Skip to content

Commit d841f75

Browse files
czczupczczup
committed
update
1 parent 8375314 commit d841f75

File tree

6 files changed

+490
-0
lines changed

6 files changed

+490
-0
lines changed

‎segmentation/configs/coco_stuff10k/README.md‎

Lines changed: 14 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,14 @@
1+
# COCO-Stuff-10K
2+
3+
<!-- [ALGORITHM] -->
4+
5+
## Introduction
6+
7+
The Common Objects in COntext-stuff (COCO-stuff) dataset is a dataset for scene understanding tasks like semantic segmentation, object detection and image captioning. It is constructed by annotating the original COCO dataset, which originally annotated things while neglecting stuff annotations. There are 10k images in COCO-Stuff-10K dataset that span over 172 categories including 80 things, 91 stuff, and 1 unlabeled class. It is split into 9,000 and 1,000 images for training and testing.
8+
9+
## Results and Models
10+
11+
| Method | Backbone | Pre-train | Batch Size | Lr schd | Crop Size | mIoU (SS) | mIoU (MS) | #Param | Config | Download |
12+
|:-----------:|:-------------:|:---------------------------------------------------------------------------------------------------------------------:|:----------:|:-------:|:---------:|:------------------------------------------------------------------------------------------:|:------------------------------------------------------------------------------------------:|:------:|:---------------------------------------------------------------------:|:----------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------------:|
13+
| UperNet | ViT-Adapter-L | [BEiT-L](https://conversationhub.blob.core.windows.net/beit-share-public/beit/beit_large_patch16_224_pt22k_ft22k.pth) | 8x2 | 80k | 512 | [51.0](https://drive.google.com/file/d/1xZodiAvOLGaLtMGx_btYVZIMC2VKrDhI/view?usp=sharing) | [51.4](https://drive.google.com/file/d/1bmFG9GA4bRqOEJfqXcO7nWYPwG3wSk2J/view?usp=sharing) | 451M | [config](./upernet_beit_adapter_large_512_80k_cocostuff10k_ss.py) | [model](https://github.com/czczup/ViT-Adapter/releases/download/v0.2.4/upernet_beit_adapter_large_512_80k_cocostuff10k.pth.tar) \| [log](https://github.com/czczup/ViT-Adapter/releases/download/v0.2.4/20220505_091358.log) |
14+
| Mask2Former | ViT-Adapter-L | [BEiT-L](https://conversationhub.blob.core.windows.net/beit-share-public/beit/beit_large_patch16_224_pt22k_ft22k.pth) | 8x2 | 40k | 512 | [53.2](https://drive.google.com/file/d/1Buewc1n7GBAcBDXeia-QarujrDZqc_Sx/view?usp=sharing) | [54.2](https://drive.google.com/file/d/1kQgJUHDeQoO3pPY6QoXRKwyF7heT7wCJ/view?usp=sharing) | 568M | [config](./mask2former_beit_adapter_large_512_40k_cocostuff10k_ss.py) | [model]() \| [log]() |

‎segmentation/configs/coco_stuff10k/mask2former_beit_adapter_large_512_40k_cocostuff10k_ms.py‎

