Amir Erfan Eshratifar
commited on
Commit
·
241b6a2
1
Parent(s):
551ee08
model checkpoints, sample input, readme
Browse files- README.md +35 -6
- jp2s/B02.jp2 +0 -0
- jp2s/B03.jp2 +0 -0
- jp2s/B04.jp2 +0 -0
- jp2s/B8A.jp2 +0 -0
- omnicloudmask/__init__.py +16 -0
- omnicloudmask/__version__.py +1 -0
- omnicloudmask/cloud_mask.py +493 -0
- omnicloudmask/data_loaders.py +159 -0
- omnicloudmask/download_models.py +92 -0
- omnicloudmask/model_utils.py +208 -0
- omnicloudmask/models/PM_model_2.2.10_RG_NIR_509_convnextv2_nano.fcmae_ft_in1k_PT_state.pth +3 -0
- omnicloudmask/models/PM_model_2.2.10_RG_NIR_509_regnety_004.pycls_in1k_PT_state.pth +3 -0
- omnicloudmask/models/model_download_links.csv +3 -0
- omnicloudmask/raster_utils.py +118 -0
README.md
CHANGED
@@ -1,6 +1,35 @@
|
|
1 |
-
---
|
2 |
-
license: mit
|
3 |
-
|
4 |
-
|
5 |
-
|
6 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
---
|
2 |
+
license: mit
|
3 |
+
datasets:
|
4 |
+
- csaybar/CloudSEN12-high
|
5 |
+
language:
|
6 |
+
- en
|
7 |
+
base_model:
|
8 |
+
- NickWright/OmniCloudMask
|
9 |
+
tags:
|
10 |
+
- remote-sensing
|
11 |
+
- cloud-detection
|
12 |
+
---
|
13 |
+
|
14 |
+
|
15 |
+
# Cloud Detection Model
|
16 |
+
|
17 |
+
This model is based on NickWright/OmniCloudMask for cloud detection in satellite imagery. It provides pixel-level segmentation with the following classes:
|
18 |
+
|
19 |
+
0 = Clear
|
20 |
+
1 = Thick Cloud
|
21 |
+
2 = Thin Cloud
|
22 |
+
3 = Cloud Shadow
|
23 |
+
|
24 |
+
## Usage
|
25 |
+
|
26 |
+
The model requires Python 3.10 or higher. To use this model:
|
27 |
+
```bash
|
28 |
+
pip install -r requirements.txt
|
29 |
+
```
|
30 |
+
```bash
|
31 |
+
python3 model.py
|
32 |
+
```
|
33 |
+
|
34 |
+
Below is a visualization of the cloud mask generated by the model:
|
35 |
+

|
jp2s/B02.jp2
ADDED
|
jp2s/B03.jp2
ADDED
|
jp2s/B04.jp2
ADDED
|
jp2s/B8A.jp2
ADDED
|
omnicloudmask/__init__.py
ADDED
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from .__version__ import __version__
|
2 |
+
from .cloud_mask import predict_from_array, predict_from_load_func
|
3 |
+
from .data_loaders import (
|
4 |
+
load_ls8,
|
5 |
+
load_multiband,
|
6 |
+
load_s2,
|
7 |
+
)
|
8 |
+
|
9 |
+
__all__ = [
|
10 |
+
"predict_from_load_func",
|
11 |
+
"predict_from_array",
|
12 |
+
"load_ls8",
|
13 |
+
"load_multiband",
|
14 |
+
"load_s2",
|
15 |
+
"__version__",
|
16 |
+
]
|
omnicloudmask/__version__.py
ADDED
@@ -0,0 +1 @@
|
|
|
|
|
1 |
+
__version__ = "1.0.9"
|
omnicloudmask/cloud_mask.py
ADDED
@@ -0,0 +1,493 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import warnings
|
2 |
+
from concurrent.futures import ThreadPoolExecutor, as_completed
|
3 |
+
from pathlib import Path
|
4 |
+
from threading import Thread
|
5 |
+
from typing import Callable, Generator, Optional, Union
|
6 |
+
|
7 |
+
import numpy as np
|
8 |
+
import torch
|
9 |
+
from rasterio.profiles import Profile
|
10 |
+
from tqdm.auto import tqdm
|
11 |
+
|
12 |
+
from .__version__ import __version__
|
13 |
+
from .download_models import get_models
|
14 |
+
from .model_utils import (
|
15 |
+
create_gradient_mask,
|
16 |
+
default_device,
|
17 |
+
get_torch_dtype,
|
18 |
+
inference_and_store,
|
19 |
+
load_model_from_weights,
|
20 |
+
)
|
21 |
+
from .raster_utils import (
|
22 |
+
get_patch,
|
23 |
+
make_patch_indexes,
|
24 |
+
mask_prediction,
|
25 |
+
save_prediction,
|
26 |
+
)
|
27 |
+
|
28 |
+
|
29 |
+
def compile_batches(
|
30 |
+
batch_size: int,
|
31 |
+
patch_size: int,
|
32 |
+
patch_indexes: list[tuple[int, int, int, int]],
|
33 |
+
input_array: np.ndarray,
|
34 |
+
no_data_value: int,
|
35 |
+
inference_device: torch.device,
|
36 |
+
inference_dtype: torch.dtype,
|
37 |
+
) -> Generator[tuple[torch.Tensor, list[tuple[int, int, int, int]]], None, None]:
|
38 |
+
"""Used to compile batches of patches from the input array and return them as a generator."""
|
39 |
+
|
40 |
+
with ThreadPoolExecutor(max_workers=batch_size) as executor:
|
41 |
+
futures = [
|
42 |
+
executor.submit(get_patch, input_array, index, no_data_value)
|
43 |
+
for index in patch_indexes
|
44 |
+
]
|
45 |
+
|
46 |
+
total_futures = len(futures)
|
47 |
+
all_indexes = set()
|
48 |
+
index_batch = []
|
49 |
+
patch_batch_array = np.zeros(
|
50 |
+
(batch_size, input_array.shape[0], patch_size, patch_size), dtype=np.float32
|
51 |
+
)
|
52 |
+
|
53 |
+
for index, future in enumerate(as_completed(futures)):
|
54 |
+
patch, new_index = future.result()
|
55 |
+
|
56 |
+
if patch is not None and new_index not in all_indexes:
|
57 |
+
index_batch.append(new_index)
|
58 |
+
patch_batch_array[len(index_batch) - 1] = patch
|
59 |
+
all_indexes.add(new_index)
|
60 |
+
|
61 |
+
if len(index_batch) == batch_size or index == total_futures - 1:
|
62 |
+
if len(index_batch) == 0:
|
63 |
+
continue
|
64 |
+
input_tensor = (
|
65 |
+
torch.tensor(patch_batch_array[: len(index_batch)])
|
66 |
+
.to(inference_device)
|
67 |
+
.to(inference_dtype)
|
68 |
+
)
|
69 |
+
yield input_tensor, index_batch
|
70 |
+
index_batch = []
|
71 |
+
|
72 |
+
|
73 |
+
def run_models_on_array(
|
74 |
+
models: list[torch.nn.Module],
|
75 |
+
input_array: np.ndarray,
|
76 |
+
pred_tracker: torch.Tensor,
|
77 |
+
grad_tracker: Union[torch.Tensor, None],
|
78 |
+
patch_size: int,
|
79 |
+
patch_overlap: int,
|
80 |
+
inference_device: torch.device,
|
81 |
+
batch_size: int = 2,
|
82 |
+
inference_dtype: torch.dtype = torch.float32,
|
83 |
+
no_data_value: int = 0,
|
84 |
+
) -> None:
|
85 |
+
"""Used to execute the model on the input array, in patches. Predictions are stored in pred_tracker and grad_tracker, updated in place."""
