Skip to content

Commit 8885a81

Browse files
yangze0930yangze0930
committed
add reppoints head, reppoints detector and the config files.
1 parent c0a4c00 commit 8885a81

17 files changed

+2574
-2
lines changed

‎configs/reppoints/bbox_r50_grid_center_fpn_1x.py‎

Lines changed: 143 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,143 @@
1+
# model settings
2+
norm_cfg = dict(type='GN', num_groups=32, requires_grad=True)
3+
4+
model = dict(
5+
type='RepPointsDetector',
6+
pretrained='torchvision://resnet50',
7+
backbone=dict(
8+
type='ResNet',
9+
depth=50,
10+
num_stages=4,
11+
out_indices=(0, 1, 2, 3),
12+
frozen_stages=1,
13+
style='pytorch'),
14+
neck=dict(
15+
type='FPN',
16+
in_channels=[256, 512, 1024, 2048],
17+
out_channels=256,
18+
start_level=1,
19+
add_extra_convs=True,
20+
num_outs=5,
21+
norm_cfg=norm_cfg),
22+
bbox_head=dict(
23+
type='RepPointsHead',
24+
num_classes=81,
25+
in_channels=256,
26+
feat_channels=256,
27+
point_feat_channels=256,
28+
stacked_convs=3,
29+
num_points=9,
30+
gradient_mul=0.1,
31+
point_strides=[8, 16, 32, 64, 128],
32+
point_base_scale=4,
33+
norm_cfg=norm_cfg,
34+
loss_cls=dict(
35+
type='FocalLoss',
36+
use_sigmoid=True,
37+
gamma=2.0,
38+
alpha=0.25,
39+
loss_weight=1.0),
40+
loss_bbox_init=dict(type='SmoothL1Loss', beta=0.11, loss_weight=0.5),
41+
loss_bbox_refine=dict(type='SmoothL1Loss', beta=0.11, loss_weight=1.0),
42+
transform_method='minmax',
43+
use_grid_points=True))
44+
# training and testing settings
45+
train_cfg = dict(
46+
init=dict(
47+
assigner=dict(type='PointAssigner', scale=4, pos_num=1),
48+
allowed_border=-1,
49+
pos_weight=-1,
50+
debug=False),
51+
refine=dict(
52+
assigner=dict(
53+
type='MaxIoUAssigner',
54+
pos_iou_thr=0.5,
55+
neg_iou_thr=0.4,
56+
min_pos_iou=0,
57+
ignore_iof_thr=-1),
58+
allowed_border=-1,
59+
pos_weight=-1,
60+
debug=False))
61+
test_cfg = dict(
62+
nms_pre=1000,
63+
min_bbox_size=0,
64+
score_thr=0.05,
65+
nms=dict(type='nms', iou_thr=0.5),
66+
max_per_img=100)
67+
# dataset settings
68+
dataset_type = 'CocoDataset'
69+
data_root = 'data/coco/'
70+
img_norm_cfg = dict(
71+
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
72+
train_pipeline = [
73+
dict(type='LoadImageFromFile'),
74+
dict(type='LoadAnnotations', with_bbox=True),
75+
dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
76+
dict(type='RandomFlip', flip_ratio=0.5),
77+
dict(type='Normalize', **img_norm_cfg),
78+
dict(type='Pad', size_divisor=32),
79+
dict(type='DefaultFormatBundle'),
80+
dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
81+
]
82+
test_pipeline = [
83+
dict(type='LoadImageFromFile'),
84+
dict(
85+
type='MultiScaleFlipAug',
86+
img_scale=(1333, 800),
87+
flip=False,
88+
transforms=[
89+
dict(type='Resize', keep_ratio=True),
90+
dict(type='RandomFlip'),
91+
dict(type='Normalize', **img_norm_cfg),
92+
dict(type='Pad', size_divisor=32),
93+
dict(type='ImageToTensor', keys=['img']),
94+
dict(type='Collect', keys=['img']),
95+
])
96+
]
97+
data = dict(
98+
imgs_per_gpu=2,
99+
workers_per_gpu=2,
100+
train=dict(
101+
type=dataset_type,
102+
ann_file=data_root + 'annotations/instances_train2017.json',
103+
img_prefix=data_root + 'train2017/',
104+
pipeline=train_pipeline),
105+
val=dict(
106+
type=dataset_type,
107+
ann_file=data_root + 'annotations/instances_val2017.json',
108+
img_prefix=data_root + 'val2017/',
109+
pipeline=test_pipeline),
110+
test=dict(
111+
type=dataset_type,
112+
ann_file=data_root + 'annotations/instances_val2017.json',
113+
img_prefix=data_root + 'val2017/',
114+
pipeline=test_pipeline))
115+
# optimizer
116+
optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)
117+
optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
118+
# learning policy
119+
lr_config = dict(
120+
policy='step',
121+
warmup='linear',
122+
warmup_iters=500,
123+
warmup_ratio=1.0 / 3,
124+
step=[8, 11])
125+
checkpoint_config = dict(interval=1)
126+
# yapf:disable
127+
log_config = dict(
128+
interval=50,
129+
hooks=[
130+
dict(type='TextLoggerHook'),
131+
# dict(type='TensorboardLoggerHook')
132+
])
133+
# yapf:enable
134+
# runtime settings
135+
total_epochs = 12
136+
device_ids = range(8)
137+
dist_params = dict(backend='nccl')
138+
log_level = 'INFO'
139+
work_dir = './work_dirs/bbox_r50_grid_center_fpn_1x'
140+
load_from = None
141+
resume_from = None
142+
auto_resume = True
143+
workflow = [('train', 1)]

