質問編集履歴
1
ソースコードを追加しました
title
CHANGED
|
File without changes
|
body
CHANGED
|
@@ -30,7 +30,149 @@
|
|
|
30
30
|
```GoogleColaboratory
|
|
31
31
|
!python make_dataset.py ../../datasets/facades/ 3 --img_size 256
|
|
32
32
|
```
|
|
33
|
+
※ソースコードを追加しました。(make_dataset.py)
|
|
33
34
|
|
|
35
|
+
```Python
|
|
36
|
+
import os
|
|
37
|
+
import cv2
|
|
38
|
+
import h5py
|
|
39
|
+
import parmap
|
|
40
|
+
import argparse
|
|
41
|
+
import numpy as np
|
|
42
|
+
from pathlib import Path
|
|
43
|
+
from tqdm import tqdm as tqdm
|
|
44
|
+
import matplotlib.pylab as plt
|
|
45
|
+
|
|
46
|
+
|
|
47
|
+
def format_image(img_path, size, nb_channels):
|
|
48
|
+
"""
|
|
49
|
+
Load img with opencv and reshape
|
|
50
|
+
"""
|
|
51
|
+
|
|
52
|
+
if nb_channels == 1:
|
|
53
|
+
img = cv2.imread(img_path, 0)
|
|
54
|
+
img = np.expand_dims(img, axis=-1)
|
|
55
|
+
else:
|
|
56
|
+
img = cv2.imread(img_path)
|
|
57
|
+
img = img[:, :, ::-1] # GBR to RGB
|
|
58
|
+
|
|
59
|
+
w = img.shape[1]
|
|
60
|
+
|
|
61
|
+
# Slice image in 2 to get both parts
|
|
62
|
+
img_full = img[:, :w // 2, :]
|
|
63
|
+
img_sketch = img[:, w // 2:, :]
|
|
64
|
+
|
|
65
|
+
img_full = cv2.resize(img_full, (size, size), interpolation=cv2.INTER_AREA)
|
|
66
|
+
img_sketch = cv2.resize(img_sketch, (size, size), interpolation=cv2.INTER_AREA)
|
|
67
|
+
|
|
68
|
+
if nb_channels == 1:
|
|
69
|
+
img_full = np.expand_dims(img_full, -1)
|
|
70
|
+
img_sketch = np.expand_dims(img_sketch, -1)
|
|
71
|
+
|
|
72
|
+
img_full = np.expand_dims(img_full, 0).transpose(0, 3, 1, 2)
|
|
73
|
+
img_sketch = np.expand_dims(img_sketch, 0).transpose(0, 3, 1, 2)
|
|
74
|
+
|
|
75
|
+
return img_full, img_sketch
|
|
76
|
+
|
|
77
|
+
|
|
78
|
+
def build_HDF5(jpeg_dir, nb_channels, data_dir, size=256):
|
|
79
|
+
"""
|
|
80
|
+
Gather the data in a single HDF5 file.
|
|
81
|
+
"""
|
|
82
|
+
|
|
83
|
+
data_dir = os.path.join(data_dir, 'processed')
|
|
84
|
+
|
|
85
|
+
# Put train data in HDF5
|
|
86
|
+
file_name = os.path.basename(jpeg_dir.rstrip("/"))
|
|
87
|
+
hdf5_file = os.path.join(data_dir, "%s_data.h5" % file_name)
|
|
88
|
+
with h5py.File(hdf5_file, "w") as hfw:
|
|
89
|
+
|
|
90
|
+
for dset_type in ["train", "test", "val"]:
|
|
91
|
+
|
|
92
|
+
list_img = [img for img in Path(jpeg_dir).glob('%s/*.jpg' % dset_type)]
|
|
93
|
+
list_img = [str(img) for img in list_img]
|
|
94
|
+
list_img.extend(list(Path(jpeg_dir).glob('%s/*.png' % dset_type)))
|
|
95
|
+
list_img = list(map(str, list_img))
|
|
96
|
+
list_img = np.array(list_img)
|
|
97
|
+
|
|
98
|
+
data_full = hfw.create_dataset("%s_data_full" % dset_type,
|
|
99
|
+
(0, nb_channels, size, size),
|
|
100
|
+
maxshape=(None, 3, size, size),
|
|
101
|
+
dtype=np.uint8)
|
|
102
|
+
|
|
103
|
+
data_sketch = hfw.create_dataset("%s_data_sketch" % dset_type,
|
|
104
|
+
(0, nb_channels, size, size),
|
|
105
|
+
maxshape=(None, 3, size, size),
|
|
106
|
+
dtype=np.uint8)
|
|
107
|
+
|
|
108
|
+
num_files = len(list_img)
|
|
109
|
+
chunk_size = 100