|
86 |
+
patch_indexes = make_patch_indexes(
|
87 |
+
array_height=input_array.shape[1],
|
88 |
+
array_width=input_array.shape[2],
|
89 |
+
patch_size=patch_size,
|
90 |
+
patch_overlap=patch_overlap,
|
91 |
+
)
|
92 |
+
|
93 |
+
gradient = create_gradient_mask(
|
94 |
+
patch_size, patch_overlap, device=inference_device, dtype=inference_dtype
|
95 |
+
)
|
96 |
+
|
97 |
+
input_tensor_gen = compile_batches(
|
98 |
+
batch_size=batch_size,
|
99 |
+
patch_size=patch_size,
|
100 |
+
patch_indexes=patch_indexes,
|
101 |
+
input_array=input_array,
|
102 |
+
no_data_value=no_data_value,
|
103 |
+
inference_device=inference_device,
|
104 |
+
inference_dtype=inference_dtype,
|
105 |
+
)
|
106 |
+
|
107 |
+
for patch_batch, index_batch in input_tensor_gen:
|
108 |
+
inference_and_store(
|
109 |
+
models=models,
|
110 |
+
patch_batch=patch_batch,
|
111 |
+
index_batch=index_batch,
|
112 |
+
pred_tracker=pred_tracker,
|
113 |
+
gradient=gradient,
|
114 |
+
grad_tracker=grad_tracker,
|
115 |
+
)
|
116 |
+
|
117 |
+
|
118 |
+
def check_patch_size(
|
119 |
+
input_array: np.ndarray, no_data_value: int, patch_size: int, patch_overlap: int
|
120 |
+
) -> tuple[int, int]:
|
121 |
+
"""Used to check the inputs and adjust the patch size and overlap if necessary."""
|
122 |
+
# check the shape of the input array
|
123 |
+
if len(input_array.shape) != 3:
|
124 |
+
raise ValueError(
|
125 |
+
f"Input array must have 3 dimensions, found {len(input_array.shape)}. The input should be in format (bands (red,green,NIR), height, width)."
|
126 |
+
)
|
127 |
+
|
128 |
+
# check the width and height are greater than 10 pixels
|
129 |
+
if min(input_array.shape[1], input_array.shape[2]) < 10:
|
130 |
+
raise ValueError(
|
131 |
+
f"Input array must have a width and height greater than 10 pixels, found shape {input_array.shape}. The input should be in format (bands (red,green,NIR), height, width)."
|
132 |
+
)
|
133 |
+
if min(input_array.shape[1], input_array.shape[2]) < 50:
|
134 |
+
warnings.warn(
|
135 |
+
f"Input width or height is less than 50 pixels, found shape {input_array.shape}. Such a small image may not provide adequate spatial context for the model."
|
136 |
+
)
|
137 |
+
|
138 |
+
# if the input has a lot of no data values and the patch size is larger than half the image size, we reduce the patch size and overlap
|
139 |
+
if np.count_nonzero(input_array == no_data_value) / input_array.size > 0.3:
|
140 |
+
if patch_size > min(input_array.shape[1], input_array.shape[2]) / 2:
|
141 |
+
patch_size = min(input_array.shape[1], input_array.shape[2]) // 2
|
142 |
+
if patch_size // 2 < patch_overlap:
|
143 |
+
patch_overlap = patch_size // 2
|
144 |
+
|
145 |
+
warnings.warn(
|
146 |
+
f"Significant no-data areas detected. Adjusting patch size to {patch_size}px and overlap to {patch_overlap}px to minimize no-data patches."
|
147 |
+
)
|
148 |
+
|
149 |
+
# if the patch size is larger than the image size, we reduce the patch size and overlap
|
150 |
+
if patch_size > min(input_array.shape[1], input_array.shape[2]):
|
151 |
+
patch_size = min(input_array.shape[1], input_array.shape[2])
|
152 |
+
if patch_size // 2 < patch_overlap:
|
153 |
+
patch_overlap = patch_size // 2
|
154 |
+
warnings.warn(
|
155 |
+
f"Patch size too large, reducing to {patch_size} and overlap to {patch_overlap}."
|
156 |
+
)
|
157 |
+
|
158 |
+
# if the patch overlap is larger than the patch size, raise an error
|
159 |
+
if patch_overlap >= patch_size:
|
160 |
+
raise ValueError(
|
161 |
+
f"Patch overlap {patch_overlap}px must be less than patch size {patch_size}px."
|
162 |
+
)
|
163 |
+
return patch_overlap, patch_size
|
164 |
+
|
165 |
+
|
166 |
+
def coordinator(
|
167 |
+
input_array: np.ndarray,
|
168 |
+
models: list[torch.nn.Module],
|
169 |
+
inference_dtype: torch.dtype,
|
170 |
+
export_confidence: bool,
|
171 |
+
softmax_output: bool,
|
172 |
+
inference_device: torch.device,
|
173 |
+
mosaic_device: torch.device,
|
174 |
+
patch_size: int,
|
175 |
+
patch_overlap: int,
|
176 |
+
batch_size: int,
|
177 |
+
profile: Profile = Profile(),
|
178 |
+
output_path: Path = Path(""),
|
179 |
+
no_data_value: int = 0,
|
180 |
+
pbar: Optional[tqdm] = None,
|
181 |
+
apply_no_data_mask: bool = False,
|
182 |
+
export_to_disk: bool = True,
|
183 |
+
save_executor: Optional[ThreadPoolExecutor] = None,
|
184 |
+
pred_classes: int = 4,
|
185 |
+
) -> np.ndarray:
|
186 |
+
"""Used to coordinate the process of predicting from an input array."""