‎configs/reppoints/bbox_r50_grid_fpn_1x.py‎

Lines changed: 148 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,148 @@
1+
# model settings
2+
norm_cfg = dict(type='GN', num_groups=32, requires_grad=True)
3+
4+
model = dict(
5+
type='RepPointsDetector',
6+
pretrained='torchvision://resnet50',
7+
backbone=dict(
8+
type='ResNet',
9+
depth=50,
10+
num_stages=4,
11+
out_indices=(0, 1, 2, 3),
12+
frozen_stages=1,
13+
style='pytorch'),
14+
neck=dict(
15+
type='FPN',
16+
in_channels=[256, 512, 1024, 2048],
17+
out_channels=256,
18+
start_level=1,
19+
add_extra_convs=True,
20+
num_outs=5,
21+
norm_cfg=norm_cfg),
22+
bbox_head=dict(
23+
type='RepPointsHead',
24+
num_classes=81,
25+
in_channels=256,
26+
feat_channels=256,
27+
point_feat_channels=256,
28+
stacked_convs=3,
29+
num_points=9,
30+
gradient_mul=0.1,
31+
point_strides=[8, 16, 32, 64, 128],
32+
point_base_scale=4,
33+
norm_cfg=norm_cfg,
34+
loss_cls=dict(
35+
type='FocalLoss',
36+
use_sigmoid=True,
37+
gamma=2.0,
38+
alpha=0.25,
39+
loss_weight=1.0),
40+
loss_bbox_init=dict(type='SmoothL1Loss', beta=0.11, loss_weight=0.5),
41+
loss_bbox_refine=dict(type='SmoothL1Loss', beta=0.11, loss_weight=1.0),
42+
transform_method='minmax',
43+
use_grid_points=True))
44+
# training and testing settings
45+
train_cfg = dict(
46+
init=dict(
47+
assigner=dict(
48+
type='MaxIoUAssigner',
49+
pos_iou_thr=0.5,
50+
neg_iou_thr=0.4,
51+
min_pos_iou=0,
52+
ignore_iof_thr=-1),
53+
allowed_border=-1,
54+
pos_weight=-1,
55+
debug=False),
56+
refine=dict(
57+
assigner=dict(
58+
type='MaxIoUAssigner',
59+
pos_iou_thr=0.5,
60+
neg_iou_thr=0.4,
61+
min_pos_iou=0,
62+
ignore_iof_thr=-1),
63+
allowed_border=-1,
64+
pos_weight=-1,
65+
debug=False))
66+
test_cfg = dict(
67+
nms_pre=1000,
68+
min_bbox_size=0,
69+
score_thr=0.05,
70+
nms=dict(type='nms', iou_thr=0.5),
71+
max_per_img=100)
72+
# dataset settings
73+
dataset_type = 'CocoDataset'
74+
data_root = 'data/coco/'
75+
img_norm_cfg = dict(
76+
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
77+
train_pipeline = [
78+
dict(type='LoadImageFromFile'),
79+
dict(type='LoadAnnotations', with_bbox=True),
80+
dict(type='Resize', img_scale=(1333, 800), keep_ratio=True),
81+
dict(type='RandomFlip', flip_ratio=0.5),
82+
dict(type='Normalize', **img_norm_cfg),
83+
dict(type='Pad', size_divisor=32),
84+
dict(type='DefaultFormatBundle'),
85+
dict(type='Collect', keys=['img', 'gt_bboxes', 'gt_labels']),
86+
]
87+
test_pipeline = [
88+
dict(type='LoadImageFromFile'),
89+
dict(
90+
type='MultiScaleFlipAug',
91+
img_scale=(1333, 800),
92+
flip=False,
93+
transforms=[
94+
dict(type='Resize', keep_ratio=True),
95+
dict(type='RandomFlip'),
96+
dict(type='Normalize', **img_norm_cfg),
97+
dict(type='Pad', size_divisor=32),
98+
dict(type='ImageToTensor', keys=['img']),
99+
dict(type='Collect', keys=['img']),
100+
])
101+
]
102+
data = dict(
103+
imgs_per_gpu=2,
104+
workers_per_gpu=2,
105+
train=dict(
106+
type=dataset_type,
107+
ann_file=data_root + 'annotations/instances_train2017.json',
108+
img_prefix=data_root + 'train2017/',
109+
pipeline=train_pipeline),
110+
val=dict(
111+
type=dataset_type,
112+
ann_file=data_root + 'annotations/instances_val2017.json',
113+
img_prefix=data_root + 'val2017/',
114+
pipeline=test_pipeline),
115+
test=dict(
116+
type=dataset_type,
117+
ann_file=data_root + 'annotations/instances_val2017.json',
118+
img_prefix=data_root + 'val2017/',
119+
pipeline=test_pipeline))
120+
# optimizer
121+
optimizer = dict(type='SGD', lr=0.01, momentum=0.9, weight_decay=0.0001)
122+
optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
123+
# learning policy
124+
lr_config = dict(
125+
policy='step',
126+
warmup='linear',
127+
warmup_iters=500,
128+
warmup_ratio=1.0 / 3,
129+
step=[8, 11])
130+
checkpoint_config = dict(interval=1)
131+
# yapf:disable
132+
log_config = dict(
133+
interval=50,
134+
hooks=[
135+
dict(type='TextLoggerHook'),
136+
# dict(type='TensorboardLoggerHook')
137+
])
138+
# yapf:enable
139+
# runtime settings
140+
total_epochs = 12
141+
device_ids = range(8)
142+
dist_params = dict(backend='nccl')
143+
log_level = 'INFO'
144+
work_dir = './work_dirs/bbox_r50_grid_fpn_1x'
145+
load_from = None
146+
resume_from = None
147+
auto_resume = True
148+
workflow = [('train', 1)]

0 commit comments

Comments
 (0)