-
Notifications
You must be signed in to change notification settings - Fork 11
/
loader.py
302 lines (276 loc) · 10.8 KB
/
loader.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
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
from torch.utils.data import Dataset
import h5py
import numpy as np
import torch as th
class LargeSample(Dataset):
def __init__(self, data_path, region, pad, data_flag, div=(5,5)):
super(LargeSample, self).__init__()
self.data_path = data_path
self.region = region
self.pad = pad
self.data_flag = data_flag
(self.row, self.col) = div
def __len__(self):
return self.row*self.col
def __getitem__(self, index):
row, col = index//self.col, index-(index//self.col)*self.col
with h5py.File(self.data_path, 'r') as f:
(_, h, w) = f[self.region][self.data_flag]['gt'].shape
h_len, w_len = h//self.row, w//self.col
sample = {
'data': f[self.region][self.data_flag]['data'][
:,
row*h_len:(row+1)*h_len+2*self.pad,
col*w_len:(col+1)*w_len+2*self.pad
],
'gt': f[self.region][self.data_flag]['gt'][
0,
row*h_len:(row+1)*h_len,
col*w_len:(col+1)*w_len
],
'index': (row, col),
'div': (self.row, self.col)
}
return sample
class LandslideTrainDataset(Dataset):
def __init__(self, path, region, stride, ws, pts, oversample_path, pad=64, feature_num=94):
super(LandslideTrainDataset, self).__init__()
self.path = path
self.ws = ws
self.stride = stride
self.region = region
self.pad = pad
self.feature_num = feature_num
self.pts = pts #nx4 matrix consisting of r1, c1, r2, c2 points, r1<r2 & c1<c2
self.data_len = self.len()
self.pts_len = self.len_oversample() #nx1 matrix containing length of the data corresponding to pts
self.oversample_path = oversample_path
def identify_idx(self, index):
cum_sum = 0
for i in range(self.pts_len.shape[0]):
cum_sum += self.pts_len[i]
if index < self.data_len+cum_sum:
return i
raise ValueError
def len_oversample(self):
stride = self.stride//4
pts_len = np.zeros((self.pts.shape[0], 1))
for i in range(self.pts.shape[0]):
h = self.pts[i,2] - self.pts[i,0]
w = self.pts[i,3] - self.pts[i,1]
hnum = (h-self.ws)//stride + 1
wnum = (w-self.ws)//stride + 1
pts_len[i] = hnum*wnum
return pts_len
def len(self):
with h5py.File(self.path, 'r') as f:
(_, h, w) = f[self.region]['train']['gt'].shape
hnum = (h-self.ws)//self.stride + 1
wnum = (w-self.ws)//self.stride + 1
return hnum*wnum
def __len__(self):
return self.data_len + int(np.sum(self.pts_len))
def get_item(self, index, dataset, gt, stride):
(_, _, wlen) = gt.shape
wnum = (wlen-self.ws)//stride + 1
row = index//wnum
col = index - row*wnum
sample = {
'data': th.tensor(
dataset[
:,
row*stride:row*stride+self.ws+2*self.pad,
col*stride:col*stride+self.ws+2*self.pad
]
),
'gt': th.tensor(
gt[
:,
row*stride:row*stride+self.ws,
col*stride:col*stride+self.ws
]
),
'index': (row, col)
}
return sample
def __getitem__(self, index):
if index < self.data_len:
with h5py.File(self.path, 'r') as f:
dataset = f[self.region]['train']['data']
gt = f[self.region]['train']['gt']
return self.get_item(index, dataset, gt, self.stride)
else:
with h5py.File(self.oversample_path, 'r') as g:
pts_idx = self.identify_idx(index)
# r1, c1, r2, c2 = self.pts[pts_idx, :]
n_index = index-self.data_len if pts_idx==0 else index-(self.data_len+int(np.sum(self.pts_len[:pts_idx])))
dataset = g[str(pts_idx)]['data']
gt = g[str(pts_idx)]['gt']
return self.get_item(n_index, dataset, gt, self.stride//4)
class SampledPixDataset(Dataset):
def __init__(self, data_path, data_indices_path, region, pad, data_flag):
super(SampledPixDataset, self).__init__()
self.data_path = data_path
self.indices = np.load(data_indices_path) # a nx2 array
self.region = region
self.pad = pad
self.data_flag = data_flag
def __len__(self):
return len(self.indices)
def __getitem__(self, index):
row, col = self.indices[index][0], self.indices[index][1]
with h5py.File(self.data_path, 'r') as f:
sample = {
'data': f[self.region][self.data_flag]['data'][:, row-self.pad:row+self.pad+1, col-self.pad:col+self.pad+1],
'gt': f[self.region][self.data_flag]['gt'][:, row, col],
'index': (row, col)
}
return sample
class PixDataset(Dataset):
def __init__(self, path, region, data_flag, pad=32):
super(PixDataset, self).__init__()
self.path = path
self.region = region
self.pad = pad
self.data_flag = data_flag
self.shape = self.get_shape()
def get_shape(self):
with h5py.File(self.path, 'r') as f:
(_, h, w) = f[self.region][self.data_flag]['gt'].shape
return (h,w)
def __len__(self):
with h5py.File(self.path, 'r') as f:
(_, h, w) = f[self.region][self.data_flag]['gt'].shape
return h*w
def __getitem__(self, index):
(_, w) = self.shape
row = index//w
col = index - (index//w)*w
with h5py.File(self.path, 'r') as f:
sample = {
'data': f[self.region][self.data_flag]['data'][
:,
row:row+2*self.pad+1,
col:col+2*self.pad+1
],
'gt': f[self.region][self.data_flag]['gt'][
:,
row,
col
],
'index': (row, col)
}
return sample
class LandslideDataset(Dataset):
'''
This class doesn't support different stride sizes and oversampling.