|
187 |
+
|
188 |
+
patch_overlap, patch_size = check_patch_size(
|
189 |
+
input_array, no_data_value, patch_size, patch_overlap
|
190 |
+
)
|
191 |
+
|
192 |
+
pred_tracker = torch.zeros(
|
193 |
+
(pred_classes, *input_array.shape[1:3]),
|
194 |
+
dtype=inference_dtype,
|
195 |
+
device=mosaic_device,
|
196 |
+
)
|
197 |
+
|
198 |
+
grad_tracker = (
|
199 |
+
torch.zeros(input_array.shape[1:3], dtype=inference_dtype, device=mosaic_device)
|
200 |
+
if export_confidence
|
201 |
+
else None
|
202 |
+
)
|
203 |
+
|
204 |
+
run_models_on_array(
|
205 |
+
models=models,
|
206 |
+
input_array=input_array,
|
207 |
+
pred_tracker=pred_tracker,
|
208 |
+
grad_tracker=grad_tracker,
|
209 |
+
inference_device=inference_device,
|
210 |
+
inference_dtype=inference_dtype,
|
211 |
+
no_data_value=no_data_value,
|
212 |
+
patch_size=patch_size,
|
213 |
+
patch_overlap=patch_overlap,
|
214 |
+
batch_size=batch_size,
|
215 |
+
)
|
216 |
+
|
217 |
+
if export_confidence:
|
218 |
+
pred_tracker_norm = pred_tracker / grad_tracker
|
219 |
+
if softmax_output:
|
220 |
+
pred_tracker = torch.clip(
|
221 |
+
(torch.nn.functional.softmax(pred_tracker_norm, 0) + 0.001),
|
222 |
+
0.001,
|
223 |
+
0.999,
|
224 |
+
)
|
225 |
+
else:
|
226 |
+
pred_tracker = pred_tracker_norm
|
227 |
+
|
228 |
+
pred_tracker_np = pred_tracker.float().numpy(force=True)
|
229 |
+
|
230 |
+
else:
|
231 |
+
pred_tracker_np = (
|
232 |
+
torch.argmax(pred_tracker, 0, keepdim=True)
|
233 |
+
.numpy(force=True)
|
234 |
+
.astype(np.uint8)
|
235 |
+
)
|
236 |
+
|
237 |
+
if apply_no_data_mask:
|
238 |
+
pred_tracker_np = mask_prediction(input_array, pred_tracker_np, no_data_value)
|
239 |
+
|
240 |
+
if export_to_disk:
|
241 |
+
export_profile = profile.copy()
|
242 |
+
export_profile.update(
|
243 |
+
dtype=pred_tracker_np.dtype,
|
244 |
+
count=pred_tracker_np.shape[0],
|
245 |
+
compress="lzw",
|
246 |
+
nodata=0,
|
247 |
+
driver="GTiff",
|
248 |
+
)
|
249 |
+
# if executer has been passed, submit the save_prediction function to it, to avoid blocking the main thread
|
250 |
+
if save_executor:
|
251 |
+
save_executor.submit(
|
252 |
+
save_prediction, output_path, export_profile, pred_tracker_np
|
253 |
+
)
|
254 |
+
# otherwise save the prediction directly
|
255 |
+
|
256 |
+
else:
|
257 |
+
save_prediction(output_path, export_profile, pred_tracker_np)
|
258 |
+
|
259 |
+
if pbar:
|
260 |
+
pbar.update(1)
|
261 |
+
return pred_tracker_np
|
262 |
+
|
263 |
+
|
264 |
+
def collect_models(
|
265 |
+
custom_models: Union[list[torch.nn.Module], torch.nn.Module],
|
266 |
+
inference_device: torch.device,
|
267 |
+
inference_dtype: torch.dtype,
|
268 |
+
source: str,
|
269 |
+
destination_model_dir: Union[str, Path, None] = None,
|
270 |
+
) -> list[torch.nn.Module]:
|
271 |
+
if not custom_models:
|
272 |
+
models = []
|
273 |
+
for model_details in get_models(model_dir=destination_model_dir, source=source):
|
274 |
+
models.append(
|
275 |
+
load_model_from_weights(
|
276 |
+
model_name=model_details["timm_model_name"],
|
277 |
+
weights_path=model_details["Path"],
|
278 |
+
device=inference_device,
|
279 |
+
dtype=inference_dtype,
|
280 |
+
)
|
281 |
+
)
|
282 |
+
else:
|
283 |
+
# if not a list, make it a list of models
|
284 |
+
if not isinstance(custom_models, list):
|
285 |
+
custom_models = [custom_models]
|
286 |
+
|
287 |
+
models = [
|
288 |
+
model.to(inference_dtype).to(inference_device) for model in custom_models
|
289 |
+
]
|
290 |
+
return models
|
291 |
+
|
292 |
+
|
293 |
+
def predict_from_array(
|
294 |
+
input_array: np.ndarray,
|
295 |
+
patch_size: int = 1000,
|
296 |
+
patch_overlap: int = 300,
|
297 |
+
batch_size: int = 1,
|
298 |
+
inference_device: Union[str, torch.device] = default_device(),
|
299 |
+
mosaic_device: Optional[Union[str, torch.device]] = None,
|
300 |
+
inference_dtype: Union[torch.dtype, str] = torch.float32,
|
301 |
+
export_confidence: bool = False,
|
302 |
+
softmax_output: bool = True,
|
303 |
+
no_data_value: int = 0,
|
304 |
+
apply_no_data_mask: bool = True,
|
305 |
+
custom_models: Union[list[torch.nn.Module], torch.nn.Module] = [],
|
306 |
+
pred_classes: int = 4,
|
307 |
+
destination_model_dir: Union[str, Path, None] = None,
|
308 |
+
model_download_source: str = "google_drive",
|
309 |
+
) -> np.ndarray:
|
310 |
+
"""Predict a cloud and cloud shadow mask from a Red, Green and NIR numpy array, with a spatial res between 10 m and 50 m.
|
311 |
+
|
312 |
+
Args:
|
313 |
+
input_array (np.ndarray): A numpy array with shape (3, height, width) representing the Red, Green and NIR bands.
|
314 |
+
patch_size (int, optional): Size of the patches for inference. Defaults to 1000.
|
315 |
+
patch_overlap (int, optional): Overlap between patches for inference. Defaults to 300.
|
316 |
+
batch_size (int, optional): Number of patches to process in a batch. Defaults to 1.
|
317 |
+
inference_device (Union[str, torch.device], optional): Device to use for inference (e.g., 'cpu', 'cuda', 'mps'). Defaults to the device returned by default_device().
|
318 |
+
mosaic_device (Union[str, torch.device], optional): Device to use for mosaicking patches. Defaults to inference device.
|
319 |
+
inference_dtype (Union[torch.dtype, str], optional): Data type for inference. Defaults to torch.float32.
|
320 |
+
export_confidence (bool, optional): If True, exports confidence maps instead of predicted classes. Defaults to False.
|
321 |
+
softmax_output (bool, optional): If True, applies a softmax to the output, only used if export_confidence = True. Defaults to True.
|
322 |
+
no_data_value (int, optional): Value within input scenes that specifies no data region. Defaults to 0.
|
323 |
+
apply_no_data_mask (bool, optional): If True, applies a no-data mask to the predictions. Defaults to True.
|
324 |
+
custom_models Union[list[torch.nn.Module], torch.nn.Module], optional): A list or singular custom torch models to use for prediction. Defaults to [].
|
325 |
+
pred_classes (int, optional): Number of classes to predict. Defaults to 4, to be used with custom models.
|
326 |
+
destination_model_dir Union[str, Path, None]: Directory to save the model weights. Defaults to None.
|
327 |
+
model_download_source (str, optional): Source from which to download the model weights. Defaults to "google_drive", can also be "hugging_face".
|
328 |
+
Returns:
|
329 |
+
np.ndarray: A numpy array with shape (1, height, width) or (4, height, width if export_confidence = True) representing the predicted cloud and cloud shadow mask.
