Paolo-Fraccaro
commited on
Commit
•
cf9d671
1
Parent(s):
adbc22d
Create sen1floods11_Prithvi_100M.py
Browse files- sen1floods11_Prithvi_100M.py +291 -0
sen1floods11_Prithvi_100M.py
ADDED
@@ -0,0 +1,291 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
|
3 |
+
# base options
|
4 |
+
dist_params = dict(backend='nccl')
|
5 |
+
log_level = 'INFO'
|
6 |
+
load_from = None
|
7 |
+
resume_from = None
|
8 |
+
cudnn_benchmark = True
|
9 |
+
|
10 |
+
custom_imports = dict(imports=["geospatial_fm"])
|
11 |
+
|
12 |
+
|
13 |
+
### Configs
|
14 |
+
# Data
|
15 |
+
# TO BE DEFINED BY USER: Data root to sen1floods11 downloaded dataset
|
16 |
+
data_root = "<path to dataset>"
|
17 |
+
|
18 |
+
dataset_type = "GeospatialDataset"
|
19 |
+
num_classes=2
|
20 |
+
num_frames = 1
|
21 |
+
img_size = 224
|
22 |
+
num_workers = 2
|
23 |
+
samples_per_gpu = 4
|
24 |
+
CLASSES=(0,1)
|
25 |
+
|
26 |
+
img_norm_cfg = dict(means=[0.14245495, 0.13921481, 0.12434631, 0.31420089, 0.20743526,0.12046503],
|
27 |
+
stds=[0.04036231, 0.04186983, 0.05267646, 0.0822221 , 0.06834774, 0.05294205])
|
28 |
+
|
29 |
+
bands = [1, 2, 3, 8, 11, 12]
|
30 |
+
tile_size = img_size
|
31 |
+
orig_nsize = 512
|
32 |
+
crop_size = (tile_size, tile_size)
|
33 |
+
|
34 |
+
img_dir = data_root + "v1.1/data/flood_events/HandLabeled/S2Hand"
|
35 |
+
ann_dir = data_root + "v1.1/data/flood_events/HandLabeled/LabelHand"
|
36 |
+
img_suffix = f"_S2Hand.tif"
|
37 |
+
seg_map_suffix = f"_LabelHand.tif"
|
38 |
+
|
39 |
+
splits = {
|
40 |
+
"train": "data_splits/train_split.txt",
|
41 |
+
"val": "data_splits/val_split.txt",
|
42 |
+
"test": "data_splits/test_split.txt",
|
43 |
+
}
|
44 |
+
splits = {k: os.path.abspath(v) for (k, v) in splits.items()}
|
45 |
+
|
46 |
+
ignore_index = 2
|
47 |
+
label_nodata = -1
|
48 |
+
image_nodata = -9999
|
49 |
+
image_nodata_replace = 0
|
50 |
+
constant = 0.0001
|
51 |
+
|
52 |
+
# Model
|
53 |
+
# TO BE DEFINED BY USER: path to pretrained backbone weights
|
54 |
+
pretrained_weights_path = "<path to pretrained weights>"
|
55 |
+
num_layers = 12
|
56 |
+
patch_size = 16
|
57 |
+
embed_dim = 768
|
58 |
+
num_heads = 12
|
59 |
+
tubelet_size = 1
|
60 |
+
|
61 |
+
# TRAINING
|
62 |
+
epochs=100
|
63 |
+
eval_epoch_interval = 5
|
64 |
+
|
65 |
+
# TO BE DEFINED BY USER: Save directory
|
66 |
+
experiment = "<experiment name>"
|
67 |
+
project_dir = "<project dir>"
|
68 |
+
work_dir = os.path.join(project_dir, experiment)
|
69 |
+
save_path = work_dir
|
70 |
+
|
71 |
+
# Pipelines
|
72 |
+
train_pipeline = [
|
73 |
+
dict(
|
74 |
+
type="LoadGeospatialImageFromFile",
|
75 |
+
to_float32=False,
|
76 |
+
nodata=image_nodata,
|
77 |
+
nodata_replace=image_nodata_replace,
|
78 |
+
channels_last=False
|
79 |
+
),
|
80 |
+
dict(
|
81 |
+
type="LoadGeospatialAnnotations",
|
82 |
+
reduce_zero_label=False,
|
83 |
+
nodata=label_nodata,
|
84 |
+
nodata_replace=ignore_index,
|
85 |
+
),
|
86 |
+
dict(type="BandsExtract", bands=bands),
|
87 |
+
dict(type="ConstantMultiply", constant=constant),
|
88 |
+
dict(type="RandomFlip", prob=0.5),
|
89 |
+
dict(type="ToTensor", keys=["img", "gt_semantic_seg"]),
|
90 |
+
# to channels first
|