Lines changed: 149 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,149 @@
1+
# Copyright (c) Shanghai AI Lab. All rights reserved.
2+
_base_ = [
3+
'../_base_/models/mask2former_beit_cocostuff.py',
4+
'../_base_/datasets/coco-stuff10k.py',
5+
'../_base_/default_runtime.py',
6+
'../_base_/schedules/schedule_40k.py'
7+
]
8+
crop_size = (512, 512)
9+
# pretrained = 'https://conversationhub.blob.core.windows.net/beit-share-public/beit/beit_large_patch16_224_pt22k_ft22k.pth'
10+
pretrained = 'pretrained/beit_large_patch16_224_pt22k_ft22k.pth'
11+
model = dict(
12+
pretrained=pretrained,
13+
backbone=dict(
14+
type='BEiTAdapter',
15+
img_size=512,
16+
patch_size=16,
17+
embed_dim=1024,
18+
depth=24,
19+
num_heads=16,
20+
mlp_ratio=4,
21+
qkv_bias=True,
22+
use_abs_pos_emb=False,
23+
use_rel_pos_bias=True,
24+
init_values=1e-6,
25+
drop_path_rate=0.3,
26+
conv_inplane=64,
27+
n_points=4,
28+
deform_num_heads=16,
29+
cffn_ratio=0.25,
30+
deform_ratio=0.5,
31+
interaction_indexes=[[0, 5], [6, 11], [12, 17], [18, 23]],
32+
),
33+
decode_head=dict(
34+
in_channels=[1024, 1024, 1024, 1024],
35+
feat_channels=1024,
36+
out_channels=1024,
37+
num_queries=100,
38+
pixel_decoder=dict(
39+
type='MSDeformAttnPixelDecoder',
40+
num_outs=3,
41+
norm_cfg=dict(type='GN', num_groups=32),
42+
act_cfg=dict(type='ReLU'),
43+
encoder=dict(
44+
type='DetrTransformerEncoder',
45+
num_layers=6,
46+
transformerlayers=dict(
47+
type='BaseTransformerLayer',
48+
attn_cfgs=dict(
49+
type='MultiScaleDeformableAttention',
50+
embed_dims=1024,
51+
num_heads=32,
52+
num_levels=3,
53+
num_points=4,
54+
im2col_step=64,
55+
dropout=0.0,
56+
batch_first=False,
57+
norm_cfg=None,
58+
init_cfg=None),
59+
ffn_cfgs=dict(
60+
type='FFN',
61+
embed_dims=1024,
62+
feedforward_channels=4096,
63+
num_fcs=2,
64+
ffn_drop=0.0,
65+
act_cfg=dict(type='ReLU', inplace=True)),
66+
operation_order=('self_attn', 'norm', 'ffn', 'norm')),
67+
init_cfg=None),
68+
positional_encoding=dict(
69+
type='SinePositionalEncoding', num_feats=512, normalize=True),
70+
init_cfg=None),
71+
positional_encoding=dict(
72+
type='SinePositionalEncoding', num_feats=512, normalize=True),
73+
transformer_decoder=dict(
74+
type='DetrTransformerDecoder',
75+
return_intermediate=True,
76+
num_layers=9,
77+
transformerlayers=dict(
78+
type='DetrTransformerDecoderLayer',
79+
attn_cfgs=dict(
80+
type='MultiheadAttention',
81+
embed_dims=1024,
82+
num_heads=32,
83+
attn_drop=0.0,
84+
proj_drop=0.0,
85+
dropout_layer=None,
86+
batch_first=False),
87+
ffn_cfgs=dict(
88+
embed_dims=1024,
89+
feedforward_channels=4096,
90+
num_fcs=2,
91+
act_cfg=dict(type='ReLU', inplace=True),
92+
ffn_drop=0.0,
93+
dropout_layer=None,
94+
add_identity=True),
95+
feedforward_channels=4096,
96+
operation_order=('cross_attn', 'norm', 'self_attn', 'norm',
97+
'ffn', 'norm')),
98+
init_cfg=None)
99+
),
100+
test_cfg=dict(mode='slide', crop_size=crop_size, stride=(341, 341))
101+
)
102+
# dataset settings
103+
img_norm_cfg = dict(
104+
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
105+
train_pipeline = [
106+
dict(type='LoadImageFromFile'),
107+
dict(type='LoadAnnotations', reduce_zero_label=True),
108+
dict(type='Resize', img_scale=(2048, 512), ratio_range=(0.5, 2.0)),
109+
dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
110+
dict(type='RandomFlip', prob=0.5),
111+
dict(type='PhotoMetricDistortion'),
112+
dict(type='Normalize', **img_norm_cfg),
113+
dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
114+
dict(type='ToMask'),
115+
dict(type='DefaultFormatBundle'),
116+
dict(type='Collect', keys=['img', 'gt_semantic_seg', 'gt_masks', 'gt_labels'])
117+
]
118+
test_pipeline = [
119+
dict(type='LoadImageFromFile'),
120+
dict(
121+
type='MultiScaleFlipAug',
122+
img_scale=(2048, 512),
123+
img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
124+
flip=True,
125+
transforms=[
126+
dict(type='SETR_Resize', keep_ratio=True,
127+
crop_size=crop_size, setr_multi_scale=True),
128+
dict(type='RandomFlip'),
129+
dict(type='Normalize', **img_norm_cfg),
130+
dict(type='ImageToTensor', keys=['img']),
131+
dict(type='Collect', keys=['img']),
132+
])
133+
]
134+
optimizer = dict(_delete_=True, type='AdamW', lr=2e-5, betas=(0.9, 0.999), weight_decay=0.05,
135+
constructor='LayerDecayOptimizerConstructor',
136+
paramwise_cfg=dict(num_layers=24, layer_decay_rate=0.90))
137+
lr_config = dict(_delete_=True,
138+
policy='poly',
139+
warmup='linear',
140+
warmup_iters=1500,
141+
warmup_ratio=1e-6,
142+
power=1.0, min_lr=0.0, by_epoch=False)
143+
data = dict(samples_per_gpu=2,
144+
train=dict(pipeline=train_pipeline),
145+
val=dict(pipeline=test_pipeline),
146+
test=dict(pipeline=test_pipeline))
147+
runner = dict(type='IterBasedRunner')
148+
checkpoint_config = dict(by_epoch=False, interval=1000, max_keep_ckpts=1)
149+
evaluation = dict(interval=4000, metric='mIoU', save_best='mIoU')

