1+ _base_ = [
2+ '../_base_/models/depthformer_swin.py' ,
3+ '../_base_/default_runtime.py'
4+ ]
5+ USEPE_FLAG = True
6+ depth_scale = 250
7+ # dataset settings
8+ dataset_type = 'DDADDataset'
9+ data_root = 'data/DDAD'
10+ img_norm_cfg = dict (
11+ mean = [123.675 , 116.28 , 103.53 ], std = [58.395 , 57.12 , 57.375 ], to_rgb = True )
12+ train_pipeline = [
13+ dict (type = 'LoadDDADImageFromFile' ,USEPE = USEPE_FLAG ,USE_DYNAMIC_PE = True ),
14+ dict (type = 'DDADDepthLoadAnnotations' ,USE_DYNAMIC_PE = True ),
15+ dict (type = 'DDADResize' ,shape = (384 , 640 ),USE_DYNAMIC_PE = True ),
16+ dict (type = 'Resize' ,ratio_range = (0.5 ,2.0 )),
17+ dict (type = 'Padding' ,img_padding_value = (0 ,0 ,0 ),depth_padding_value = 255 ,pe_k = True ,ori_h = 384 ,ori_w = 640 ),
18+ dict (type = 'RandomRotate' , prob = 0.5 , degree = 2.5 ),
19+ dict (type = 'RandomFlip' , prob = 0.0 ),
20+ dict (type = 'RandomCrop' , crop_size = (384 , 640 )),
21+ dict (type = 'ColorAug' , prob = 0.5 , gamma_range = [0.9 , 1.1 ], brightness_range = [0.9 , 1.1 ], color_range = [0.9 , 1.1 ]),
22+ dict (type = 'Normalize' , depth_scale = depth_scale , ** img_norm_cfg ),
23+ dict (type = 'DefaultFormatBundle' ),
24+ dict (type = 'Collect' ,
25+ keys = ['img' , 'depth_gt' ,'pe_k_gt' ,'height' ],
26+ meta_keys = ('filename' , 'ori_filename' , 'ori_shape' ,
27+ 'img_shape' , 'pad_shape' , 'scale_factor' ,
28+ 'flip' , 'flip_direction' , 'img_norm_cfg' ,)),
29+ ]
30+ test_pipeline = [
31+ dict (type = 'LoadDDADImageFromFile' ,USEPE = USEPE_FLAG ,USE_DYNAMIC_PE = True ),
32+ dict (type = 'DDADResize' ,shape = (384 , 640 ),depth = False ),
33+ dict (
34+ type = 'MultiScaleFlipAug' ,
35+ img_scale = (384 , 640 ),
36+ flip = False ,
37+ flip_direction = 'horizontal' ,
38+ transforms = [
39+ dict (type = 'Normalize' , depth_scale = depth_scale ,** img_norm_cfg ),
40+ dict (type = 'ImageToTensor' , keys = ['img' ]),
41+ dict (type = 'Collect' ,
42+ keys = ['img' ,'height' ,'test' ],
43+ meta_keys = ('filename' , 'ori_filename' , 'ori_shape' ,
44+ 'img_shape' , 'pad_shape' , 'scale_factor' ,
45+ 'flip' , 'flip_direction' , 'img_norm_cfg' ,)),
46+ ])
47+ ]
48+ data = dict (
49+ samples_per_gpu = 4 ,
50+ workers_per_gpu = 4 ,
51+ train_dataloader = dict (
52+ shuffle = True ,
53+ drop_last = True ,
54+ persistent_workers = False ),
55+ val_dataloader = dict (
56+ shuffle = False ,
57+ persistent_workers = True ),
58+ test_dataloader = dict (
59+ shuffle = False ,
60+ persistent_workers = False ),
61+ train = dict (
62+ type = dataset_type ,
63+ split = 'splits/ddad_train_split.txt' ,
64+ pipeline = train_pipeline ,
65+ min_depth = 1e-3 ,
66+ max_depth = 200 ,
67+ cameras = ['CAMERA_%02d' % idx for idx in [1 ,5 ,6 ,9 ]],
68+ ),
