-
Notifications
You must be signed in to change notification settings - Fork 11
/
preprocess.py
201 lines (185 loc) · 7.81 KB
/
preprocess.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
import numpy as np
import argparse
# import torch as th
# from os import listdir
import os
import h5py
import json
from PIL import Image
from utils import args
from time import ctime
# from torchvision import transforms
Image.MAX_IMAGE_PIXELS = 1000000000
def get_args():
parser = argparse.ArgumentParser(description="Data Preparation")
parser.add_argument("--data_dir", action="append", type=str)
parser.add_argument("--name", type=str, default="landslide.h5")
parser.add_argument("--save_to", type=str, default="../image_data/")
parser.add_argument("--feature_num", type=int, default=94)
parser.add_argument('--shape', action='append', type=args.shape)
# parser.add_argument("--feature_names", type=str, default="litho, landcover, slope")
# parser.add_argument("--ground_truth_name", type=str, default="polygon_shallow_soil_slide.tif")
parser.add_argument("--data_format", type=str, default=".tif")
parser.add_argument("--pad", type=int, default=64)
# parser.add_argument("--label_pos", nargs='+', type=args.pos)
# parser.add_argument("--img_size", type=(int,int), default=(20340, 26591))
return parser.parse_args()
def convert_nodata(np_img):
'''
convert a binary image with no data pts (value=127) to a binary image with no data values being the mean.
'''
data = np_img.all()==1 + np_img.all()==0
nodata = 1-data
mean = np.mean(np_img[data])
np_img[nodata] = mean
# import ipdb; ipdb.set_trace()
return np_img
def normalize(np_img, f = 'slope'):
if f == 'slope':
np_img[np_img < 0] = 0
np_img[np_img > 180] = 0
mean = np.mean(np_img)
std = np.std(np_img)
print('mean, std before normalizing: %f, %f' %(mean, std))
np_img = (np_img - mean)/std
print('after: %f, %f' %(np.mean(np_img), np.std(np_img)))
return np_img
def zero_one(np_img):
# ones = np_img==1
ones = np_img!=0
np_img[ones]=1
return np_img
def initialize(f, key):
(n, h, w) = f[key]['train/data'].shape
(_, hg, wg) = f[key]['test/data'].shape
zero_train = np.zeros((h, w))
zero_test = np.zeros((hg, wg))
for i in range(n):
f[key]['train/data'][i] = zero_train
f[key]['test/data'][i] = zero_test
print('%s -> %d/%d' %(ctime(), i+1, n), end='\r')
return f
def process_data():
args = get_args()
g = open('data_dict.json', 'r')
data_dict = json.load(g)
g.close()
f = h5py.File(args.save_to+args.name, 'a')
for data_path in args.data_dir:
name = data_path.split('/')[-2]
for n, h, w in args.shape:
hv = h//5
if n == name and not name in f.keys():
f.create_dataset(
name+'/test/data',
(args.feature_num, hv+args.pad*2, w+args.pad*2),
dtype='f',
compression='lzf'
)
f.create_dataset(name+'/test/gt', (1, hv, w), dtype='f', compression='lzf')
f.create_dataset(
name+'/train/data',
(args.feature_num, h-hv+args.pad*2, w+args.pad*2),
dtype='f',
compression='lzf'
)
f.create_dataset(name+'/train/gt', (1, h-hv, w), dtype='f', compression='lzf')
print('created data and gt in %s' %name)
break
f = initialize(f, name)
print(list(f.keys()))
for data_path in args.data_dir:
name = data_path.split('/')[-2]
images = os.listdir(data_path)
for img in images:
if args.data_format in img and not '.xml' in img and not 'gt' in img:
t = np.array(Image.open(data_path+img))