|
|
110
|
+
num_chunks = num_files / chunk_size
|
|
111
|
+
arr_chunks = np.array_split(np.arange(num_files), num_chunks)
|
|
112
|
+
|
|
113
|
+
for chunk_idx in tqdm(arr_chunks):
|
|
114
|
+
|
|
115
|
+
list_img_path = list_img[chunk_idx].tolist()
|
|
116
|
+
output = parmap.map(format_image, list_img_path, size, nb_channels, pm_parallel=False)
|
|
117
|
+
|
|
118
|
+
arr_img_full = np.concatenate([o[0] for o in output], axis=0)
|
|
119
|
+
arr_img_sketch = np.concatenate([o[1] for o in output], axis=0)
|
|
120
|
+
|
|
121
|
+
# Resize HDF5 dataset
|
|
122
|
+
data_full.resize(data_full.shape[0] + arr_img_full.shape[0], axis=0)
|
|
123
|
+
data_sketch.resize(data_sketch.shape[0] + arr_img_sketch.shape[0], axis=0)
|
|
124
|
+
|
|
125
|
+
data_full[-arr_img_full.shape[0]:] = arr_img_full.astype(np.uint8)
|
|
126
|
+
data_sketch[-arr_img_sketch.shape[0]:] = arr_img_sketch.astype(np.uint8)
|
|
127
|
+
|
|
128
|
+
def check_HDF5(jpeg_dir, nb_channels):
|
|
129
|
+
"""
|
|
130
|
+
Plot images with landmarks to check the processing
|
|
131
|
+
"""
|
|
132
|
+
|
|
133
|
+
# Get hdf5 file
|
|
134
|
+
file_name = os.path.basename(jpeg_dir.rstrip("/"))
|
|
135
|
+
hdf5_file = os.path.join(data_dir, "%s_data.h5" % file_name)
|
|
136
|
+
|
|
137
|
+
with h5py.File(hdf5_file, "r") as hf:
|
|
138
|
+
data_full = hf["train_data_full"]
|
|
139
|
+
data_sketch = hf["train_data_sketch"]
|
|
140
|
+
for i in range(data_full.shape[0]):
|
|
141
|
+
plt.figure()
|
|
142
|
+
img = data_full[i, :, :, :].transpose(1,2,0)
|
|
143
|
+
img2 = data_sketch[i, :, :, :].transpose(1,2,0)
|
|
144
|
+
img = np.concatenate((img, img2), axis=1)
|
|
145
|
+
if nb_channels == 1:
|
|
146
|
+
plt.imshow(img[:, :, 0], cmap="gray")
|
|
147
|
+
else:
|
|
148
|
+
plt.imshow(img)
|
|
149
|
+
plt.show()
|
|
150
|
+
plt.clf()
|
|
151
|
+
plt.close()
|
|
152
|
+
|
|
153
|
+
|
|
154
|
+
if __name__ == '__main__':
|
|
155
|
+
|
|
156
|
+
parser = argparse.ArgumentParser(description='Build dataset')
|
|
157
|
+
parser.add_argument('jpeg_dir', type=str, help='path to jpeg images')
|
|
158
|
+
parser.add_argument('nb_channels', type=int, help='number of image channels')
|
|
159
|
+
parser.add_argument('--img_size', default=256, type=int,
|
|
160
|
+
help='Desired Width == Height')
|
|
161
|
+
parser.add_argument('--do_plot', action="store_true",
|
|
162
|
+
help='Plot the images to make sure the data processing went OK')
|
|
163
|
+
parser.add_argument('--data_dir', default='../../data', type=str, help='Data directory')
|
|
164
|
+
args = parser.parse_args()
|
|
165
|
+
|
|
166
|
+
build_HDF5(args.jpeg_dir,
|
|
167
|
+
args.nb_channels,
|
|
168
|
+
args.data_dir,
|
|
169
|
+
size=args.img_size)
|
|
170
|
+
|
|
171
|
+
if args.do_plot:
|
|
172
|
+
check_HDF5(args.jpeg_dir, args.nb_channels)
|
|
173
|
+
|
|
174
|
+
|
|
175
|
+
```
|
|
34
176
|
### 試したこと
|
|
35
177
|
|
|
36
178
|
データセットの作り方に問題があると考え調べてみたところ、データセットの画像には何かしら意味付けがされているのではないか、単に編集アプリで作るだけではいけないのではないかと考えました。が、理解が浅く、どうすればよいのか分からない状態です。
|