wangyi111 commited on
Commit
0aaf21e
·
verified ·
1 Parent(s): 65f48be

Upload senbench_so2sat_wrapper.py

Browse files
so2sat_s1s2/senbench_so2sat_wrapper.py ADDED
@@ -0,0 +1,268 @@
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
+ import kornia.augmentation as K
2
+ import torch
3
+ from torchgeo.datasets import So2Sat
4
+ import os
5
+ from collections.abc import Callable, Sequence
6
+ from torch import Tensor
7
+ import numpy as np
8
+ import rasterio
9
+ from pyproj import Transformer
10
+ import h5py
11
+ from typing import TypeAlias, ClassVar
12
+ import pathlib
13
+ Path: TypeAlias = str | os.PathLike[str]
14
+
15
+ class SenBenchSo2Sat(So2Sat):
16
+
17
+ versions = ('3_culture_10')
18
+ filenames_by_version: ClassVar[dict[str, dict[str, str]]] = {
19
+ # '2': {
20
+ # 'train': 'training.h5',
21
+ # 'validation': 'validation.h5',
22
+ # 'test': 'testing.h5',
23
+ # },
24
+ # '3_random': {'train': 'random/training.h5', 'test': 'random/testing.h5'},
25
+ # '3_block': {'train': 'block/training.h5', 'test': 'block/testing.h5'},
26
+ '3_culture_10': {
27
+ 'train': 'culture_10/train-new.h5',
28
+ 'val': 'culture_10/val-new.h5',
29
+ 'test': 'culture_10/test-new.h5',
30
+ },
31
+ }
32
+
33
+ classes = (
34
+ 'Compact high rise',
35
+ 'Compact mid rise',
36
+ 'Compact low rise',
37
+ 'Open high rise',
38
+ 'Open mid rise',
39
+ 'Open low rise',
40
+ 'Lightweight low rise',
41
+ 'Large low rise',
42
+ 'Sparsely built',
43
+ 'Heavy industry',
44
+ 'Dense trees',
45
+ 'Scattered trees',
46
+ 'Bush, scrub',
47
+ 'Low plants',
48
+ 'Bare rock or paved',
49
+ 'Bare soil or sand',
50
+ 'Water',
51
+ )
52
+
53
+ all_s1_band_names = (
54
+ 'S1_B1', # VH real
55
+ 'S1_B2', # VH imaginary
56
+ 'S1_B3', # VV real
57
+ 'S1_B4', # VV imaginary
58
+ 'S1_B5', # VH intensity
59
+ 'S1_B6', # VV intensity
60
+ 'S1_B7', # PolSAR covariance matrix off-diagonal real
61
+ 'S1_B8', # PolSAR covariance matrix off-diagonal imaginary
62
+ )
63
+ all_s2_band_names = (
64
+ 'S2_B02',
65
+ 'S2_B03',
66
+ 'S2_B04',
67
+ 'S2_B05',
68
+ 'S2_B06',
69
+ 'S2_B07',
70
+ 'S2_B08',
71
+ 'S2_B8A',
72
+ 'S2_B11',
73
+ 'S2_B12',
74
+ )
75
+ all_band_names = all_s1_band_names + all_s2_band_names
76
+
77
+ rgb_bands = ('S2_B04', 'S2_B03', 'S2_B02')
78
+
79
+ BAND_SETS: ClassVar[dict[str, tuple[str, ...]]] = {
80
+ 'all': all_band_names,
81
+ 's1': all_s1_band_names,
82
+ 's2': all_s2_band_names,
83
+ 'rgb': rgb_bands,
84
+ }
85
+
86
+ def __init__(
87
+ self,
88
+ root: Path = 'data',
89
+ version: str = '3_culture_10', # only supported version now
90
+ split: str = 'train',
91
+ bands: Sequence[str] = BAND_SETS['s2'], # only supported bands now
92
+ transforms: Callable[[dict[str, Tensor]], dict[str, Tensor]] | None = None,
93
+ download: bool = False,
94
+ ) -> None:
95
+
96
+ #h5py = lazy_import('h5py')
97
+
98
+ assert version in self.versions
99
+ assert split in self.filenames_by_version[version]
100
+
101
+ self._validate_bands(bands)
102
+ self.s1_band_indices: np.typing.NDArray[np.int_] = np.array(
103
+ [
104
+ self.all_s1_band_names.index(b)
105
+ for b in bands
106
+ if b in self.all_s1_band_names
107
+ ]
108
+ ).astype(int)
109
+
110
+ self.s1_band_names = [self.all_s1_band_names[i] for i in self.s1_band_indices]
111
+
112
+ self.s2_band_indices: np.typing.NDArray[np.int_] = np.array(
113
+ [
114
+ self.all_s2_band_names.index(b)
115
+ for b in bands
116
+ if b in self.all_s2_band_names
117
+ ]
118
+ ).astype(int)
119
+
120
+ self.s2_band_names = [self.all_s2_band_names[i] for i in self.s2_band_indices]
121
+
122
+ self.bands = bands
123
+
124
+ self.root = root
125
+ self.version = version
126
+ self.split = split
127
+ self.transforms = transforms
128
+ # self.checksum = checksum
129
+
130
+ self.fn = os.path.join(self.root, self.filenames_by_version[version][split])
131
+
132
+ # if not self._check_integrity():
133
+ # raise DatasetNotFoundError(self)
134
+
135
+ with h5py.File(self.fn, 'r') as f:
136
+ self.size: int = f['label'].shape[0]
137
+
138
+ self.patch_area = (16*10/1000)**2 # patchsize 16 pix, gsd 10m
139
+
140
+
141
+ def __getitem__(self, index: int) -> dict[str, Tensor]:
142
+ """Return an index within the dataset.
