-
Notifications
You must be signed in to change notification settings - Fork 0
/
Copy pathGenDatasetClass.py
237 lines (207 loc) · 8.87 KB
/
GenDatasetClass.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
from torch.utils.data import Dataset, DataLoader
from PIL import Image, ImageEnhance
from osgeo import gdal
from torchvision import transforms
import glob
import torch as tc
import numpy as np
class FarmDataset(Dataset):
def __init__(self, istrain=True, isaug=True):
self.istrain = istrain
self.trainxformat = './data/train/data1500/*.bmp'
self.trainyformat = './data/train/label1500/*.bmp'
self.testxformat = './data/test/*.png'
self.fns = glob.glob(self.trainxformat) if istrain else glob.glob(self.testxformat)
self.length = len(self.fns)
self.transforms = transforms
self.isaug = isaug
def __len__(self):
# total length is 2217
return self.length
def __getitem__(self, idx):
if self.istrain:
imgxname = self.fns[idx]
sampleimg = Image.open(imgxname)
imgyname = imgxname.replace('data1500', 'label1500')
targetimg = Image.open(imgyname).convert('L')
# sampleimg.save('original.bmp')
# data augmentation
if self.isaug:
sampleimg, targetimg = self.imgtrans(sampleimg, targetimg)
# check the result of dataautmentation
# sampleimg.save('sampletmp.bmp')
# targetimg.save('targettmp.bmp')
sampleimg = transforms.ToTensor()(sampleimg)
# targetimg=transforms.ToTensor()(targetimg).squeeze(0).long()
targetimg = np.array(targetimg)
targetimg = tc.from_numpy(targetimg).long() # to tensor
# print(sampleimg.shape,targetimg.shape)
return sampleimg, targetimg
else:
return gdal.Open(self.fns[idx])
def imgtrans(self, x, y, outsize=1024):
'''input is a PIL image
image dataaugumentation
return also aPIL image。
'''
# rotate should consider y
degree = np.random.randint(360)
x = x.rotate(degree, resample=Image.NEAREST, fillcolor=0)
y = y.rotate(degree, resample=Image.NEAREST, fillcolor=0) # here should be carefull, in case of label damage
# random do the input image augmentation
if np.random.random() > 0.5:
# sharpness
factor = 0.5 + np.random.random()
enhancer = ImageEnhance.Sharpness(x)
x = enhancer.enhance(factor)
if np.random.random() > 0.5:
# color augument
factor = 0.5 + np.random.random()
enhancer = ImageEnhance.Color(x)
x = enhancer.enhance(factor)
if np.random.random() > 0.5:
# contrast augument
factor = 0.5 + np.random.random()
enhancer = ImageEnhance.Contrast(x)
x = enhancer.enhance(factor)
if np.random.random() > 0.5:
# brightness
factor = 0.5 + np.random.random()
enhancer = ImageEnhance.Brightness(x)
x = enhancer.enhance(factor)
# img flip
transtypes = [Image.FLIP_LEFT_RIGHT, Image.FLIP_TOP_BOTTOM,
Image.ROTATE_90, Image.ROTATE_180, Image.ROTATE_270]
transtype = transtypes[np.random.randint(len(transtypes))]
x = x.transpose(transtype)
y = y.transpose(transtype)
# img resize between 0.8-1.2
w, h = x.size
factor = 1 + np.random.normal() / 5
if factor > 1.2: factor = 1.2
if factor < 0.8: factor = 0.8
# print(factor,x.size)
x = x.resize((int(w * factor), int(h * factor)), Image.NEAREST)
y = y.resize((int(w * factor), int(h * factor)), Image.NEAREST)
# random crop
w, h = x.size
stx = np.random.randint(w - outsize)
sty = np.random.randint(h - outsize)
# print((stx,sty,outsize,outsize))
x = x.crop((stx, sty, stx + outsize, sty + outsize)) # stx,sty,width,height
y = y.crop((stx, sty, stx + outsize, sty + outsize))
# print(x.size,y.size)
return x, y # return outsized pil image