|
330 |
+
|
331 |
+
"""
|
332 |
+
|
333 |
+
inference_device = torch.device(inference_device)
|
334 |
+
if mosaic_device is None:
|
335 |
+
mosaic_device = inference_device
|
336 |
+
else:
|
337 |
+
mosaic_device = torch.device(mosaic_device)
|
338 |
+
|
339 |
+
inference_dtype = get_torch_dtype(inference_dtype)
|
340 |
+
# if no custom model paths are provided, use the default models
|
341 |
+
models = collect_models(
|
342 |
+
custom_models=custom_models,
|
343 |
+
inference_device=inference_device,
|
344 |
+
inference_dtype=inference_dtype,
|
345 |
+
source=model_download_source,
|
346 |
+
destination_model_dir=destination_model_dir,
|
347 |
+
)
|
348 |
+
|
349 |
+
pred_tracker = coordinator(
|
350 |
+
input_array=input_array,
|
351 |
+
models=models,
|
352 |
+
inference_device=inference_device,
|
353 |
+
mosaic_device=mosaic_device,
|
354 |
+
inference_dtype=inference_dtype,
|
355 |
+
export_confidence=export_confidence,
|
356 |
+
softmax_output=softmax_output,
|
357 |
+
patch_size=patch_size,
|
358 |
+
patch_overlap=patch_overlap,
|
359 |
+
batch_size=batch_size,
|
360 |
+
no_data_value=no_data_value,
|
361 |
+
export_to_disk=False,
|
362 |
+
apply_no_data_mask=apply_no_data_mask,
|
363 |
+
pred_classes=pred_classes,
|
364 |
+
)
|
365 |
+
|
366 |
+
return pred_tracker
|
367 |
+
|
368 |
+
|
369 |
+
def predict_from_load_func(
|
370 |
+
scene_paths: Union[list[Path], list[str]],
|
371 |
+
load_func: Callable,
|
372 |
+
patch_size: int = 1000,
|
373 |
+
patch_overlap: int = 300,
|
374 |
+
batch_size: int = 1,
|
375 |
+
inference_device: Union[str, torch.device] = default_device(),
|
376 |
+
mosaic_device: Optional[Union[str, torch.device]] = None,
|
377 |
+
inference_dtype: Union[torch.dtype, str] = torch.float32,
|
378 |
+
export_confidence: bool = False,
|
379 |
+
softmax_output: bool = True,
|
380 |
+
no_data_value: int = 0,
|
381 |
+
overwrite: bool = True,
|
382 |
+
apply_no_data_mask: bool = True,
|
383 |
+
output_dir: Optional[Union[Path, str]] = None,
|
384 |
+
custom_models: Union[list[torch.nn.Module], torch.nn.Module] = [],
|
385 |
+
destination_model_dir: Union[str, Path, None] = None,
|
386 |
+
model_download_source: str = "google_drive",
|
387 |
+
) -> list[Path]:
|
388 |
+
"""
|
389 |
+
Predicts cloud and cloud shadow masks for a list of scenes using a specified loading function.
|
390 |
+
|
391 |
+
Args:
|
392 |
+
scene_paths (Union[list[Path], list[str]]): A list of paths to the scene files to be processed.
|
393 |
+
load_func (Callable): A function to load the scene data. This function should take an input_path parameter and return a R,G,NIR numpy array and a rasterio for export profile, several load func are provided within data_loaders.py
|
394 |
+
patch_size (int, optional): Size of the patches for inference. Defaults to 1000.
|
395 |
+
patch_overlap (int, optional): Overlap between patches for inference. Defaults to 300.
|
396 |
+
batch_size (int, optional): Number of patches to process in a batch. Defaults to 1.
|
397 |
+
inference_device (Union[str, torch.device], optional): Device to use for inference (e.g., 'cpu', 'cuda', 'mps'). Defaults to the device returned by default_device().
|
398 |
+
mosaic_device (Union[str, torch.device], optional): Device to use for mosaicking patches. Defaults to inference device.
|
399 |
+
inference_dtype (Union[torch.dtype, str], optional): Data type for inference. Defaults to torch.float32.
|
400 |
+
export_confidence (bool, optional): If True, exports confidence maps instead of predicted classes. Defaults to False.
|
401 |
+
softmax_output (bool, optional): If True, applies a softmax to the output, only used if export_confidence = True. Defaults to True.
|
402 |
+
no_data_value (int, optional): Value within input scenes that specifies no data region. Defaults to 0.
|
403 |
+
overwrite (bool, optional): If False, skips scenes that already have a prediction file. Defaults to True.
|
404 |
+
apply_no_data_mask (bool, optional): If True, applies a no-data mask to the predictions. Defaults to True.
|
405 |
+
output_dir (Optional[Union[Path, str]], optional): Directory to save the prediction files. Defaults to None. If None, the predictions will be saved in the same directory as the input scene.
|
406 |
+
custom_models Union[list[torch.nn.Module], torch.nn.Module], optional): A list or singular custom torch models to use for prediction. Defaults to [].
|
407 |
+
destination_model_dir Union[str, Path, None]: Directory to save the model weights. Defaults to None.
|
408 |
+
model_download_source (str, optional): Source from which to download the model weights. Defaults to "google_drive", can also be "hugging_face".
|
409 |
+
|
410 |
+
Returns:
|
411 |
+
list[Path]: A list of paths to the output prediction files.
|
412 |
+
|
413 |
+
"""
|
414 |
+
pred_paths = []
|
415 |
+
inf_thread = Thread()
|
416 |
+
save_executor = ThreadPoolExecutor(max_workers=1)
|
417 |
+
|
418 |
+
inference_device = torch.device(inference_device)
|
419 |
+
if mosaic_device is None:
|
420 |
+
mosaic_device = inference_device
|
421 |
+
else:
|
422 |
+
mosaic_device = torch.device(mosaic_device)
|
423 |
+
|
424 |
+
inference_dtype = get_torch_dtype(inference_dtype)
|
425 |
+
|
426 |
+
models = collect_models(
|
427 |
+
custom_models=custom_models,
|
428 |
+
inference_device=inference_device,
|
429 |
+
inference_dtype=inference_dtype,
|
430 |
+
destination_model_dir=destination_model_dir,
|
431 |
+
source=model_download_source,
|
432 |
+
)
|
433 |
+
|
434 |
+
pbar = tqdm(
|
435 |
+
total=len(scene_paths),
|
436 |
+
desc=f"Running inference using {inference_device.type} {str(inference_dtype).split('.')[-1]}",
|
437 |
+
)
|
438 |
+
|
439 |
+
for scene_path in scene_paths:
|
440 |
+
scene_path = Path(scene_path)
|
441 |
+
file_name = f"{scene_path.stem}_OCM_v{__version__.replace('.','_')}.tif"
|
442 |
+
|
443 |
+
if output_dir is None:
|
444 |
+
output_path = scene_path.parent / file_name
|
445 |
+
else:
|
446 |
+
Path(output_dir).mkdir(parents=True, exist_ok=True)
|
447 |
+
output_path = Path(output_dir) / file_name
|
448 |
+
|
449 |
+
pred_paths.append(output_path)
|
450 |
+
|
451 |
+
if output_path.exists() and not overwrite:
|
452 |
+
pbar.update(1)
|
453 |
+
pbar.refresh()
|
454 |
+
continue
|
455 |
+
|
456 |
+
input_array, profile = load_func(input_path=scene_path)
|
457 |
+
|
458 |
+
while inf_thread.is_alive():
|
459 |
+
inf_thread.join()
|
460 |
+
|
461 |
+
inf_thread = Thread(
|
462 |
+
target=coordinator,
|
463 |
+
kwargs={
|
464 |
+
"input_array": input_array,
|
465 |
+
"profile": profile,
|
466 |
+
"output_path": output_path,
|
467 |
+
"models": models,
|
468 |
+
"inference_dtype": inference_dtype,
|
469 |
+
"export_confidence": export_confidence,
|
470 |
+
"softmax_output": softmax_output,
|
471 |
+
"inference_device": inference_device,
|
472 |
+
"mosaic_device": mosaic_device,
|
473 |
+
"patch_size": patch_size,
|
474 |
+
"patch_overlap": patch_overlap,
|
475 |
+
"batch_size": batch_size,
|
476 |
+
"no_data_value": no_data_value,
|
477 |
+
"pbar": pbar,
|
478 |
+
"apply_no_data_mask": apply_no_data_mask,
|
479 |
+
"save_executor": save_executor,
|
480 |
+
},
|
481 |
+
)
|
482 |
+
inf_thread.start()
|
483 |
+
|
484 |
+
while inf_thread.is_alive():
|
485 |
+
inf_thread.join()
|
486 |
+
|
487 |
+
if inference_device.type.startswith("cuda"):
|
488 |
+
torch.cuda.empty_cache()
|
489 |
+
|
490 |
+
save_executor.shutdown(wait=True)
|
491 |
+
pbar.refresh()
|
492 |
+
|
493 |
+
return pred_paths
|
omnicloudmask/data_loaders.py
ADDED
@@ -0,0 +1,159 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import warnings
|
2 |
+
from pathlib import Path
|
3 |
+
from typing import Optional, Union
|
4 |
+
|
5 |
+
import numpy as np
|
6 |
+
import rasterio as rio
|
7 |
+
from rasterio.profiles import Profile
|
8 |
+
|
9 |
+
|
10 |
+
def load_s2(
|
11 |
+
input_path: Union[Path, str],
|
12 |
+
resolution: float = 10.0,
|
13 |
+
required_bands: list[str] = ["B04", "B03", "B8A"],
|
14 |
+
) -> tuple[np.ndarray, Profile]:
|
15 |
+
"""Load a Sentinel-2 (L1C or L2A) image from a SAFE folder containing the bands"""
|
16 |
+
if not 10 <= resolution <= 50:
|
17 |
+
raise ValueError("Resolution must be between 10 and 50")
|
18 |
+
input_path = Path(input_path)
|
19 |
+
processing_level = find_s2_processing_level(input_path)
|
20 |
+
return open_s2_bands(input_path, processing_level, resolution, required_bands)
|
21 |
+
|
22 |
+
|
23 |
+
def find_s2_processing_level(
|
24 |
+
input_path: Path,
|
25 |
+
) -> str:
|
26 |
+
"""Derive the processing level of a Sentinel-2 image from the folder name."""
