Upload senbench_clouds2_wrapper.py
Browse files
cloud_s2/senbench_clouds2_wrapper.py
ADDED
@@ -0,0 +1,155 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import kornia as K
|
2 |
+
import torch
|
3 |
+
from torchgeo.datasets import CloudCoverDetection
|
4 |
+
from typing import ClassVar
|
5 |
+
from collections.abc import Callable, Sequence
|
6 |
+
from torch import Tensor
|
7 |
+
from datetime import date
|
8 |
+
import os
|
9 |
+
import pandas as pd
|
10 |
+
import numpy as np
|
11 |
+
import rasterio
|
12 |
+
from pyproj import Transformer
|
13 |
+
from typing import TypeAlias
|
14 |
+
Path: TypeAlias = str | os.PathLike[str]
|
15 |
+
|
16 |
+
class SenBenchCloudS2(CloudCoverDetection):
|
17 |
+
url = None
|
18 |
+
all_bands = ('B02', 'B03', 'B04', 'B08')
|
19 |
+
splits: ClassVar[dict[str, str]] = {'train': 'public', 'val': 'private', 'test': 'private'}
|
20 |
+
|
21 |
+
def __init__(
|
22 |
+
self,
|
23 |
+
root: Path = 'data',
|
24 |
+
split: str = 'train',
|
25 |
+
bands: Sequence[str] = all_bands,
|
26 |
+
transforms: Callable[[dict[str, Tensor]], dict[str, Tensor]] | None = None,
|
27 |
+
download: bool = False,
|
28 |
+
) -> None:
|
29 |
+
|
30 |
+
#super().__init__(root=root, split=split, bands=bands, transforms=transforms, download=download)
|
31 |
+
assert split in self.splits
|
32 |
+
assert set(bands) <= set(self.all_bands)
|
33 |
+
|
34 |
+
self.root = root
|
35 |
+
self.split = split
|
36 |
+
self.bands = bands
|
37 |
+
self.transforms = transforms
|
38 |
+
self.download = download
|
39 |
+
|
40 |
+
self.csv = os.path.join(self.root, self.split, f'{self.split}_metadata.csv')
|
41 |
+
self._verify()
|
42 |
+
|
43 |
+
self.metadata = pd.read_csv(self.csv)
|
44 |
+
|
45 |
+
self.reference_date = date(1970, 1, 1)
|
46 |
+
self.patch_area = (16*10)**2 # patchsize 16 pix, gsd 10m
|
47 |
+
|
48 |
+
def __getitem__(self, index: int) -> dict[str, Tensor]:
|
49 |
+
"""Returns a sample from dataset.
|
50 |
+
|
51 |
+
Args:
|
52 |
+
index: index to return
|
53 |
+
|
54 |
+
Returns:
|
55 |
+
data, metadata (lon,lat,days,area) and label at given index
|
56 |
+
"""
|
57 |
+
chip_id = self.metadata.iat[index, 0]
|
58 |
+
date_str = self.metadata.iat[index, 2]
|
59 |
+
date_obj = date(int(date_str[:4]), int(date_str[5:7]), int(date_str[8:10]))
|
60 |
+
delta = (date_obj - self.reference_date).days
|
61 |
+
|
62 |
+
image, coord = self._load_image(chip_id)
|
63 |
+
label = self._load_target(chip_id)
|
64 |
+
|
65 |
+
meta_info = np.array([coord[0], coord[1], delta, self.patch_area]).astype(np.float32)
|
66 |
+
|
67 |
+
sample = {'image': image, 'mask': label, 'meta': torch.from_numpy(meta_info)}
|
68 |
+
|
69 |
+
if self.transforms is not None:
|
70 |
+
sample = self.transforms(sample)
|
71 |
+
|
72 |
+
# # add metadata
|
73 |
+
# sample['meta'] = torch.from_numpy(meta_info)
|
74 |
+
|
75 |
+
return sample
|
76 |
+
|
77 |
+
def _load_image(self, chip_id: str) -> Tensor:
|
78 |
+
"""Load all source images for a chip.
|
79 |
+
|
80 |
+
Args:
|
81 |
+
chip_id: ID of the chip.