69+ val = dict (
70+ type = dataset_type ,
71+ split = 'splits/ddad_test_split.txt' ,
72+ pipeline = test_pipeline ,
73+ min_depth = 1e-3 ,
74+ max_depth = 200 ,
75+ cameras = ['CAMERA_%02d' % idx for idx in [1 ,5 ,6 ,9 ]],
76+ ),
77+ test = dict (
78+ type = dataset_type ,
79+ split = 'splits/ddad_test_split.txt' ,
80+ pipeline = test_pipeline ,
81+ min_depth = 1e-3 ,
82+ max_depth = 200 ,
83+ cameras = ['CAMERA_%02d' % idx for idx in [1 ,5 ,6 ,9 ]],
84+ ))
85+
86+
87+
88+ model = dict (
89+ pretrained = None ,
90+ depth_scale = depth_scale ,
91+ backbone = dict (
92+ embed_dims = 192 ,
93+ depths = [2 , 2 , 18 , 2 ],
94+ num_heads = [6 , 12 , 24 , 48 ],
95+ window_size = 7 ,
96+ USEPE = USEPE_FLAG ),
97+ neck = dict (
98+ type = 'HAHIHeteroNeck' ,
99+ positional_encoding = dict (
100+ type = 'SinePositionalEncoding' , num_feats = 256 ),
101+ in_channels = [64 , 192 , 384 , 768 , 1536 ],
102+ out_channels = [64 , 192 , 384 , 768 , 1536 ],
103+ embedding_dim = 512 ,
104+ scales = [1 , 1 , 1 , 1 , 1 ]),
105+ pe_mask_neck = dict (
106+ type = 'LightPEMASKNeck'
107+ ),
108+ dynamic_pe_neck = dict (
109+ type = 'DynamicPENeckSOFT'
110+ ),
111+ decode_head = dict (
112+ type = 'DenseDepthHead' ,
113+ act_cfg = dict (type = 'LeakyReLU' , inplace = True ),
114+ in_channels = [64 , 192 , 384 , 768 , 1536 ],
115+ up_sample_channels = [64 , 192 , 384 , 768 , 1536 ],
116+ channels = 64 ,
117+ min_depth = 1e-3 ,
118+ max_depth = 200 ,
119+ ))
120+ # schedules
121+ # optimizer
122+ max_lr = 1e-4
123+ optimizer = dict (
124+ type = 'AdamW' ,
125+ lr = max_lr ,
126+ betas = (0.9 , 0.999 ),
127+ weight_decay = 0.01 ,
128+ paramwise_cfg = dict (
129+ custom_keys = {
130+ 'absolute_pos_embed' : dict (decay_mult = 0. ),
131+ 'relative_position_bias_table' : dict (decay_mult = 0. ),
132+ 'norm' : dict (decay_mult = 0. ),
133+ }))
134+ # learning policy
135+ lr_config = dict (
136+ policy = 'CosineAnnealing' ,
137+ min_lr_ratio = 1e-8 ,
138+ by_epoch = False ) # test add by_epoch false
139+ optimizer_config = dict (grad_clip = dict (max_norm = 35 , norm_type = 2 ))
140+ # runtime settings
141+ runner = dict (type = 'IterBasedRunner' , max_iters = 38400 )
142+ checkpoint_config = dict (by_epoch = False , max_keep_ckpts = 2 , interval = 800 )
143+ evaluation = dict (by_epoch = False ,
144+ start = 0 ,
145+ interval = 800 ,
146+ pre_eval = True ,
147+ rule = 'less' ,
148+ save_best = 'abs_rel' ,
149+ greater_keys = ("a1" , "a2" , "a3" ),
150+ less_keys = ("abs_rel" , "rmse" ))
151+ # iter runtime
152+ log_config = dict (
153+ _delete_ = True ,
154+ interval = 50 ,
155+ hooks = [
156+ dict (type = 'TextLoggerHook' , by_epoch = False ),
157+ dict (type = 'TensorboardLoggerHook' )
158+ ])
0 commit comments