|
27 |
+
|
28 |
+
folder_name = Path(input_path).name
|
29 |
+
processing_level = folder_name.split("_")[1][3:6]
|
30 |
+
|
31 |
+
if processing_level not in ["L1C", "L2A"]:
|
32 |
+
raise ValueError(
|
33 |
+
f"Processing level {processing_level} not recognized, expected L1C or L2A"
|
34 |
+
)
|
35 |
+
return processing_level
|
36 |
+
|
37 |
+
|
38 |
+
def open_s2_bands(
|
39 |
+
input_path: Path,
|
40 |
+
processing_level: str,
|
41 |
+
resolution: float,
|
42 |
+
required_bands: list[str],
|
43 |
+
) -> tuple[np.ndarray, Profile]:
|
44 |
+
bands = []
|
45 |
+
for band_name in required_bands:
|
46 |
+
if processing_level == "L1C":
|
47 |
+
try:
|
48 |
+
band = list(input_path.rglob(f"*IMG_DATA/*{band_name}.jp2"))[0]
|
49 |
+
|
50 |
+
except IndexError:
|
51 |
+
raise ValueError(f"Band {band_name} not found in {input_path}")
|
52 |
+
else:
|
53 |
+
band = None
|
54 |
+
for search_resolution in [10, 20, 60]:
|
55 |
+
band_paths = list(
|
56 |
+
input_path.rglob(f"*{band_name}_{search_resolution}m.jp2")
|
57 |
+
)
|
58 |
+
if band_paths:
|
59 |
+
band = band_paths[0]
|
60 |
+
break
|
61 |
+
if not band:
|
62 |
+
raise ValueError(f"Band {band_name} not found in {input_path}")
|
63 |
+
|
64 |
+
with rio.open(band) as src:
|
65 |
+
profile = src.profile
|
66 |
+
native_resolution = int(src.res[0])
|
67 |
+
scale_factor = native_resolution / resolution
|
68 |
+
if native_resolution == resolution:
|
69 |
+
bands.append(src.read(1))
|
70 |
+
else:
|
71 |
+
bands.append(
|
72 |
+
src.read(
|
73 |
+
1,
|
74 |
+
out_shape=(
|
75 |
+
int(src.height * scale_factor),
|
76 |
+
int(src.width * scale_factor),
|
77 |
+
),
|
78 |
+
)
|
79 |
+
)
|
80 |
+
profile["transform"] = rio.transform.from_origin( # type: ignore
|
81 |
+
profile["transform"][2],
|
82 |
+
profile["transform"][5],
|
83 |
+
resolution,
|
84 |
+
resolution,
|
85 |
+
)
|
86 |
+
data = np.array(bands)
|
87 |
+
profile["height"] = data.shape[1]
|
88 |
+
profile["width"] = data.shape[2]
|
89 |
+
return data, profile
|
90 |
+
|
91 |
+
|
92 |
+
def load_multiband(
|
93 |
+
input_path: Union[Path, str],
|
94 |
+
resample_res: Optional[float] = None,
|
95 |
+
band_order: Optional[list[int]] = None,
|
96 |
+
) -> tuple[np.ndarray, Profile]:
|
97 |
+
"""Load a multiband image and resample it to requested resolution."""
|
98 |
+
if band_order is None:
|
99 |
+
warnings.warn(
|
100 |
+
"No band order provided, using default [1, 2, 3] (RGN)", UserWarning
|
101 |
+
)
|
102 |
+
band_order = [1, 2, 3]
|
103 |
+
input_path = Path(input_path)
|
104 |
+
|
105 |
+
with rio.open(input_path) as src:
|
106 |
+
if resample_res:
|
107 |
+
current_res = src.res
|
108 |
+
desired_res = (resample_res, resample_res)
|
109 |
+
scale_factor = (
|
110 |
+
current_res[0] / desired_res[0],
|
111 |
+
current_res[1] / desired_res[1],
|
112 |
+
)
|
113 |
+
else:
|
114 |
+
scale_factor = (1, 1)
|
115 |
+
|
116 |
+
data = src.read(
|
117 |
+
band_order,
|
118 |
+
out_shape=(
|
119 |
+
len(band_order),
|
120 |
+
int(src.height * scale_factor[0]),
|
121 |
+
int(src.width * scale_factor[1]),
|
122 |
+
),
|
123 |
+
resampling=rio.enums.Resampling.nearest, # type: ignore
|
124 |
+
)
|
125 |
+
profile = src.profile
|
126 |
+
|
127 |
+
return data, profile
|
128 |
+
|
129 |
+
|
130 |
+
def load_ls8(
|
131 |
+
input_path: Union[Path, str],
|
132 |
+
resolution: int = 30,
|
133 |
+
required_bands=["B4", "B3", "B5"],
|
134 |
+
) -> tuple[np.ndarray, Profile]:
|
135 |
+
"""Load a Landsat 8 image from a folder containing the bands"""
|
136 |
+
if resolution != 30:
|
137 |
+
raise ValueError("Resolution must be 30")
|
138 |
+
|
139 |
+
input_path = Path(input_path)
|
140 |
+
|
141 |
+
band_files = {}
|
142 |
+
for band_name in required_bands:
|
143 |
+
try:
|
144 |
+
band = list(input_path.rglob(f"*{band_name}.TIF"))[0]
|
145 |
+
|
146 |
+
except IndexError:
|
147 |
+
raise ValueError(f"Band {band_name} not found in {input_path}")
|
148 |
+
band_files[band_name] = band
|
149 |
+
|
150 |
+
data = []
|
151 |
+
profile = Profile()
|
152 |
+
for band_name in required_bands:
|
153 |
+
with rio.open(band_files[band_name]) as src:
|
154 |
+
if not profile:
|
155 |
+
profile = src.profile
|
156 |
+
data.append(src.read(1))
|
157 |
+
|
158 |
+
data = np.array(data)
|
159 |
+
return data, profile
|
omnicloudmask/download_models.py
ADDED
@@ -0,0 +1,92 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from pathlib import Path
|
2 |
+
from typing import Union
|
3 |
+
|
4 |
+
import gdown
|
5 |
+
import pandas as pd
|
6 |
+
import torch
|
7 |
+
from huggingface_hub import hf_hub_download
|
8 |
+
from safetensors.torch import load_file
|
9 |
+
|
10 |
+
|
11 |
+
def download_file_from_google_drive(file_id: str, destination: Path) -> None:
|
12 |
+
"""
|
13 |
+
Downloads a file from Google Drive and saves it at the given destination using gdown.