|
82 |
+
|
83 |
+
Returns:
|
84 |
+
a tensor of stacked source image data, coord (lon,lat)
|
85 |
+
"""
|
86 |
+
path = os.path.join(self.root, self.split, f'{self.split}_features', chip_id)
|
87 |
+
images = []
|
88 |
+
coords = None
|
89 |
+
for band in self.bands:
|
90 |
+
with rasterio.open(os.path.join(path, f'{band}.tif')) as src:
|
91 |
+
images.append(src.read(1).astype(np.float32))
|
92 |
+
if coords is None:
|
93 |
+
cx,cy = src.xy(src.height // 2, src.width // 2)
|
94 |
+
if src.crs.to_string() != 'EPSG:4326':
|
95 |
+
crs_transformer = Transformer.from_crs(src.crs, 'epsg:4326', always_xy=True)
|
96 |
+
lon, lat = crs_transformer.transform(cx,cy)
|
97 |
+
else:
|
98 |
+
lon, lat = cx, cy
|
99 |
+
|
100 |
+
return torch.from_numpy(np.stack(images, axis=0)), (lon,lat)
|
101 |
+
|
102 |
+
|
103 |
+
|
104 |
+
class SegDataAugmentation(torch.nn.Module):
|
105 |
+
def __init__(self, split, size):
|
106 |
+
super().__init__()
|
107 |
+
|
108 |
+
mean = torch.Tensor([0.0])
|
109 |
+
std = torch.Tensor([1.0])
|
110 |
+
|
111 |
+
self.norm = K.augmentation.Normalize(mean=mean, std=std)
|
112 |
+
|
113 |
+
if split == "train":
|
114 |
+
self.transform = K.augmentation.AugmentationSequential(
|
115 |
+
K.augmentation.Resize(size=size, align_corners=True),
|
116 |
+
K.augmentation.RandomRotation(degrees=90, p=0.5, align_corners=True),
|
117 |
+
K.augmentation.RandomHorizontalFlip(p=0.5),
|
118 |
+
K.augmentation.RandomVerticalFlip(p=0.5),
|
119 |
+
data_keys=["input", "mask"],
|
120 |
+
)
|
121 |
+
else:
|
122 |
+
self.transform = K.augmentation.AugmentationSequential(
|
123 |
+
K.augmentation.Resize(size=size, align_corners=True),
|
124 |
+
data_keys=["input", "mask"],
|
125 |
+
)
|
126 |
+
|
127 |
+
@torch.no_grad()
|
128 |
+
def forward(self, batch: dict[str,]):
|
129 |
+
"""Torchgeo returns a dictionary with 'image' and 'label' keys, but engine expects a tuple"""
|
130 |
+
x,mask = batch["image"], batch["mask"]
|
131 |
+
x = self.norm(x)
|
132 |
+
x_out, mask_out = self.transform(x, mask)
|
133 |
+
return x_out.squeeze(0), mask_out.squeeze(0).squeeze(0), batch["meta"]
|
134 |
+
|
135 |
+
class SenBenchCloudS2Dataset:
|
136 |
+
def __init__(self, config):
|
137 |
+
self.dataset_config = config
|
138 |
+
self.img_size = (config.image_resolution, config.image_resolution)
|
139 |
+
self.root_dir = config.data_path
|
140 |
+
|
141 |
+
def create_dataset(self):
|
142 |
+
train_transform = SegDataAugmentation(split="train", size=self.img_size)
|
143 |
+
eval_transform = SegDataAugmentation(split="test", size=self.img_size)
|
144 |
+
|
145 |
+
dataset_train = SenBenchCloudS2(
|
146 |
+
root=self.root_dir, split="train", transforms=train_transform
|
147 |
+
)
|
148 |
+
dataset_val = SenBenchCloudS2(
|
149 |
+
root=self.root_dir, split="val", transforms=eval_transform
|
150 |
+
)
|
151 |
+
dataset_test = SenBenchCloudS2(
|
152 |
+
root=self.root_dir, split="test", transforms=eval_transform
|
153 |
+
)
|
154 |
+
|
155 |
+
return dataset_train, dataset_val, dataset_test
|