File size: 7,089 Bytes
f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 95b8d0f 5769604 f95bf59 5769604 f95bf59 5769604 95b8d0f 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 f95bf59 5769604 |
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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 |
import os
dist_params = dict(backend='nccl')
log_level = 'INFO'
load_from = None
resume_from = None
cudnn_benchmark = True
custom_imports = dict(imports=['geospatial_fm'])
num_frames = 3
img_size = 224
num_workers = 2
# model
# TO BE DEFINED BY USER: model path
pretrained_weights_path = '<path to pretrained weights>'
num_layers = 6
patch_size = 16
embed_dim = 768
num_heads = 8
tubelet_size = 1
max_epochs = 80
eval_epoch_interval = 5
loss_weights_multi = [
0.386375, 0.661126, 0.548184, 0.640482, 0.876862, 0.925186, 3.249462,
1.542289, 2.175141, 2.272419, 3.062762, 3.626097, 1.198702
]
loss_func = dict(
type='CrossEntropyLoss',
use_sigmoid=False,
class_weight=loss_weights_multi,
avg_non_ignore=True)
output_embed_dim = embed_dim*num_frames
# TO BE DEFINED BY USER: Save directory
experiment = '<experiment name>'
project_dir = '<project directory name>'
work_dir = os.path.join(project_dir, experiment)
save_path = work_dir
gpu_ids = range(0, 1)
dataset_type = 'GeospatialDataset'
# TO BE DEFINED BY USER: data directory
data_root = '<path to data root>'
splits = dict(
train='<path to train split>',
val= '<path to val split>',
test= '<path to test split>'
)
img_norm_cfg = dict(
means=[
494.905781, 815.239594, 924.335066, 2968.881459, 2634.621962,
1739.579917, 494.905781, 815.239594, 924.335066, 2968.881459,
2634.621962, 1739.579917, 494.905781, 815.239594, 924.335066,
2968.881459, 2634.621962, 1739.579917
],
stds=[
284.925432, 357.84876, 575.566823, 896.601013, 951.900334, 921.407808,
284.925432, 357.84876, 575.566823, 896.601013, 951.900334, 921.407808,
284.925432, 357.84876, 575.566823, 896.601013, 951.900334, 921.407808
])
bands = [0, 1, 2, 3, 4, 5]
tile_size = 224
orig_nsize = 512
crop_size = (tile_size, tile_size)
train_pipeline = [
dict(type='LoadGeospatialImageFromFile', to_float32=True),
dict(type='LoadGeospatialAnnotations', reduce_zero_label=True),
dict(type='RandomFlip', prob=0.5),
dict(type='ToTensor', keys=['img', 'gt_semantic_seg']),
# to channels first
dict(type="TorchPermute", keys=["img"], order=(2, 0, 1)),
dict(type='TorchNormalize', **img_norm_cfg),
dict(type='TorchRandomCrop', crop_size=crop_size),
dict(type='Reshape', keys=['img'], new_shape=(len(bands), num_frames, tile_size, tile_size)),
dict(type='Reshape', keys=['gt_semantic_seg'], new_shape=(1, tile_size, tile_size)),
dict(type='CastTensor', keys=['gt_semantic_seg'], new_type="torch.LongTensor"),
dict(type='Collect', keys=['img', 'gt_semantic_seg']),
]
test_pipeline = [
dict(type='LoadGeospatialImageFromFile', to_float32=True),
dict(type='ToTensor', keys=['img']),
# to channels first
dict(type="TorchPermute", keys=["img"], order=(2, 0, 1)),
dict(type='TorchNormalize', **img_norm_cfg),
dict(type='Reshape', keys=['img'], new_shape=(len(bands), num_frames, -1, -1), look_up = {'2': 1, '3': 2}),
dict(type='CastTensor', keys=['img'], new_type="torch.FloatTensor"),