|
14 |
+
|
15 |
+
Args:
|
16 |
+
file_id (str): The ID of the file on Google Drive.
|
17 |
+
destination (Path): The local path where the file should be saved.
|
18 |
+
"""
|
19 |
+
url = f"https://drive.google.com/uc?id={file_id}"
|
20 |
+
gdown.download(url, str(destination), quiet=False)
|
21 |
+
|
22 |
+
|
23 |
+
def download_file_from_hugging_face(destination: Path) -> None:
|
24 |
+
"""
|
25 |
+
Downloads a file from Hugging Face and saves it at the given destination using hf_hub_download.
|
26 |
+
Loads the resulting safetensors file and saves it as a PyTorch model state for compatibility with the rest of the codebase.
|
27 |
+
|
28 |
+
Args:
|
29 |
+
file_id (str): The ID of the file on Hugging Face.
|
30 |
+
destination (Path): The local path where the file should be saved.
|
31 |
+
"""
|
32 |
+
file_name = destination.stem
|
33 |
+
safetensor_path = hf_hub_download(
|
34 |
+
repo_id="NickWright/OmniCloudMask",
|
35 |
+
filename=f"{file_name}.safetensors",
|
36 |
+
force_download=True,
|
37 |
+
cache_dir=destination.parent,
|
38 |
+
)
|
39 |
+
model_state = load_file(safetensor_path)
|
40 |
+
torch.save(model_state, destination)
|
41 |
+
|
42 |
+
|
43 |
+
def download_file(file_id: str, destination: Path, source: str) -> None:
|
44 |
+
if source == "google_drive":
|
45 |
+
download_file_from_google_drive(file_id, destination)
|
46 |
+
elif source == "hugging_face":
|
47 |
+
download_file_from_hugging_face(destination)
|
48 |
+
else:
|
49 |
+
raise ValueError(
|
50 |
+
"Invalid source. Supported sources are 'google_drive' and 'hugging_face'."
|
51 |
+
)
|
52 |
+
|
53 |
+
|
54 |
+
def get_models(
|
55 |
+
force_download: bool = False,
|
56 |
+
model_dir: Union[str, Path, None] = None,
|
57 |
+
source: str = "google_drive",
|
58 |
+
) -> list[dict]:
|
59 |
+
"""
|
60 |
+
Downloads the model weights from Google Drive and saves them locally.
|
61 |
+
|
62 |
+
Args:
|
63 |
+
force_download (bool): Whether to force download the model weights even if they already exist locally.
|
64 |
+
model_dir (Union[str, Path, None]): The directory where the model weights should be saved.
|
65 |
+
source (str): The source from which the model weights should be downloaded. Currently, only "google_drive" or "hugging_face" are supported.
|
66 |
+
"""
|
67 |
+
|
68 |
+
df = pd.read_csv(
|
69 |
+
Path(__file__).resolve().parent / "models/model_download_links.csv"
|
70 |
+
)
|
71 |
+
model_paths = []
|
72 |
+
|
73 |
+
for _, row in df.iterrows():
|
74 |
+
file_id = str(row["google_drive_id"])
|
75 |
+
|
76 |
+
if model_dir is not None:
|
77 |
+
model_dir = Path(model_dir)
|
78 |
+
else:
|
79 |
+
model_dir = Path(__file__).resolve().parent / "models"
|
80 |
+
|
81 |
+
model_dir.mkdir(exist_ok=True)
|
82 |
+
destination = model_dir / str(row["file_name"])
|
83 |
+
timm_model_name = row["timm_model_name"]
|
84 |
+
|
85 |
+
if not destination.exists() or force_download:
|
86 |
+
download_file(file_id=file_id, destination=destination, source=source)
|
87 |
+
|
88 |
+
elif destination.stat().st_size <= 1024 * 1024:
|
89 |
+
download_file(file_id=file_id, destination=destination, source=source)
|
90 |
+
|
91 |
+
model_paths.append({"Path": destination, "timm_model_name": timm_model_name})
|
92 |
+
return model_paths
|
omnicloudmask/model_utils.py
ADDED
@@ -0,0 +1,208 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from functools import partial
|
2 |
+
from pathlib import Path
|
3 |
+
from typing import Optional, Union
|
4 |
+
|
5 |
+
import numpy as np
|
6 |
+
import timm
|
7 |
+
import torch
|
8 |
+
from fastai.vision.learner import create_unet_model
|
9 |
+
|
10 |
+
|
11 |
+
def get_torch_dtype(dtype: Union[torch.dtype, str]) -> torch.dtype:
|
12 |
+
"""Return a torch.dtype from a string or torch.dtype."""
|
13 |
+
if isinstance(dtype, str):
|
14 |
+
dtype_mapping = {
|
15 |
+
"float16": torch.float16,
|
16 |
+
"half": torch.float16,
|
17 |
+
"fp16": torch.float16,
|
18 |
+
"float32": torch.float32,
|
19 |
+
"float": torch.float32,
|
20 |
+
"bfloat16": torch.bfloat16,
|
21 |
+
"bf16": torch.bfloat16,
|
22 |
+
}
|
23 |
+
try:
|
24 |
+
return dtype_mapping[dtype.lower()]
|
25 |
+
except KeyError:
|
26 |
+
raise ValueError(
|
27 |
+
f"Invalid dtype: {dtype}. Must be one of {list(dtype_mapping.keys())}"
|
28 |
+
)
|
29 |
+
elif isinstance(dtype, torch.dtype):
|
30 |
+
return dtype
|
31 |
+
else:
|
32 |
+
raise TypeError(
|
33 |
+
f"Expected dtype to be a str or torch.dtype, but got {type(dtype)}"
|
34 |
+
)
|
35 |
+
|
36 |
+
|
37 |
+
def create_gradient_mask(
|
38 |
+
patch_size: int, patch_overlap: int, device: torch.device, dtype: torch.dtype
|
39 |
+
) -> torch.Tensor:
|
40 |
+
"""Create a gradient mask for a given patch size and overlap."""
