-
Notifications
You must be signed in to change notification settings - Fork 2
/
dataloaders.py
251 lines (219 loc) · 10 KB
/
dataloaders.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
import os
import cv2
import torchvision.transforms as tfs
import torch.utils.data
import numpy as np
from skimage import io
import pickle
def get_train_val_dataloaders(cfgs):
batch_size = cfgs.get('batch_size', 8)
num_workers = cfgs.get('num_workers', 4)
image_size = cfgs.get('image_size', 256)
crop = cfgs.get('crop', None)
use_mask = cfgs.get('use_mask', False)
normalize = cfgs.get('normalize', True)
data_dir = cfgs.get('data_dir', 'data/gbh/')
data_split_dirs = cfgs.get('data_split_dirs', "data/split1+/")
if isinstance(data_split_dirs, str):
data_split_dirs = [data_split_dirs]
data_train = {os.path.basename(data_split_dir): os.path.join(data_split_dir, 'train.txt') for data_split_dir in data_split_dirs}
data_val = {os.path.basename(data_split_dir): os.path.join(data_split_dir, 'val.txt') for data_split_dir in data_split_dirs}
image_paths = get_image_list(data_dir, list(data_train.values())+list(data_val.values()), False, use_vis)
if not os.path.exists(os.path.join(data_dir, 'image_stats.pickle')):
online_get_image_stats(data_dir, image_paths)
ndsm_paths = get_image_list(data_dir, list(data_train.values())+list(data_val.values()), True)
if not os.path.exists(os.path.join(data_dir, 'ndsm_stats.pickle')):
online_get_ndsm_stats(data_dir, ndsm_paths)
overfit = cfgs.get('overfit', False)
train_loader = val_loader = None
get_loader = lambda **kargs: get_tri_image_loader(**kargs, data_dir=data_dir, use_mask=use_mask, num_workers=num_workers, image_size=image_size, crop=crop, normalize=normalize, overfit=overfit)
assert os.path.isdir(data_dir), "Data directory does not exist: %s" %data_dir
print(f"Loading training data from {data_train}")
train_loader = get_loader(data_split=data_train, is_validation=False, batch_size=batch_size)
print(f"Loading validation data from {data_val}")
if overfit:
val_loader = train_loader
else:
val_loader = get_loader(data_split=data_val, is_validation=True, batch_size=batch_size)
return train_loader, val_loader
def get_test_dataloaders(cfgs):
num_workers = cfgs.get('num_workers', 4)
image_size = cfgs.get('image_size', 256)
crop = cfgs.get('crop', None)
normalize = cfgs.get('normalize', True)
use_mask = cfgs.get('test_use_mask', True)
data_dir = cfgs.get('data_dir', 'data/gbh/')
data_split_dirs = cfgs.get('test_data_split_dirs', ["data/SAO+", "data/MUC+", "data/GUA+", "data/split1+"])
if isinstance(data_split_dirs, str):
data_split_dirs = [data_split_dirs]
data_test = {os.path.basename(data_split_dir): os.path.join(data_split_dir, 'test.txt') for data_split_dir in data_split_dirs}
test_loader = {}
get_loader = lambda **kargs: get_tri_image_loader(**kargs, data_dir=data_dir, use_mask=use_mask, num_workers=num_workers, image_size=image_size, crop=crop, normalize=normalize)
assert os.path.isdir(data_dir), "Data directory does not exist: %s" %data_dir
print(f"Loading testing data from {data_test}")
for name, data in data_test.items():
test_loader.update({name: get_loader(data_split={name: data}, is_validation=True, batch_size=1)})
return test_loader
def compute_stats_per_image(mask, ndsm, res):
connected_components = cv2.connectedComponentsWithStats(mask, 4, cv2.CV_8U)
num_instances = connected_components[0]
instances = connected_components[1]
for i in range(num_instances):
m = (instances == i+1).astype(np.uint8)
if m.max()<1:
continue
h_gt = np.median(ndsm[m>0])
size = m.sum()
if size in res:
res[size].append(h_gt)
else:
res.update({size: [h_gt]})
return res
def online_get_image_stats(data_dir, image_paths):
sum_x = np.zeros(3)
sum_x2 = np.zeros(3)
sum_size = 0
for image_path in image_paths:
img = io.imread(image_path).astype(int).reshape(-1,3)
sum_x += img.sum(axis=0)
sum_x2 += (img*img).sum(axis=0)
sum_size += img[:, 0].size
mean = sum_x / sum_size
std = list(np.sqrt(sum_x2/sum_size - mean*mean))
mean = list(mean)
torch.save([mean, std], os.path.join(data_dir, 'image_stats.pickle'))
return mean, std
def online_get_ndsm_stats(data_dir, image_paths):
sum_x = 0
sum_x2 = 0
sum_size = 0
minh = 0
maxh = 0
for image_path in image_paths:
img = io.imread(image_path).flatten()
img = np.nan_to_num(img)
sum_x += img.sum()
sum_x2 += (img**2).sum()
sum_size += img.size
minh = img.min() if img.min()<minh else minh
maxh = img.max() if img.max()>maxh else maxh
minh = 0 if minh < 0 else minh
mean = sum_x / sum_size