91 |
+
dict(type="TorchPermute", keys=["img"], order=(2, 0, 1)),
|
92 |
+
dict(type="TorchNormalize", **img_norm_cfg),
|
93 |
+
dict(type="TorchRandomCrop", crop_size=crop_size),
|
94 |
+
dict(
|
95 |
+
type="Reshape",
|
96 |
+
keys=["img"],
|
97 |
+
new_shape=(len(bands), num_frames, tile_size, tile_size),
|
98 |
+
),
|
99 |
+
dict(type="Reshape", keys=["gt_semantic_seg"], new_shape=(1, tile_size, tile_size)),
|
100 |
+
dict(type="CastTensor", keys=["gt_semantic_seg"], new_type="torch.LongTensor"),
|
101 |
+
dict(type="Collect", keys=["img", "gt_semantic_seg"]),
|
102 |
+
]
|
103 |
+
|
104 |
+
|
105 |
+
test_pipeline = [
|
106 |
+
dict(
|
107 |
+
type="LoadGeospatialImageFromFile",
|
108 |
+
to_float32=False,
|
109 |
+
nodata=image_nodata,
|
110 |
+
nodata_replace=image_nodata_replace,
|
111 |
+
channels_last=False
|
112 |
+
),
|
113 |
+
dict(type="BandsExtract", bands=bands),
|
114 |
+
dict(type="ConstantMultiply", constant=constant),
|
115 |
+
dict(type="ToTensor", keys=["img"]),
|
116 |
+
# to channels first
|
117 |
+
dict(type="TorchPermute", keys=["img"], order=(2, 0, 1)),
|
118 |
+
dict(type="TorchNormalize", **img_norm_cfg),
|
119 |
+
dict(
|
120 |
+
type="Reshape",
|
121 |
+
keys=["img"],
|
122 |
+
new_shape=(len(bands), num_frames, -1, -1),
|
123 |
+
look_up={'2': 1, '3': 2}
|
124 |
+
),
|
125 |
+
dict(type="CastTensor", keys=["img"], new_type="torch.FloatTensor"),
|
126 |
+
dict(
|
127 |
+
type="CollectTestList",
|
128 |
+
keys=["img"],
|
129 |
+
meta_keys=[
|
130 |
+
"img_info",
|
131 |
+
"seg_fields",
|
132 |
+
"img_prefix",
|
133 |
+
"seg_prefix",
|
134 |
+
"filename",
|
135 |
+
"ori_filename",
|
136 |
+
"img",
|
137 |
+
"img_shape",
|
138 |
+
"ori_shape",
|
139 |
+
"pad_shape",
|
140 |
+
"scale_factor",
|
141 |
+
"img_norm_cfg",
|
142 |
+
],
|
143 |
+
),
|
144 |
+
]
|
145 |
+
|
146 |
+
# Dataset
|
147 |
+
data = dict(
|
148 |
+
samples_per_gpu=samples_per_gpu,
|
149 |
+
workers_per_gpu=num_workers,
|
150 |
+
train=dict(
|
151 |
+
type=dataset_type,
|
152 |
+
CLASSES=CLASSES,
|
153 |
+
data_root=data_root,
|
154 |
+
img_dir=img_dir,
|
155 |
+
ann_dir=ann_dir,
|
156 |
+
img_suffix=img_suffix,
|
157 |
+
seg_map_suffix=seg_map_suffix,
|
158 |
+
pipeline=train_pipeline,
|
159 |
+
ignore_index=ignore_index,
|
160 |
+
split=splits["train"],
|
161 |
+
),
|
162 |
+
val=dict(
|
163 |
+
type=dataset_type,
|
164 |
+
CLASSES=CLASSES,
|
165 |
+
data_root=data_root,
|
166 |
+
img_dir=img_dir,
|
167 |
+
ann_dir=ann_dir,
|
168 |
+
img_suffix=img_suffix,
|
169 |
+
seg_map_suffix=seg_map_suffix,
|
170 |
+
pipeline=test_pipeline,
|
171 |
+
ignore_index=ignore_index,
|
172 |
+
split=splits["val"],
|
173 |
+
gt_seg_map_loader_cfg=dict(nodata=label_nodata, nodata_replace=ignore_index)
|
174 |
+
),
|
175 |
+
test=dict(
|
176 |
+
type=dataset_type,
|
177 |
+
CLASSES=CLASSES,
|
178 |
+
data_root=data_root,
|
179 |
+
img_dir=img_dir,
|
180 |
+
ann_dir=ann_dir,
|
181 |
+
img_suffix=img_suffix,
|
182 |
+
seg_map_suffix=seg_map_suffix,
|
183 |
+
pipeline=test_pipeline,
|
184 |
+
ignore_index=ignore_index,
|
185 |
+
split=splits["test"],
|
186 |
+
gt_seg_map_loader_cfg=dict(nodata=label_nodata, nodata_replace=ignore_index),
