22
+ raise NotImplementedError
23
+
24
+
25
+ def check_dtype(self, result) -> None:
26
+ raise NotImplementedError
27
+
28
+
29
+ def check_keys(self, result) -> None:
30
+ raise NotImplementedError
31
+
32
+
33
+ def check_details(self, result) -> None:
34
+ raise NotImplementedError
35
+
36
+
37
+ def validate(self, result) -> None:
38
+ self.check_data(result)
39
+ self.check_samples(result)
40
+ self.check_dtype(result)
41
+ self.check_keys(result)
42
+ self.check_details(result)
43
+
44
+
45
+ def get_data(self) -> None:
46
+ return self.data
47
+
48
+
49
+ class ImageFolderValidator(Validator):
50
+ def check_data(self, result) -> None:
51
+ msg = ' Checking data...'
52
+ print(msg, end='\r')
53
+ for category in tqdm(os.listdir(result)):
54
+ if not os.path.isdir(os.path.join(result, category)):
55
+ raise NotADirectoryError('Not a directory.')
56
+ if len(os.listdir(os.path.join(result, category))) == 0:
57
+ raise NullError('No data in {}'.format(category))
58
+ print(msg+' Done')
59
+
60
+
61
+ def check_samples(self, result) -> None:
62
+ msg = ' Checking samples...'
63
+ print(msg, end='\r')
64
+ for category in tqdm(os.listdir(result)):
65
+ for image_path in os.listdir(os.path.join(result, category)):
66
+ try:
67
+ img = Image.open(os.path.join(result, category, image_path))
68
+ except:
69
+ raise SampleError('Missing samples or invalid samples found.')
70
+ print(msg+' Done')
71
+
72
+
73
+ def check_dtype(self, result) -> None:
74
+ msg = ' Checking dtype...'
75
+ print(msg, end='\r')
76
+ num_channels = None
77
+ size = None
78
+ count = 0
79
+ for category in tqdm(os.listdir(result)):
80
+ for image_path in os.listdir(os.path.join(result, category)):
81
+ img = Image.open(os.path.join(result, category, image_path))
82
+ c = len(img.getbands())
83
+ if num_channels != c or size != img.size:
84
+ count += 1
85
+ num_channels = c
86
+ size = img.size
87
+ if count >= 2:
88
+ raise DimError('Dim mismatch found.')
89
+ print(msg+' Done')
90
+
91
+
92
+ def check_keys(self, result) -> None:
93
+ pass
94
+
95
+
96
+ def check_details(self, result) -> None:
97
+ msg = ' Checking details...'
98
+ num_categories = self.data_format['num_categories']
99
+ num_images = self.data_format['num_images']
100
+ if len(os.listdir(result))!=num_categories:
101
+ raise NumCategoryError('Number of categories is not {}'.format(num_categories))
102
+ for category in tqdm(os.listdir(result)):
103
+ image_paths = os.listdir(os.path.join(result, category))
104
+ if len(image_paths)!=num_images:
105
+ raise NumImageError('Number of images is not {} in {}'.format(num_images, category))
106
+ print(msg+' Done')
107
+
108
+
109
+ class NotADirectoryError(Exception):
110
+ pass
111
+
112
+
113
+ class SampleError(Exception):
114
+ pass
115
+
116
+
117
+ class DimError(Exception):
118
+ pass
119
+
120
+
121
+ class DtypeError(Exception):
122
+ pass
123
+
124
+
125
+ class ExtentionError(Exception):
126
+ pass
127
+
128
+
129
+ class DelimiterError(Exception):
130
+ pass
131
+
132
+
133
+ class NumColumnsError(Exception):
134
+ pass
135
+
136
+
137
+ class NullError(Exception):
138
+ pass
139
+
140
+
141
+ class DiscreteDataError(Exception):
142
+ pass
143
+
144
+
145
+ class MaximumExceedError(Exception):
146
+ pass
147
+
148
+
149
+ class InstanceError(Exception):
150
+ pass
151
+
152
+
153
+ class NumCategoryError(Exception):
154
+ pass
155
+
156
+
157
+ class NumImageError(Exception):
158
+ pass