std = np.sqrt(sum_x2/sum_size - mean**2)
count = np.zeros(int(np.ceil(maxh)))
for image_path in image_paths:
img = io.imread(image_path).flatten()
img = np.nan_to_num(img)
img = np.clip(np.floor(img), a_min=0, a_max=None).astype(np.int)
res = np.bincount(img)
count[:res.size] += res
torch.save([mean, std, minh, maxh, count], os.path.join(data_dir, 'ndsm_stats.pickle'))
return mean, std, minh, maxh, count
class GBHDataset(torch.utils.data.Dataset):
def __init__(self, data_dir, data_mode, image_size=256, use_mask=False, crop=None, hflip=False, \
normalize=True, is_validation=False, overfit=False):
super(GBHDataset, self).__init__()
self.root = data_dir
self.paths = make_gbh_dataset(data_dir, data_mode, overfit)
self.size = len(self.paths)
self.image_size = image_size
self.crop = crop
self.hflip = hflip
self.normalize = normalize
self.use_mask = use_mask
self.is_validation = is_validation
image_stats_file = os.path.join(data_dir, 'image_stats.pickle')
self.mean, self.std = torch.load(image_stats_file)
def transform_ndsm(self, img):
img_shifted = np.log(img - self.ndsm_min + 1)
cls = np.floor(img_shifted / self.slope * self.num_classes)
return cls
def get_normal(self, ndsm):
zx = cv2.Sobel(ndsm, cv2.CV_32F, 1, 0, ksize=5)
zy = cv2.Sobel(ndsm, cv2.CV_32F, 0, 1, ksize=5)
norm_tile = np.stack((-zx, -zy, np.ones_like(ndsm)))
n = np.linalg.norm(norm_tile, axis=0)
norm_tile /= n
return norm_tile
def transform(self, img):
if self.crop:
if isinstance(self.crop, int):
img = tfs.CenterCrop(self.crop)(img)
else:
assert len(self.crop) == 4, 'Crop size must be an integer for center crop, or a list of 4 integers (y0,x0,h,w)'
img = tfs.functional.crop(img, *self.crop)
if not self.rcnn:
img = tfs.functional.resize(img, (self.image_size, self.image_size))
if self.hflip:
img = tfs.functional.hflip(img)
if self.normalize:
img = tfs.functional.normalize(img, self.mean, self.std)
return img
def __getitem__(self, index):
image_path, ndsm_path, mask_path = self.paths[index]
file_idx = os.path.basename(image_path).split('_IMG')[0]
image = torch.tensor(io.imread(image_path).astype(np.float32).transpose(2, 0, 1))
ndsm = np.nan_to_num(np.float32(io.imread(ndsm_path))).clip(0)
gt_dict = {"ndsm": torch.tensor(ndsm)[None, :, :]}
if self.use_mask:
mask = np.float32(io.imread(mask_path)>0)
gt_dict.update({"mask": torch.tensor(mask)[None, :, :]})
return file_idx, self.transform(image), gt_dict
def __len__(self):
return self.size
def name(self):
return 'GBHDataset'
## multiple image dataset with every subfolder in the data_dir ##
def make_gbh_dataset(dir, mode, overfit=False):
folder_name = ['image', 'ndsm', 'mask']
file_suffix = ['_IMG.tif', '_AGL.tif', '_BLG.tif']
images = []
with open(mode, 'r') as f:
lines = f.readlines()
if overfit:
lines = lines[:2]
for line in lines:
image = []
line = line.rstrip()
for folder, suffix in zip(folder_name, file_suffix):
image.append(os.path.join(dir, folder, line+suffix))
images.append(image)
return images
def get_image_list(dir, mode, ndsm=False, vis=False):
images = []
for m in mode:
with open(m, 'r') as f:
lines = f.readlines()
for line in lines:
line = line.rstrip()
if ndsm:
images.append(os.path.join(dir, 'ndsm', line+'_AGL.tif'))
elif vis:
images.append(os.path.join(dir, 'visualization', line+'_VIS.tif'))
else:
images.append(os.path.join(dir, 'image', line+'_IMG.tif'))
return images
def get_tri_image_loader(data_dir, data_split, is_validation=False, batch_size=8, num_workers=4, \
image_size=256, crop=None, normalize=True, use_mask=False, use_normal=False, overfit=False, \
rcnn=False, multitask=False, collate_fn=None, ordinal=False, num_classes=1000, sup_mode=None,
tasks=[True, False], use_vis=False, extended=False, noise=False, prototype=False, drop_last=False):
datasets = []
dataset_class = GBHDataset
if sup_mode is None:
sup_mode = [False] * len(data_split)
for ds, sm, task in zip(data_split.values(), sup_mode, tasks):
datasets.append(dataset_class(data_dir, ds, image_size=image_size, crop=crop, normalize=normalize, \
use_mask=use_mask, use_normal=use_normal, is_validation=is_validation, overfit=overfit, rcnn=rcnn, \
multitask=multitask, ordinal=ordinal, num_classes=num_classes, sup_mode=sm, is_h=task, use_vis=use_vis, \
noise=noise, prototype=prototype))
dataset = torch.utils.data.ConcatDataset(datasets)
loader = torch.utils.data.DataLoader(
dataset,
batch_size=batch_size,
shuffle=not is_validation,
num_workers=num_workers,
pin_memory=True,
collate_fn=collate_fn,
drop_last=drop_last
)
return loader