|
41 |
+
if patch_overlap > 0:
|
42 |
+
if patch_overlap * 2 > patch_size:
|
43 |
+
patch_overlap = patch_size // 2
|
44 |
+
|
45 |
+
gradient_strength = 1
|
46 |
+
gradient = (
|
47 |
+
torch.ones((patch_size, patch_size), dtype=torch.int, device=device)
|
48 |
+
* patch_overlap
|
49 |
+
)
|
50 |
+
gradient[:, :patch_overlap] = torch.tile(
|
51 |
+
torch.arange(1, patch_overlap + 1),
|
52 |
+
(patch_size, 1),
|
53 |
+
)
|
54 |
+
gradient[:, -patch_overlap:] = torch.tile(
|
55 |
+
torch.arange(patch_overlap, 0, -1),
|
56 |
+
(patch_size, 1),
|
57 |
+
)
|
58 |
+
gradient = gradient / patch_overlap
|
59 |
+
rotated_gradient = torch.rot90(gradient)
|
60 |
+
combined_gradient = rotated_gradient * gradient
|
61 |
+
|
62 |
+
combined_gradient = (combined_gradient * gradient_strength) + (
|
63 |
+
1 - gradient_strength
|
64 |
+
)
|
65 |
+
else:
|
66 |
+
combined_gradient = torch.ones(
|
67 |
+
(patch_size, patch_size), dtype=torch.int, device=device
|
68 |
+
)
|
69 |
+
return combined_gradient.to(dtype)
|
70 |
+
|
71 |
+
|
72 |
+
def channel_norm(patch: np.ndarray, nodata_value: Optional[int] = 0) -> np.ndarray:
|
73 |
+
"""Normalize each band of the input array by subtracting the nonzero mean and dividing
|
74 |
+
by the nonzero standard deviation then fill nodata values with 0."""
|
75 |
+
out_array = np.zeros(patch.shape).astype(np.float32)
|
76 |
+
for id, band in enumerate(patch):
|
77 |
+
# Mask for non-zero values
|
78 |
+
mask = band != nodata_value
|
79 |
+
# Check if there are any non-zero values
|
80 |
+
if np.any(mask):
|
81 |
+
mean = band[mask].mean()
|
82 |
+
std = band[mask].std()
|
83 |
+
if std == 0:
|
84 |
+
std = 1 # Prevent division by zero
|
85 |
+
# Normalize only non-zero values
|
86 |
+
out_array[id][mask] = (band[mask] - mean) / std
|
87 |
+
else:
|
88 |
+
continue
|
89 |
+
# Fill original nodata values with 0
|
90 |
+
out_array[id][~mask] = 0
|
91 |
+
return out_array
|
92 |
+
|
93 |
+
|
94 |
+
def store_results(
|
95 |
+
pred_batch: torch.Tensor,
|
96 |
+
index_batch: list[tuple],
|
97 |
+
pred_tracker: torch.Tensor,
|
98 |
+
gradient: torch.Tensor,
|
99 |
+
grad_tracker: Optional[torch.Tensor] = None,
|
100 |
+
) -> None:
|
101 |
+
"""Store the results of the model inference in the pred_tracker and grad_tracker tensors."""
|
102 |
+
# Store the predictions in the pred_tracker tensor
|
103 |
+
assert pred_batch.ndim == 4, "pred_batch must have 4 dimensions, (B, class, H, W)"
|
104 |
+
assert pred_batch.shape[0] == len(index_batch), "Batch size must match index_batch"
|
105 |
+
assert pred_batch.shape[1] == pred_tracker.shape[0], "Number of classes must match"
|
106 |
+
assert pred_batch.shape[2] == gradient.shape[0], "Height must match gradient"
|
107 |
+
assert pred_batch.shape[3] == gradient.shape[1], "Width must match gradient"
|
108 |
+
|
109 |
+
pred_batch *= gradient[None, None, :, :]
|
110 |
+
|
111 |
+
for pred, index in zip(pred_batch.to(pred_tracker.device), index_batch):
|
112 |
+
pred_tracker[:, index[0] : index[1], index[2] : index[3]] += pred
|
113 |
+
if grad_tracker is not None:
|
114 |
+
grad_tracker[index[0] : index[1], index[2] : index[3]] += gradient.to(
|
115 |
+
grad_tracker.device
|
116 |
+
)
|
117 |
+
|
118 |
+
|
119 |
+
def inference_and_store(
|
120 |
+
models: list[torch.nn.Module],
|
121 |
+
patch_batch: torch.Tensor,
|
122 |
+
index_batch: list[tuple],
|
123 |
+
pred_tracker: torch.Tensor,
|
124 |
+
gradient: torch.Tensor,
|
125 |
+
grad_tracker: Optional[torch.Tensor] = None,
|
126 |
+
) -> None:
|
127 |
+
"""Perform inference on the patch_batch and store the results in the pred_tracker and grad_tracker tensors."""
|
128 |
+
# pre-initialize the all_preds tensor to store the predictions from each model
|
129 |
+
all_preds = torch.zeros(
|
130 |
+
len(models),
|
131 |
+
patch_batch.shape[0],
|
132 |
+
pred_tracker.shape[0],
|
133 |
+
patch_batch.shape[2],
|
134 |
+
patch_batch.shape[3],
|
135 |
+
device=patch_batch.device,
|
136 |
+
dtype=patch_batch.dtype,
|
137 |
+
)
|
138 |
+
for index, model in enumerate(models):
|
139 |
+
with torch.no_grad():
|
140 |
+
all_preds[index] = model(patch_batch)
|
141 |
+
|
142 |
+
mean_preds = all_preds.mean(dim=0)
|
143 |
+
|
144 |
+
store_results(
|
145 |
+
pred_batch=mean_preds,
|
146 |
+
index_batch=index_batch,
|
147 |
+
pred_tracker=pred_tracker,
|
148 |
+
gradient=gradient,
|
149 |
+
grad_tracker=grad_tracker,
|
150 |
+
)
|
151 |
+
|
152 |
+
|
153 |
+
def default_device() -> torch.device:
|
154 |
+
"""Return the default device for model inference"""
|
155 |
+
if torch.cuda.is_available():
|
156 |
+
return torch.device("cuda")
|
157 |
+
elif torch.backends.mps.is_available():
|
158 |
+
return torch.device("mps")
|
159 |
+
return torch.device("cpu")
|
160 |
+
|
161 |
+
|
162 |
+
def load_model(
|
163 |
+
model_path: Union[Path, str],
|
164 |
+
device: torch.device,
|
165 |
+
dtype: torch.dtype = torch.float32,
|
166 |
+
) -> torch.nn.Module:
|
167 |
+
"""Load a PyTorch model from a file and move it to the specified device and dtype."""