|
187 |
+
),
|
188 |
+
)
|
189 |
+
|
190 |
+
# Training
|
191 |
+
optimizer = dict(type="Adam", lr=6e-5, weight_decay=0.05)
|
192 |
+
optimizer_config = dict(grad_clip=None)
|
193 |
+
lr_config = dict(
|
194 |
+
policy="poly",
|
195 |
+
warmup="linear",
|
196 |
+
warmup_iters=1500,
|
197 |
+
warmup_ratio=1e-6,
|
198 |
+
power=1.0,
|
199 |
+
min_lr=0.0,
|
200 |
+
by_epoch=False,
|
201 |
+
)
|
202 |
+
|
203 |
+
log_config = dict(
|
204 |
+
interval=10,
|
205 |
+
hooks=[
|
206 |
+
dict(type='TextLoggerHook', by_epoch=True),
|
207 |
+
dict(type='TensorboardLoggerHook', by_epoch=True),
|
208 |
+
])
|
209 |
+
|
210 |
+
checkpoint_config = dict(
|
211 |
+
by_epoch=True, interval=10, out_dir=save_path
|
212 |
+
)
|
213 |
+
|
214 |
+
evaluation = dict(
|
215 |
+
interval=eval_epoch_interval, metric="mIoU", pre_eval=True, save_best="mIoU", by_epoch=True
|
216 |
+
)
|
217 |
+
|
218 |
+
runner = dict(type="EpochBasedRunner", max_epochs=epochs)
|
219 |
+
|
220 |
+
workflow = [("train", 1),("val", 1)]
|
221 |
+
|
222 |
+
norm_cfg = dict(type="BN", requires_grad=True)
|
223 |
+
|
224 |
+
ce_weights = [0.3, 0.7]
|
225 |
+
|
226 |
+
model = dict(
|
227 |
+
type="TemporalEncoderDecoder",
|
228 |
+
frozen_backbone=False,
|
229 |
+
backbone=dict(
|
230 |
+
type="TemporalViTEncoder",
|
231 |
+
pretrained=pretrained_weights_path,
|
232 |
+
img_size=img_size,
|
233 |
+
patch_size=patch_size,
|
234 |
+
num_frames=num_frames,
|
235 |
+
tubelet_size=1,
|
236 |
+
in_chans=len(bands),
|
237 |
+
embed_dim=embed_dim,
|
238 |
+
depth=num_layers,
|
239 |
+
num_heads=num_heads,
|
240 |
+
mlp_ratio=4.0,
|
241 |
+
norm_pix_loss=False,
|
242 |
+
),
|
243 |
+
neck=dict(
|
244 |
+
type="ConvTransformerTokensToEmbeddingNeck",
|
245 |
+
embed_dim=num_frames*embed_dim,
|
246 |
+
output_embed_dim=embed_dim,
|
247 |
+
drop_cls_token=True,
|
248 |
+
Hp=img_size // patch_size,
|
249 |
+
Wp=img_size // patch_size,
|
250 |
+
),
|
251 |
+
decode_head=dict(
|
252 |
+
num_classes=num_classes,
|
253 |
+
in_channels=embed_dim,
|
254 |
+
type="FCNHead",
|
255 |
+
in_index=-1,
|
256 |
+
ignore_index=ignore_index,
|
257 |
+
channels=256,
|
258 |
+
num_convs=1,
|
259 |
+
concat_input=False,
|
260 |
+
dropout_ratio=0.1,
|
261 |
+
norm_cfg=norm_cfg,
|
262 |
+
align_corners=False,
|
263 |
+
loss_decode=dict(
|
264 |
+
type="CrossEntropyLoss",
|
265 |
+
use_sigmoid=False,
|
266 |
+
loss_weight=1,
|
267 |
+
class_weight=ce_weights,
|
268 |
+
),
|
269 |
+
),
|
270 |
+
auxiliary_head=dict(
|
271 |
+
num_classes=num_classes,
|
272 |
+
in_channels=embed_dim,
|
273 |
+
ignore_index=ignore_index,
|
274 |
+
type="FCNHead",
|
275 |
+
in_index=-1,
|
276 |
+
channels=256,
|
277 |
+
num_convs=2,
|
278 |
+
concat_input=False,
|
279 |
+
dropout_ratio=0.1,
|
280 |
+
norm_cfg=norm_cfg,
|
281 |
+
align_corners=False,
|
282 |
+
loss_decode=dict(
|
283 |
+
type="CrossEntropyLoss",
|
284 |
+
use_sigmoid=False,
|
285 |
+
loss_weight=1,
|
286 |
+
class_weight=ce_weights,
|
287 |
+
),
|
288 |
+
),
|
289 |
+
train_cfg=dict(),
|
290 |
+
test_cfg=dict(mode="slide", stride=(int(tile_size/2), int(tile_size/2)), crop_size=(tile_size, tile_size)),
|
291 |
+
)
|