# class valDataset(Dataset):
# def __init__(self):
# self.valxformat = './data/val/data1500/*.bmp'
# self.valyformat = './data/val/label1500/*.bmp'
# self.fns = glob.glob(self.valxformat)
# self.length = len(self.fns)
# self.transforms = transforms
#
# def __len__(self):
# # total length is 2217
# return self.length
#
# def __getitem__(self, idx):
# imgxname = self.fns[idx]
# sampleimg = Image.open(imgxname)
# imgyname = imgxname.replace('data1500', 'label1500')
# targetimg = Image.open(imgyname).convert('L')
# sampleimg = transforms.ToTensor()(sampleimg)
# targetimg = np.array(targetimg)
# targetimg = tc.from_numpy(targetimg).long() # to tensor
# return sampleimg, targetimg
class valDataset(Dataset):
def __init__(self, istrain=True, isaug=True):
self.istrain = istrain
self.trainxformat = './data/val/data1500/*.bmp'
self.trainyformat = './data/val/label1500/*.bmp'
self.testxformat = './data/test/*.png'
self.fns = glob.glob(self.trainxformat) if istrain else glob.glob(self.testxformat)
self.length = len(self.fns)
self.transforms = transforms
self.isaug = isaug
def __len__(self):
# total length is 2217
return self.length
def __getitem__(self, idx):
if self.istrain:
imgxname = self.fns[idx]
sampleimg = Image.open(imgxname)
imgyname = imgxname.replace('data1500', 'label1500')
targetimg = Image.open(imgyname).convert('L')
# sampleimg.save('original.bmp')
# data augmentation
if self.isaug:
sampleimg, targetimg = self.imgtrans(sampleimg, targetimg)
# check the result of dataautmentation
# sampleimg.save('sampletmp.bmp')
# targetimg.save('targettmp.bmp')
sampleimg = transforms.ToTensor()(sampleimg)
# targetimg=transforms.ToTensor()(targetimg).squeeze(0).long()
targetimg = np.array(targetimg)
targetimg = tc.from_numpy(targetimg).long() # to tensor
# print(sampleimg.shape,targetimg.shape)
return sampleimg, targetimg
else:
return gdal.Open(self.fns[idx])
def imgtrans(self, x, y, outsize=1024):
'''input is a PIL image
image dataaugumentation
return also aPIL image。
'''
# rotate should consider y
degree = np.random.randint(360)
x = x.rotate(degree, resample=Image.NEAREST, fillcolor=0)
y = y.rotate(degree, resample=Image.NEAREST, fillcolor=0) # here should be carefull, in case of label damage
# random do the input image augmentation
if np.random.random() > 0.5:
# sharpness
factor = 0.5 + np.random.random()
enhancer = ImageEnhance.Sharpness(x)
x = enhancer.enhance(factor)
if np.random.random() > 0.5:
# color augument
factor = 0.5 + np.random.random()
enhancer = ImageEnhance.Color(x)
x = enhancer.enhance(factor)
if np.random.random() > 0.5:
# contrast augument
factor = 0.5 + np.random.random()
enhancer = ImageEnhance.Contrast(x)
x = enhancer.enhance(factor)
if np.random.random() > 0.5:
# brightness
factor = 0.5 + np.random.random()
enhancer = ImageEnhance.Brightness(x)
x = enhancer.enhance(factor)
# img flip
transtypes = [Image.FLIP_LEFT_RIGHT, Image.FLIP_TOP_BOTTOM,
Image.ROTATE_90, Image.ROTATE_180, Image.ROTATE_270]
transtype = transtypes[np.random.randint(len(transtypes))]
x = x.transpose(transtype)
y = y.transpose(transtype)
# img resize between 0.8-1.2
w, h = x.size
factor = 1 + np.random.normal() / 5
if factor > 1.2: factor = 1.2
if factor < 0.8: factor = 0.8
# print(factor,x.size)
x = x.resize((int(w * factor), int(h * factor)), Image.NEAREST)
y = y.resize((int(w * factor), int(h * factor)), Image.NEAREST)
# random crop
w, h = x.size
stx = np.random.randint(w - outsize)
sty = np.random.randint(h - outsize)
# print((stx,sty,outsize,outsize))
x = x.crop((stx, sty, stx + outsize, sty + outsize)) # stx,sty,width,height
y = y.crop((stx, sty, stx + outsize, sty + outsize))
# print(x.size,y.size)
return x, y # return outsized pil image
if __name__ == '__main__':
d = FarmDataset(istrain=True)
x, y = d[2216]
print(x.shape)
print(y.shape)