‎segmentation/configs/coco_stuff10k/mask2former_beit_adapter_large_512_40k_cocostuff10k_ss.py‎

Lines changed: 149 additions & 0 deletions
Original file line numberDiff line numberDiff line change
@@ -0,0 +1,149 @@
1+
# Copyright (c) Shanghai AI Lab. All rights reserved.
2+
_base_ = [
3+
'../_base_/models/mask2former_beit_cocostuff.py',
4+
'../_base_/datasets/coco-stuff10k.py',
5+
'../_base_/default_runtime.py',
6+
'../_base_/schedules/schedule_40k.py'
7+
]
8+
crop_size = (512, 512)
9+
# pretrained = 'https://conversationhub.blob.core.windows.net/beit-share-public/beit/beit_large_patch16_224_pt22k_ft22k.pth'
10+
pretrained = 'pretrained/beit_large_patch16_224_pt22k_ft22k.pth'
11+
model = dict(
12+
pretrained=pretrained,
13+
backbone=dict(
14+
type='BEiTAdapter',
15+
img_size=512,
16+
patch_size=16,
17+
embed_dim=1024,
18+
depth=24,
19+
num_heads=16,
20+
mlp_ratio=4,
21+
qkv_bias=True,
22+
use_abs_pos_emb=False,
23+
use_rel_pos_bias=True,
24+
init_values=1e-6,
25+
drop_path_rate=0.3,
26+
conv_inplane=64,
27+
n_points=4,
28+
deform_num_heads=16,
29+
cffn_ratio=0.25,
30+
deform_ratio=0.5,
31+
interaction_indexes=[[0, 5], [6, 11], [12, 17], [18, 23]],
32+
),
33+
decode_head=dict(
34+
in_channels=[1024, 1024, 1024, 1024],
35+
feat_channels=1024,
36+
out_channels=1024,
37+
num_queries=100,
38+
pixel_decoder=dict(
39+
type='MSDeformAttnPixelDecoder',
40+
num_outs=3,
41+
norm_cfg=dict(type='GN', num_groups=32),
42+
act_cfg=dict(type='ReLU'),
43+
encoder=dict(
44+
type='DetrTransformerEncoder',
45+
num_layers=6,
46+
transformerlayers=dict(
47+
type='BaseTransformerLayer',
48+
attn_cfgs=dict(
49+
type='MultiScaleDeformableAttention',
50+
embed_dims=1024,
51+
num_heads=32,
52+
num_levels=3,
53+
num_points=4,
54+
im2col_step=64,
55+
dropout=0.0,
56+
batch_first=False,
57+
norm_cfg=None,
58+
init_cfg=None),
59+
ffn_cfgs=dict(
60+
type='FFN',
61+
embed_dims=1024,
62+
feedforward_channels=4096,
63+
num_fcs=2,
64+
ffn_drop=0.0,
65+
act_cfg=dict(type='ReLU', inplace=True)),
66+
operation_order=('self_attn', 'norm', 'ffn', 'norm')),
67+
init_cfg=None),
68+
positional_encoding=dict(
69+
type='SinePositionalEncoding', num_feats=512, normalize=True),
70+
init_cfg=None),
71+
positional_encoding=dict(
72+
type='SinePositionalEncoding', num_feats=512, normalize=True),
73+
transformer_decoder=dict(
74+
type='DetrTransformerDecoder',
75+
return_intermediate=True,
76+
num_layers=9,
77+
transformerlayers=dict(
78+
type='DetrTransformerDecoderLayer',
79+
attn_cfgs=dict(
80+
type='MultiheadAttention',
81+