dict(type='CollectTestList', keys=['img'],
meta_keys=['img_info', 'seg_fields', 'img_prefix', 'seg_prefix', 'filename', 'ori_filename', 'img',
'img_shape', 'ori_shape', 'pad_shape', 'scale_factor', 'img_norm_cfg']),
]
CLASSES = ('Natural Vegetation',
'Forest',
'Corn',
'Soybeans',
'Wetlands',
'Developed/Barren',
'Open Water',
'Winter Wheat',
'Alfalfa',
'Fallow/Idle Cropland',
'Cotton',
'Sorghum',
'Other')
dataset = 'GeospatialDataset'
data = dict(
samples_per_gpu=8,
workers_per_gpu=4,
train=dict(
type=dataset,
CLASSES=CLASSES,
reduce_zero_label=True,
data_root=data_root,
img_dir='training_chips',
ann_dir='training_chips',
pipeline=train_pipeline,
img_suffix='_merged.tif',
seg_map_suffix='.mask.tif',
split=splits['train']),
val=dict(
type=dataset,
CLASSES=CLASSES,
reduce_zero_label=True,
data_root=data_root,
img_dir='validation_chips',
ann_dir='validation_chips',
pipeline=test_pipeline,
img_suffix='_merged.tif',
seg_map_suffix='.mask.tif',
split=splits['val']
),
test=dict(
type=dataset,
CLASSES=CLASSES,
reduce_zero_label=True,
data_root=data_root,
img_dir='validation_chips',
ann_dir='validation_chips',
pipeline=test_pipeline,
img_suffix='_merged.tif',
seg_map_suffix='.mask.tif',
split=splits['val']
))
optimizer = dict(
type='Adam', lr=1.5e-05, betas=(0.9, 0.999), weight_decay=0.05)
optimizer_config = dict(grad_clip=None)
lr_config = dict(
policy='poly',
warmup='linear',
warmup_iters=1500,
warmup_ratio=1e-06,
power=1.0,
min_lr=0.0,
by_epoch=False)
log_config = dict(
interval=10,
hooks=[dict(type='TextLoggerHook'),
dict(type='TensorboardLoggerHook')])
checkpoint_config = dict(
by_epoch=True,
interval=100,
out_dir=save_path)
evaluation = dict(interval=eval_epoch_interval, metric='mIoU', pre_eval=True, save_best='mIoU', by_epoch=True)
reduce_train_set = dict(reduce_train_set=False)
reduce_factor = dict(reduce_factor=1)
runner = dict(type='EpochBasedRunner', max_epochs=max_epochs)
workflow = [('train', 1)]
norm_cfg = dict(type='BN', requires_grad=True)
model = dict(
type='TemporalEncoderDecoder',
frozen_backbone=False,
backbone=dict(
type='TemporalViTEncoder',
pretrained=pretrained_weights_path,
img_size=img_size,
patch_size=patch_size,
num_frames=num_frames,
tubelet_size=1,
in_chans=len(bands),
embed_dim=embed_dim,
depth=6,
num_heads=num_heads,
mlp_ratio=4.0,
norm_pix_loss=False),
neck=dict(
type='ConvTransformerTokensToEmbeddingNeck',
embed_dim=embed_dim*num_frames,
output_embed_dim=output_embed_dim,
drop_cls_token=True,
Hp=14,
Wp=14),
decode_head=dict(
num_classes=len(CLASSES),
in_channels=output_embed_dim,
type='FCNHead',
in_index=-1,
channels=256,
num_convs=1,
concat_input=False,
dropout_ratio=0.1,
norm_cfg=dict(type='BN', requires_grad=True),
align_corners=False,
loss_decode=loss_func),
auxiliary_head=dict(
num_classes=len(CLASSES),
in_channels=output_embed_dim,
type='FCNHead',
in_index=-1,
channels=256,
num_convs=2,
concat_input=False,
dropout_ratio=0.1,
norm_cfg=dict(type='BN', requires_grad=True),
align_corners=False,
loss_decode=loss_func),
train_cfg=dict(),
test_cfg=dict(mode='slide', stride=(int(tile_size/2), int(tile_size/2)), crop_size=(tile_size, tile_size)))
auto_resume = False
|