-
Notifications
You must be signed in to change notification settings - Fork 1
/
Copy pathlpcv_train.yaml
260 lines (242 loc) · 6.69 KB
/
lpcv_train.yaml
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
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
num_classes: &num_classes 14
runtime:
task_names: seg
seg_rand_resize: &seg_rand_resize
type: seg_rand_resize
kwargs:
scale: [ 0.5, 2.0 ]
seg_rand_resize_after: &seg_rand_resize_after
type: seg_rand_resize
kwargs:
scale: [ 0.8, 1.2 ]
seg_resize: &seg_resize
type: seg_resize
kwargs:
size: [ 512, 512 ]
color_jitter: &color_jitter
type: color_jitter_mmseg
kwargs:
color_type: &color_type RGB
seg_crop_train: &seg_crop_train
type: seg_crop
kwargs:
size: [ 512, 512 ]
crop_type: rand
seg_cutout: &seg_cutout
type: seg_cutout
kwargs:
n_holes: 2
length: 0.1
ignore_index: 255
seg_flip: &flip
type: seg_random_flip
seg_flip_v: &flip_v
type: seg_random_flip_vertical
seg_rand_rotate: &seg_rand_rotate
type: seg_rand_rotate_lpcv
kwargs:
angle: 40.
prob: 0.8
seg_crop_test: &seg_crop_test
type: seg_crop
kwargs:
size: [ 512, 512 ]
crop_type: center
to_tensor: &to_tensor
type: custom_to_tensor
normalize: &normalize
type: normalize
kwargs:
mean: [ 123.675, 116.28, 103.53 ] # ImageNet pretrained statics
std: [ 58.395, 57.12, 57.375 ]
dataset: # Required.
train:
dataset:
type: seg
kwargs:
meta_file: train_meta_repeat_100.txt
image_reader:
type: fs_opencv
kwargs:
image_dir: LPCVC_Train_Updated/IMG/train
color_mode: RGB
seg_label_reader:
type: fs_opencv
kwargs:
image_dir: LPCVC_Train_Updated/GT_Updated/train
color_mode: GRAY
transformer: [ *seg_rand_rotate,*seg_rand_resize, *flip, *flip_v,*seg_crop_train, *seg_cutout,*color_jitter, *to_tensor, *normalize ]
num_classes: *num_classes
ignore_label: 255
batch_sampler:
type: base
kwargs:
sampler:
type: dist
kwargs: { }
batch_size: 16
dataloader:
type: seg_base
kwargs:
num_workers: 24
pin_memory: True
test:
dataset:
type: seg_lpcv
kwargs:
meta_file: val_meta.txt
image_reader:
type: fs_opencv
kwargs:
image_dir: LPCVC_Val/IMG/val
color_mode: RGB
seg_label_reader:
type: fs_opencv
kwargs:
image_dir: LPCVC_Val/GT/val
color_mode: GRAY
transformer: [ *seg_resize, *to_tensor, *normalize ]
num_classes: *num_classes
ignore_label: 255
output_pred: True
output_gt: True
evaluator:
type: seg_with_dice # choices = {'COCO', 'VOC', 'MR'}
kwargs:
num_classes: *num_classes
cmp_key: dice
batch_sampler:
type: base
kwargs:
sampler:
type: dist
kwargs: { }
batch_size: 1
dataloader:
type: seg_base
kwargs:
num_workers: 2
pin_memory: False
trainer: # Required.
max_iter: &max_iter 20000
test_freq: 1
save_freq: 20000
only_save_latest: True
optimizer:
register_type: segformer
type: AdamW
kwargs:
lr: 4.0e-6
betas: !!python/tuple [ 0.9, 0.999 ]
weight_decay: 0.01
special_param_group: [ { 'key': 'decoder', 'lr': 4.0e-5, 'weight_decay': 0.01 },
{ 'key': 'norm', 'lr': 4.0e-6, 'weight_decay': 0.0 },
{ 'key': 'bn', 'lr': 4.0e-6, 'weight_decay': 0.0 } ]
lr_scheduler:
warmup_iter: 1500 # 1000 iterations of warmup
warmup_type: linear
warmup_ratio: 0.000001
type: polylr
kwargs:
power: 1.0
max_iter: *max_iter
ema:
enable: True
ema_type: exp
kwargs:
decay: 0.9998
saver: # Required.
save_dir: checkpoints/LPCV_2023_Seg # dir to save checkpoints
results_dir: results_dir/LPCV_2023_Seg # dir to save detection results. i.e., bboxes, masks, keypoints
pretrain_model: models/pretrain_weight.pth
auto_resume: True # find last checkpoint from save_dir and resume from it automatically
# this option has the highest priority (auto_resume > opts > resume_model > pretrain_model)
hooks:
- type: auto_save_best
- type: yolox_noaug
kwargs:
no_aug_epoch: 3
test_freq: 1
save_freq: 999999
max_epoch: 25
transformer: [ *seg_rand_resize_after, *flip, *seg_crop_train, *to_tensor, *normalize ]
net: &subnet
- name: backbone # backbone = resnet50(frozen_layers, out_layers, out_strides)
type: LPCV_2023_Seg_Backbone
kwargs:
out_layers: [ 0,1,2,3,4 ]
out_strides: [ 4,8,16,32,64 ]
frozen_layers: [ ]
more_act: True
force_connect: False
input_resize: True
input_resize_ratio: 0.5
# input_resize_mode: bilinear
# input_resize_align_corners: True
width: [ [ 3, 16, 16, ],
[ 16, 32, 32 ],
[ 32, 64, 64, ],
[ 64, 128, 128, ],
[ 128, 192, 192, 192, 192, ],
[ 192, 256, 256, 256, 256, 256, ],
]
expand_ratio: [ [ 1, 1, ],
[ 1, 1, ],
[ 1, 1, ],
[ 1, 1, ],
[ 1, 1, 1, 1, ],
[ 1, 1, 1, 1, 1, ], ]
dbb: True
dbb_mid_expand_factor: 1.
dropout_prob: 0.05
dense_connect: True
normalize:
type: sync_bn
kwargs:
group_size: 8
initializer:
method: xavier
- name: neck
prev: backbone
type: LPCV_2023_Seg_Neck # up.tasks.det.models.necks.fpn.FPN
kwargs:
num_repeat: 1
outplanes: -1
align_channels: False
start_level: 3
num_level: 5 # if num_level>len(backbone.out_layers), additional conv with be stacked.
out_strides: [ 4,8,16,32,64 ] # strides of output features. aka., anchor strides for roi_head
downsample: conv # method to downsample, for FPN, it's pool, for RetienaNet, it's conv
upsample: deconv # method to interp, nearest or bilinear
split_deconv: True
padding: [ [ 1, 1 ], [ 1, 1 ], [ 1, 1 ], [ 1, 1 ], [ 1, 1 ] ]
output_padding: [ [ 0, 0 ], [ 0, 0 ], [ 0, 0 ], [ 0, 0 ], [ 0, 0 ] ]
deconv_kernel: [ 4, 4, 4, 4, 4 ]
use_dbb: True
dbb_ratio: 1.
normalize:
type: sync_bn
kwargs:
group_size: 8
initializer:
method: xavier
- name: decoder
prev: neck
type: LPCV_2023_Seg_Decoder
kwargs:
num_classes: *num_classes
head_planes: 16
simple: False
antialias: False
upsample_mode: bicubic
normalize:
type: sync_bn
kwargs:
group_size: 8
loss:
type: seg_ohem
kwargs:
aux_weight: 0.4
thresh: 0.7
min_kept: 40000
ignore_index: 255