embed_dims=1024,
82+
num_heads=32,
83+
attn_drop=0.0,
84+
proj_drop=0.0,
85+
dropout_layer=None,
86+
batch_first=False),
87+
ffn_cfgs=dict(
88+
embed_dims=1024,
89+
feedforward_channels=4096,
90+
num_fcs=2,
91+
act_cfg=dict(type='ReLU', inplace=True),
92+
ffn_drop=0.0,
93+
dropout_layer=None,
94+
add_identity=True),
95+
feedforward_channels=4096,
96+
operation_order=('cross_attn', 'norm', 'self_attn', 'norm',
97+
'ffn', 'norm')),
98+
init_cfg=None)
99+
),
100+
test_cfg=dict(mode='slide', crop_size=crop_size, stride=(341, 341))
101+
)
102+
# dataset settings
103+
img_norm_cfg = dict(
104+
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
105+
train_pipeline = [
106+
dict(type='LoadImageFromFile'),
107+
dict(type='LoadAnnotations', reduce_zero_label=True),
108+
dict(type='Resize', img_scale=(2048, 512), ratio_range=(0.5, 2.0)),
109+
dict(type='RandomCrop', crop_size=crop_size, cat_max_ratio=0.75),
110+
dict(type='RandomFlip', prob=0.5),
111+
dict(type='PhotoMetricDistortion'),
112+
dict(type='Normalize', **img_norm_cfg),
113+
dict(type='Pad', size=crop_size, pad_val=0, seg_pad_val=255),
114+
dict(type='ToMask'),
115+
dict(type='DefaultFormatBundle'),
116+
dict(type='Collect', keys=['img', 'gt_semantic_seg', 'gt_masks', 'gt_labels'])
117+
]
118+
test_pipeline = [
119+
dict(type='LoadImageFromFile'),
120+
dict(
121+
type='MultiScaleFlipAug',
122+
img_scale=(2048, 512),
123+
# img_ratios=[0.5, 0.75, 1.0, 1.25, 1.5, 1.75],
124+
flip=False,
125+
transforms=[
126+
dict(type='Resize', keep_ratio=True),
127+
dict(type='ResizeToMultiple', size_divisor=32),
128+
dict(type='RandomFlip'),
129+
dict(type='Normalize', **img_norm_cfg),
130+
dict(type='ImageToTensor', keys=['img']),
131+
dict(type='Collect', keys=['img']),
132+
])
133+
]
134+
optimizer = dict(_delete_=True, type='AdamW', lr=2e-5, betas=(0.9, 0.999), weight_decay=0.05,
135+
constructor='LayerDecayOptimizerConstructor',
136+
paramwise_cfg=dict(num_layers=24, layer_decay_rate=0.90))
137+
lr_config = dict(_delete_=True,
138+
policy='poly',
139+
warmup='linear',
140+
warmup_iters=1500,
141+
warmup_ratio=1e-6,
142+
power=1.0, min_lr=0.0, by_epoch=False)
143+
data = dict(samples_per_gpu=2,
144+
train=dict(pipeline=train_pipeline),
145+
val=dict(pipeline=test_pipeline),
146+
test=dict(pipeline=test_pipeline))
147+
runner = dict(type='IterBasedRunner')
148+
checkpoint_config = dict(by_epoch=False, interval=1000, max_keep_ckpts=1)
149+
evaluation = dict(interval=4000, metric='mIoU', save_best='mIoU')

0 commit comments

Comments
 (0)