n_ = img.split('.')[0]
if int(data_dict[n_]) == 0:
print('normalizing slope')
t = normalize(t, 'slope')
elif int(data_dict[n_]) == args.feature_num-1:
t = normalize(t, 'DEM')
# else:
# # t = convert_nodata(zero_one(t))
# t = zero_one(t)
print(data_dict[n_], type(data_dict[n_]))
hlen = t.shape[0]//5
f[name+'/train/data'][int(data_dict[n_])] = np.pad(np.concatenate((t[0:hlen, :], t[2*hlen:, :]), 0), args.pad, 'constant')
f[name+'/test/data'][int(data_dict[n_])] = np.pad(t[hlen:2*hlen, :], args.pad, 'constant')
gt = np.array(Image.open(data_path+'gt'+args.data_format))
hlen = gt.shape[0]//5
f[name+'/train/gt'][0] = np.concatenate((gt[0:hlen, :], gt[2*hlen:, :]), 0)
f[name+'/test/gt'][0] = gt[hlen:2*hlen, :]
f.close()
process_data()
# def oversample(args, directory_path, data, pad):
# print('%s --- oversampling pos images ...' % ctime())
# (h, w) = data.shape
# h, w = h-pad*2, w-pad*2
# lpos = args.label_pos
# cnt = 0
# for e in lpos:
# (l0, l1, u0, u1) = e
# l0, l1, u0, u1 = l0//10, l1//10, (u0//10)+1, (u1//10)+1
# for row in range(l0, u0):
# for col in range(l1, u1):
# if (row+20)*10+pad*2 > data.shape[0] or (col+20)*10+pad*2 > data.shape[1]:
# print('ignoring this batch')
# continue
# np.save(
# directory_path+str(row)+'_'+str(col)+'_10'+'.npy',
# data[row*10:(row+20)*10+pad*2, col*10:(col+20)*10+pad*2]
# )
# cnt += 1
# #if (row+10)*20+pad*2 > data.shape[0] or (col+10)*20+pad*2 > data.shape[2]:
# # raise ValueError
# print('%s --- oversampling done: %d.' %(ctime(), cnt))
# def write(directory_path, data, feature_num, pad):
# print('%s --- writing images for feature %s.' %(ctime(), str(feature_num)))
# if not os.path.exists(directory_path+str(feature_num)):
# os.mkdir(directory_path+str(feature_num))
# dir_name = directory_path+str(feature_num)+'/'
# (h, w) = data.shape
# h, w = h-pad*2, w-pad*2
# for i in range(h//100-1):
# for j in range(w//100-1):
# np.save(
# dir_name+str(i)+'_'+str(j)+'_100'+'.npy',
# data[i*100:(i+2)*100+pad*2, j*100:(j+2)*100+pad*2]
# )
# print('%s --- wrote images with stride 100: %d.' %(ctime(), (h//100-1)*(w//100-1)))
# def preprocess():
# args = get_args()
# fn = args.feature_names.split(", ")
# gtn = args.ground_truth_name
# path = args.img_dir_path
# files = os.listdir(path)
# data_directory = args.save_to + 'data_' + str(args.pad*2)+'/'
# if not os.path.exists(data_directory):
# os.mkdir(data_directory)
# cnt = 0
# for feature_name in fn:
# for img_path in files:
# if feature_name in img_path:
# im = Image.open(path+img_path)
# if feature_name=="litho" or feature_name=="landcover":
# im = np.asarray(im, dtype=np.float16)
# im = convert_nodata(im)
# else: # it's either slope or DEM, they don't have a no-data pt and should be normalized
# im = np.array(im)
# im = normalize(im)
# if args.pad != 0:
# im = np.pad(im, args.pad, 'constant') # pads with zeros
# write(data_directory, im, cnt, args.pad)
# if args.label_pos:
# oversample(args, data_directory+str(cnt)+'/', im, args.pad)
# cnt += 1
# print("%d feature(s) are loaded ..." % cnt, end="\r")
# print(">> all %s features have been loaded." % feature_name)
# gt = np.array(Image.open(path+gtn))
# gt = zero_one(gt)
# write(data_directory, gt, 'gt', 0)
# oversample(args, data_directory+'gt/', gt, 0)
# print("all images are saved in %s." % args.save_to)
# preprocess()