-
Notifications
You must be signed in to change notification settings - Fork 8
/
swinv2l_480reso_parallel_depthonly.py
136 lines (124 loc) · 3.79 KB
/
swinv2l_480reso_parallel_depthonly.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
checkpoint_config = dict(interval=5050)
log_config = dict(
interval=20,
hooks=[
dict(type='TextLoggerHook'),
dict(type='TensorboardLoggerHook')
])
custom_hooks = [dict(type='NumClassCheckHook')]
dist_params = dict(backend='nccl')
log_level = 'INFO'
load_from = None
resume_from = None
workflow = [('train', 1)]
num_bins = 2000
num_classes = 80
num_embeddings = 128
num_embeddings_depth = 128
num_vocal = num_bins+1 + num_classes + 2 + num_embeddings + num_embeddings_depth
model = dict(
type='AiT',
backbone=dict(
type="SwinV2TransformerRPE2FC",
pretrain_img_size=192,
embed_dim=192,
depths=[2, 2, 18, 2],
num_heads=[6, 12, 24, 48],
window_size=[30, 30, 30, 15],
use_shift=[True, True, False, False],
pretrain_window_size=[12, 12, 12, 6],
ape=False,
drop_path_rate=0.1,
patch_norm=True,
use_checkpoint=False,
out_indices=(3,),
init_cfg=dict(type='Pretrained',
checkpoint='swin_v2_large_densesimmim.pth'),
),
transformer=dict(
type='ARTransformer',
in_chans=1536,
d_model=256,
drop_path=0.1,
drop_out=0.1,
nhead=8,
dim_feedforward=1024,
num_encoder_layers=6,
num_decoder_layers=6,
num_vocal=num_vocal,
num_bins=num_bins,
num_classes=num_classes,
num_embeddings=num_embeddings,
num_embeddings_depth=num_embeddings_depth,
dec_length=2100,
n_rows=15,
n_cols=15,
pos_enc='sine',
pred_eos=False,
parallel=True,
soft_vae=True,
soft_transformer=False,
top_p=0.
),
task_heads=dict(
depth=dict(
type='DepthHead',
task_id=2,
loss_weight=1.,
depth_token_offset=num_bins+1 + num_classes + 2 + num_embeddings,
vae_cfg=dict(
type='VQVAE',
use_norm=False,
token_length=15*15,
mask_size=480,
embedding_dim=512,
hidden_dim=256,
num_resnet_blocks=2,
num_embeddings=num_embeddings_depth,
tau=0.8,
pretrained='vqvae_depth.pt',
freeze=True
),
decoder_loss_weight=1.0,
soft_vae=True),
))
img_norm_cfg = dict(
mean=[123.675, 116.28, 103.53], std=[58.395, 57.12, 57.375], to_rgb=True)
file_client_args = dict(backend='disk')
runner = dict(type='IterBasedRunnerMultitask', max_iters=25250)
evaluation = dict(interval=25250)
# learning policy
lr_config = dict(
policy='Step',
step=[18180], # 18/25 epoch
by_epoch=False,
warmup_ratio=0.1,
warmup='linear',
warmup_by_epoch=False,
warmup_iters=500,
)
# optimizer
optimizer = dict(
type='AdamW',
lr=1e-4,
weight_decay=0.075,
constructor='SwinLayerDecayOptimizerConstructor',
paramwise_cfg=dict(
num_layers=[2, 2, 18, 2], layer_decay_rate=0.9,
no_decay_names=['relative_position_bias_table', 'rpe_mlp', 'logit_scale',
'det_embed', 'voc_embed', 'enc_embed', 'dec_embed', 'mask_embed'],
))
optimizer_config = dict(grad_clip={'max_norm': 10, 'norm_type': 2})
task = dict(
depth=dict( # len=24231
times=1,
data=dict(
train=dict(type='nyudepthv2', data_path='data', filenames_path='code/dataset/depth/filenames/',
is_train=True, crop_size=(480, 480), samples_total_gpu=24),
val=dict(type='nyudepthv2', data_path='data', filenames_path='code/dataset/depth/filenames/',
is_train=False, crop_size=(480, 480), samples_per_gpu=2, workers_per_gpu=8),
)
),
)
# enable fp16
fp16 = dict(loss_scale='dynamic')