143
+
144
+ Args:
145
+ index: index to return
146
+
147
+ Returns:
148
+ data and label at that index
149
+ """
150
+ #h5py = lazy_import('h5py')
151
+ with h5py.File(self.fn, 'r') as f:
152
+ #s1 = f['sen1'][index].astype(np.float32)
153
+ #s1 = np.take(s1, indices=self.s1_band_indices, axis=2)
154
+ s2 = f['sen2'][index].astype(np.float32)
155
+ s2 = np.take(s2, indices=self.s2_band_indices, axis=2)
156
+
157
+ # convert one-hot encoding to int64 then torch int
158
+ label = torch.tensor(f['label'][index].argmax())
159
+
160
+ #s1 = np.rollaxis(s1, 2, 0) # convert to CxHxW format
161
+ s2 = np.rollaxis(s2, 2, 0) # convert to CxHxW format
162
+
163
+ #s1 = torch.from_numpy(s1)
164
+ s2 = torch.from_numpy(s2)
165
+
166
+ meta_info = np.array([np.nan, np.nan, np.nan, self.patch_area]).astype(np.float32)
167
+
168
+ sample = {'image': s2, 'label': label, 'meta': torch.from_numpy(meta_info)}
169
+
170
+ if self.transforms is not None:
171
+ sample = self.transforms(sample)
172
+
173
+ return sample
174
+
175
+
176
+ class ClsDataAugmentation(torch.nn.Module):
177
+ BAND_STATS = {
178
+ 'mean': {
179
+ 'B01': 1353.72696296,
180
+ 'B02': 1117.20222222,
181
+ 'B03': 1041.8842963,
182
+ 'B04': 946.554,
183
+ 'B05': 1199.18896296,
184
+ 'B06': 2003.00696296,
185
+ 'B07': 2374.00874074,
186
+ 'B08': 2301.22014815,
187
+ 'B8A': 2599.78311111,
188
+ 'B09': 732.18207407,
189
+ 'B10': 12.09952894,
190
+ 'B11': 1820.69659259,
191
+ 'B12': 1118.20259259,
192
+ #'VV': -12.54847273,
193
+ #'VH': -20.19237134
194
+ },
195
+ 'std': {
196
+ 'B01': 897.27143653,
197
+ 'B02': 736.01759721,
198
+ 'B03': 684.77615743,
199
+ 'B04': 620.02902871,
200
+ 'B05': 791.86263829,
201
+ 'B06': 1341.28018273,
202
+ 'B07': 1595.39989386,
203
+ 'B08': 1545.52915718,
204
+ 'B8A': 1750.12066835,
205
+ 'B09': 475.11595216,
206
+ 'B10': 98.26600935,
207
+ 'B11': 1216.48651476,
208
+ 'B12': 736.6981037,
209
+ #'VV': 5.25697717,
210
+ #'VH': 5.91150917
211
+ }
212
+ }
213
+
214
+ def __init__(self, split, size, bands):
215
+ super().__init__()
216
+
217
+ mean = []
218
+ std = []
219
+ for band in bands:
220
+ band = band[3:]
221
+ mean.append(self.BAND_STATS['mean'][band])
222
+ std.append(self.BAND_STATS['std'][band])
223
+ mean = torch.Tensor(mean)
224
+ std = torch.Tensor(std)
225
+
226
+ if split == "train":
227
+ self.transform = torch.nn.Sequential(
228
+ K.Normalize(mean=mean, std=std),
229
+ K.Resize(size=size, align_corners=True),
230
+ K.RandomHorizontalFlip(p=0.5),
231
+ K.RandomVerticalFlip(p=0.5),
232
+ )
233
+ else:
234
+ self.transform = torch.nn.Sequential(
235
+ K.Normalize(mean=mean, std=std),
236
+ K.Resize(size=size, align_corners=True),
237
+ )
238
+
239
+ @torch.no_grad()
240
+ def forward(self, batch: dict[str,]):
241
+ """Torchgeo returns a dictionary with 'image' and 'label' keys, but engine expects a tuple"""
242
+ x_out = self.transform(batch["image"]).squeeze(0)
243
+ return x_out, batch["label"], batch["meta"]
244
+
245
+
246
+ class SenBenchSo2SatDataset:
247
+ def __init__(self, config):
248
+ self.dataset_config = config
249
+ self.img_size = (config.image_resolution, config.image_resolution)
250
+ self.root_dir = config.data_path
251
+ self.bands = config.band_names
252
+ self.version = config.version
253
+
254
+ def create_dataset(self):
255
+ train_transform = ClsDataAugmentation(split="train", size=self.img_size, bands=self.bands)
256
+ eval_transform = ClsDataAugmentation(split="test", size=self.img_size, bands=self.bands)
257
+
258
+ dataset_train = SenBenchSo2Sat(
259
+ root=self.root_dir, version=self.version, split="train", bands=self.bands, transforms=train_transform
260
+ )
261
+ dataset_val = SenBenchSo2Sat(
262
+ root=self.root_dir, version=self.version, split="val", bands=self.bands, transforms=eval_transform
263
+ )
264
+ dataset_test = SenBenchSo2Sat(
265
+ root=self.root_dir, version=self.version, split="test", bands=self.bands, transforms=eval_transform
266
+ )
267
+
268
+ return dataset_train, dataset_val, dataset_test