|
168 |
+
model_path = Path(model_path)
|
169 |
+
if not model_path.is_file():
|
170 |
+
raise FileNotFoundError(f"Model file not found at: {model_path}")
|
171 |
+
|
172 |
+
try:
|
173 |
+
model = torch.load(model_path, map_location="cpu")
|
174 |
+
except Exception as e:
|
175 |
+
raise RuntimeError(f"Error loading model: {e}")
|
176 |
+
|
177 |
+
model.eval()
|
178 |
+
return model.to(dtype).to(device)
|
179 |
+
|
180 |
+
|
181 |
+
def load_model_from_weights(
|
182 |
+
model_name: str,
|
183 |
+
weights_path: Union[Path, str],
|
184 |
+
device: torch.device,
|
185 |
+
dtype: torch.dtype = torch.float32,
|
186 |
+
in_chans: int = 3,
|
187 |
+
n_out: int = 4,
|
188 |
+
) -> torch.nn.Module:
|
189 |
+
"""Build Fastai DynamicUnet model from timm model and load weights from file"""
|
190 |
+
timm_model = partial(
|
191 |
+
timm.create_model,
|
192 |
+
model_name=model_name,
|
193 |
+
pretrained=False,
|
194 |
+
in_chans=in_chans,
|
195 |
+
)
|
196 |
+
|
197 |
+
model = create_unet_model(
|
198 |
+
arch=timm_model,
|
199 |
+
n_out=n_out,
|
200 |
+
img_size=(509, 509),
|
201 |
+
act_cls=torch.nn.Mish,
|
202 |
+
pretrained=False,
|
203 |
+
)
|
204 |
+
|
205 |
+
model.load_state_dict(torch.load(weights_path, weights_only=True))
|
206 |
+
model.eval()
|
207 |
+
|
208 |
+
return model.to(dtype).to(device)
|
omnicloudmask/models/PM_model_2.2.10_RG_NIR_509_convnextv2_nano.fcmae_ft_in1k_PT_state.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:7d83ddef55797fd443cb30fdd545edf4f070c76cfb031ab53af2cd01f51d6d0f
|
3 |
+
size 130226202
|
omnicloudmask/models/PM_model_2.2.10_RG_NIR_509_regnety_004.pycls_in1k_PT_state.pth
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
version https://git-lfs.github.com/spec/v1
|
2 |
+
oid sha256:be9e29fa69464a286d40e71b5af10894cabe2a258a6f1de4e869500ae704c7bd
|
3 |
+
size 72458313
|
omnicloudmask/models/model_download_links.csv
ADDED
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
1 |
+
file_name,timm_model_name,google_drive_id
|
2 |
+
PM_model_2.2.10_RG_NIR_509_regnety_004.pycls_in1k_PT_state.pth,regnety_004,1tGJh9nnrH-apjmV70AcK8VtXnbBtRb67
|
3 |
+
PM_model_2.2.10_RG_NIR_509_convnextv2_nano.fcmae_ft_in1k_PT_state.pth,convnextv2_nano,1QXQ_oPhLKEowC9fxlZGLOACt8gCNMbWP
|
omnicloudmask/raster_utils.py
ADDED
@@ -0,0 +1,118 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
from pathlib import Path
|
2 |
+
from typing import Optional
|
3 |
+
|
4 |
+
import numpy as np
|
5 |
+
import rasterio as rio
|
6 |
+
from rasterio.profiles import Profile
|
7 |
+
|
8 |
+
from .model_utils import channel_norm
|
9 |
+
|
10 |
+
|
11 |
+
def get_patch(
|
12 |
+
input_array: np.ndarray,
|
13 |
+
index: tuple,
|
14 |
+
no_data_value: Optional[int] = 0,
|
15 |
+
) -> tuple[Optional[np.ndarray], Optional[tuple[int, int, int, int]]]:
|
16 |
+
"""Extract a patch from a 3D array and normalize it. If the patch is entirely nodata, return None.
|
17 |
+
If the patch contains nodata, try to move patches to reduce nodata regions in patches.
|
18 |
+
"""
|
19 |
+
assert input_array.ndim == 3, "Input array must have 3 dimensions"
|
20 |
+
|
21 |
+
top, bottom, left, right = index
|
22 |
+
patch = input_array[:, top:bottom, left:right].astype(np.float32)
|
23 |
+
|
24 |
+
if patch.sum() == 0:
|
25 |
+
return None, None
|
26 |
+
|
27 |
+
if no_data_value is None:
|
28 |
+
if np.all(patch == no_data_value):
|
29 |
+
return None, None
|
30 |
+
|
31 |
+
if np.any(patch == 0):
|
32 |
+
max_bottom, max_right = input_array.shape[1:3]
|
33 |
+
|
34 |
+
if np.any(patch[:, 0, :]) or np.any(patch[:, -1, :]):
|
35 |
+
while not np.any(patch[:, 0, :]) and bottom < max_bottom: # check top row
|
36 |
+
patch = patch[:, 1:, :]
|
37 |
+
top += 1
|
38 |
+
bottom += 1
|
39 |
+
|
40 |
+
while not np.any(patch[:, -1, :]) and top > 0:
|
41 |
+
patch = patch[:, :-1, :]
|
42 |
+
bottom -= 1
|
43 |
+
top -= 1
|
44 |
+
|
45 |
+
# Both sides are not zero-filled
|
46 |
+
if np.any(patch[:, :, 0]) or np.any(patch[:, :, -1]):
|
47 |
+
while not np.any(patch[:, :, 0]) and right < max_right: # check left column
|
48 |
+
patch = patch[:, :, 1:]
|
49 |
+
left += 1
|
50 |
+
right += 1
|
51 |
+
|
52 |
+
while not np.any(patch[:, :, -1]) and left > 0: # check right column
|
53 |
+
patch = patch[:, :, :-1]
|
54 |
+
right -= 1
|
55 |
+
left -= 1
|
56 |
+
patch = input_array[:, top:bottom, left:right].astype(np.float32)
|
57 |
+
index = (top, bottom, left, right)
|
58 |
+
|
59 |
+
# trim index bottom and right to match patch shape
|
60 |
+
index = (top, top + patch.shape[1], left, left + patch.shape[2])
|
61 |
+
return channel_norm(patch, no_data_value), index
|
62 |
+
|
63 |
+
|
64 |
+
def mask_prediction(
|
65 |
+
scene: np.ndarray, pred_tracker_np: np.ndarray, no_data_value: int = 0
|
66 |
+
) -> np.ndarray:
|
67 |
+
"""Create a no data mask from a raster scene."""
|
68 |
+
assert scene.ndim == 3, "Scene must have 3 dimensions"
|
69 |
+
assert pred_tracker_np.ndim == 3, "Prediction tracker must have 3 dimensions"
|
70 |
+
assert (
|
71 |
+
scene.shape[1:] == pred_tracker_np.shape[1:]
|
72 |
+
), "Scene and prediction tracker must have the same shape"
|
73 |
+
mask = np.all(scene != no_data_value, axis=0).astype(np.uint8)
|
74 |
+
pred_tracker_np *= mask
|
75 |
+
return pred_tracker_np
|
76 |
+
|
77 |
+
|
78 |
+
def make_patch_indexes(
|
79 |
+
array_width: int,
|
80 |
+
array_height: int,
|
81 |
+
patch_size: int = 1000,
|
82 |
+
patch_overlap: int = 300,
|
83 |
+
) -> list[tuple[int, int, int, int]]:
|
84 |
+
"""Create a list of patch indexes for a given shape and patch size."""
|
85 |
+
assert patch_size > patch_overlap, "Patch size must be greater than patch overlap"
|
86 |
+
assert patch_overlap >= 0, "Patch overlap must be greater than or equal to 0"
|
87 |
+
assert patch_size > 0, "Patch size must be greater than 0"
|
88 |
+
assert (
|
89 |
+
patch_size <= array_width
|
90 |
+
), "Patch size must be less than or equal to array width"
|
91 |
+
assert (
|
92 |
+
patch_size <= array_height
|
93 |
+
), "Patch size must be less than or equal to array height"
|
94 |
+
|
95 |
+
stride = patch_size - patch_overlap
|
96 |
+
|
97 |
+
max_bottom = array_height - patch_size
|
98 |
+
max_right = array_width - patch_size
|
99 |
+
|
100 |
+
patch_indexes = []
|
101 |
+
for top in range(0, array_height, stride):
|
102 |
+
if top > max_bottom:
|
103 |
+
top = max_bottom
|
104 |
+
bottom = top + patch_size
|
105 |
+
for left in range(0, array_width, stride):
|
106 |
+
if left > max_right:
|
107 |
+
left = max_right
|
108 |
+
right = left + patch_size
|
109 |
+
patch_indexes.append((top, bottom, left, right))
|
110 |
+
|
111 |
+
return patch_indexes
|
112 |
+
|
113 |
+
|
114 |
+
def save_prediction(
|
115 |
+
output_path: Path, export_profile: Profile, pred_tracker_np: np.ndarray
|
116 |
+
) -> None:
|
117 |
+
with rio.open(output_path, "w", **export_profile) as dst:
|
118 |
+
dst.write(pred_tracker_np)
|