-
Notifications
You must be signed in to change notification settings - Fork 105
/
Copy pathbinsformer_swint_w7_nyu.py
203 lines (195 loc) · 6.51 KB
/
binsformer_swint_w7_nyu.py
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
_base_ = [
'../_base_/models/binsformer.py',
'../_base_/default_runtime.py',
]
model = dict(
pretrained='https://github.com/SwinTransformer/storage/releases/download/v1.0.0/swin_tiny_patch4_window7_224.pth', # noqa
backbone=dict(
embed_dims=96,
depths=[2, 2, 6, 2],
num_heads=[3, 6, 12, 24],
window_size=7,
use_abs_pos_embed=False,
drop_path_rate=0.3,
patch_norm=True,
pretrain_style='official'),
decode_head=dict(
type='BinsFormerDecodeHead',
class_num=25,
in_channels=[96, 192, 384, 768],
conv_dim=256,
min_depth=1e-3,
max_depth=10,
n_bins=64,
index=[0, 1, 2, 3],
trans_index=[1, 2, 3], # select index for cross-att
loss_decode=dict(type='SigLoss', valid_mask=True, loss_weight=10),
with_loss_chamfer=False, # do not use chamfer loss
loss_chamfer=dict(type='BinsChamferLoss', loss_weight=1e-1),
classify=True, # class embedding
loss_class=dict(type='CrossEntropyLoss', loss_weight=1e-2),
norm_cfg=dict(type='BN', requires_grad=True),
transformer_encoder=dict( # default settings
type='PureMSDEnTransformer',
num_feature_levels=3,
encoder=dict(
type='DetrTransformerEncoder',
num_layers=6,
transformerlayers=dict(
type='BaseTransformerLayer',
attn_cfgs=dict(
type='MultiScaleDeformableAttention',
embed_dims=256,
num_levels=3,
num_points=8),
feedforward_channels=1024,
ffn_dropout=0.1,
operation_order=('self_attn', 'norm', 'ffn', 'norm')))),
positional_encoding=dict(
type='SinePositionalEncoding', num_feats=128, normalize=True),
transformer_decoder=dict(
type='PixelTransformerDecoder',
return_intermediate=True,
num_layers=9,
num_feature_levels=3,
hidden_dim=256,
operation='//',
transformerlayers=dict(
type='PixelTransformerDecoderLayer',
attn_cfgs=dict(
type='MultiheadAttention',
embed_dims=256,
num_heads=8,
dropout=0.0),
ffn_cfgs=dict(
feedforward_channels=2048,
ffn_drop=0.0),
operation_order=('cross_attn', 'norm', 'self_attn', 'norm', 'ffn', 'norm')))),
train_cfg=dict(
aux_loss = True,
aux_index = [2, 5],
aux_weight = [1/4, 1/2]
),
test_cfg=dict(mode='whole')
)
# dataset settings
dataset_type = 'NYUBinFormerDataset'
data_root = 'data/nyu/'
img_norm_cfg = dict(
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
crop_size= (416, 544)
train_pipeline = [
dict(type='LoadImageFromFile'),
dict(type='DepthLoadAnnotations'),
dict(type='NYUCrop', depth=True),
dict(type='RandomRotate', prob=0.5, degree=2.5),
dict(type='RandomFlip', prob=0.5),
dict(type='RandomCrop', crop_size=(416, 544)),
dict(type='ColorAug', prob=0.5, gamma_range=[0.9, 1.1], brightness_range=[0.75, 1.25], color_range=[0.9, 1.1]),
dict(type='Normalize', **img_norm_cfg),
dict(type='DefaultFormatBundle'),
dict(type='Collect', keys=['img', 'depth_gt', 'class_label']),
]
test_pipeline = [
dict(type='LoadImageFromFile'),
dict(
type='MultiScaleFlipAug',
img_scale=(480, 640),
flip=True,
flip_direction='horizontal',
transforms=[
dict(type='RandomFlip', direction='horizontal'),
dict(type='Normalize', **img_norm_cfg),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect', keys=['img']),
])
]
# for visualization of pc
eval_pipeline = [
dict(type='LoadImageFromFile'),
dict(type='RandomFlip', prob=0.0), # set to zero
dict(type='Normalize', **img_norm_cfg),
dict(type='ImageToTensor', keys=['img']),
dict(type='Collect',
keys=['img'],
meta_keys=('filename', 'ori_filename', 'ori_shape',
'img_shape', 'pad_shape', 'scale_factor',
'flip', 'flip_direction', 'img_norm_cfg',
'cam_intrinsic')),
]
data = dict(
samples_per_gpu=2,
workers_per_gpu=2,
train=dict(
type=dataset_type,
data_root=data_root,
depth_scale=1000,
split='nyu_train.txt',
pipeline=train_pipeline,
garg_crop=False,
eigen_crop=True,
min_depth=1e-3,
max_depth=10),
val=dict(
type=dataset_type,
data_root=data_root,
depth_scale=1000,
split='nyu_test.txt',
pipeline=test_pipeline,
garg_crop=False,
eigen_crop=True,
min_depth=1e-3,
max_depth=10),
test=dict(
type=dataset_type,
data_root=data_root,
depth_scale=1000,
split='nyu_test.txt',
pipeline=test_pipeline,
garg_crop=False,
eigen_crop=True,
min_depth=1e-3,
max_depth=10))
# AdamW optimizer, no weight decay for position embedding & layer norm
# in backbone
max_lr = 1e-4
optimizer = dict(
type='AdamW',
lr=max_lr,
betas=(0.9, 0.999),
weight_decay=0.01,
paramwise_cfg=dict(
custom_keys={
'absolute_pos_embed': dict(decay_mult=0.),
'relative_position_bias_table': dict(decay_mult=0.),
'norm': dict(decay_mult=0.),
}))
# learning policy
lr_config = dict(
policy='OneCycle',
max_lr=max_lr,
warmup_iters=1600 * 8,
div_factor=25,
final_div_factor=100,
by_epoch=False)
optimizer_config = dict(grad_clip=dict(max_norm=35, norm_type=2))
# runtime settings
runner = dict(type='IterBasedRunner', max_iters=1600 * 24)
checkpoint_config = dict(by_epoch=False, max_keep_ckpts=2, interval=1600)
evaluation = dict(by_epoch=False,
start=0,
interval=1600,
pre_eval=True,
rule='less',
save_best='abs_rel',
greater_keys=("a1", "a2", "a3"),
less_keys=("abs_rel", "rmse"))
# iter runtime
log_config = dict(
_delete_=True,
interval=50,
hooks=[
dict(type='TextLoggerHook', by_epoch=False),
dict(type='TensorboardLoggerHook') # TensorboardImageLoggerHook
])
find_unused_parameters=True