When testing, we don't need to have stride smaller than ws.
Also, we don't need to oversample.
'''
def __init__(self, data_path, indices, region, ws, pad, prune):
super(LandslideDataset, self).__init__()
self.indices = indices # validation and train indices should be handled in the indices path
self.data_path = data_path
self.ws = ws
self.region = region
self.pad = pad
self.prune = prune
def __len__(self):
return self.indices.shape[0]
def __getitem__(self, index):
with h5py.File(self.data_path, 'r') as f:
entry = self.indices[index, :]
row, col = int(entry[0]), int(entry[1])
sample = {
'data': th.tensor(
f[self.region]['data'][
:,
row*self.ws+self.pad-self.prune:(row+1)*self.ws+self.pad+self.prune,
col*self.ws+self.pad-self.prune:(col+1)*self.ws+self.pad+self.prune
]
),
'gt': th.tensor(
f[self.region]['gt'][
:,
row*self.ws:(row+1)*self.ws,
col*self.ws:(col+1)*self.ws
]
),
'index': (row, col)
}
return sample
class DistLandslideDataset(LandslideDataset):
def __init__(self, data_path, indices, region, ws, pad, prune, dist_num):
super(DistLandslideDataset, self).__init__(data_path, indices, region, ws, pad, prune)
self.dist_num = dist_num
def __getitem__(self, index):
with h5py.File(self.data_path, 'r') as f:
entry = self.indices[index, :]
row, col = int(entry[0]), int(entry[1])
data = f[self.region]['data/dist0'][
:,
row*self.ws+self.pad-self.prune:(row+1)*self.ws+self.pad+self.prune,
col*self.ws+self.pad-self.prune:(col+1)*self.ws+self.pad+self.prune
]
for i in range(self.dist_num):
data = np.concatenate(
(
data,
f[self.region]['data/dist{}'.format(str(i+1))][
:,
row*self.ws+self.pad-self.prune:(row+1)*self.ws+self.pad+self.prune,
col*self.ws+self.pad-self.prune:(col+1)*self.ws+self.pad+self.prune
]
),
axis=0
)
sample = {
'data': th.tensor(data),
'gt': th.tensor(
f[self.region]['gt'][
:,
row*self.ws:(row+1)*self.ws,
col*self.ws:(col+1)*self.ws
]
),
'index': (row, col)
}
del data
return sample
def initilize_data_oversample(args, fw):
with h5py.File(args.data_path, 'r') as f:
dataset = f[args.region]['train']['data']
gt = f[args.region]['train']['gt']
for idx in range(args.oversample_pts.shape[0]):
r1, c1, r2, c2 = args.oversample_pts[idx, :]
h, w = r2-r1, c2-c1
fw[str(idx)]['data'][:, 0:args.pad, :] = 0
fw[str(idx)]['data'][:, h+args.pad:, :] = 0
fw[str(idx)]['data'][:, :, 0:args.pad] = 0
fw[str(idx)]['data'][:, :, w+args.pad:] = 0
fw[str(idx)]['data'][:, args.pad:args.pad+h, args.pad:args.pad+w] = dataset[:, r1:r2, c1:c2]
fw[str(idx)]['gt'][:][:, :, :] = gt[:, r1:r2, c1:c2]
return fw
def create_oversample_data(args):
path = '/'.join(args.data_path.split('/')[:-1])+'/'+args.region+'_oversample.h5'
# path = '/home/ainaz/Projects/Landslides/image_data/'+args.region+'_oversample.h5'
if not args.oversample:
return path
f = h5py.File(path, 'a')
for i in range(args.oversample_pts.shape[0]):
r1, c1, r2, c2 = args.oversample_pts[i, :]
f.create_dataset(
str(i)+'/data',
(args.feature_num, r2-r1+2*args.pad, c2-c1+2*args.pad),
dtype='f',
compression='lzf'
)
f.create_dataset(
str(i)+'/gt',
(1, r2-r1, c2-c1),
dtype='f',
compression='lzf'
)
f = initilize_data_oversample(args, f)
f.close()
return path