-
Notifications
You must be signed in to change notification settings - Fork 78
/
lift.sh
143 lines (136 loc) · 4.21 KB
/
lift.sh
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
#!/bin/bash
DATA_DIR="/mnt/fsx/nuscenes"
# DATA_DIR="../nuscenes/full_v1.0"
# DATA_DIR="../nuscenes"
# lif00: debug
# lif01: show me feat_bev
# lif02: ncams=6, batch1, grad8
# lif03: ncams=6, batch1, grad1
# lif04: 1k iters
# lif05: 100, liftnet
EXP_NAME=lif06 # uniform dist
EXP_NAME=lif07 # 4 to Z
EXP_NAME=lif08 # 4 to Z+4
EXP_NAME=lif09 # 5 to Z+5
EXP_NAME=lif10 # div by ones
EXP_NAME=lif11 # non-uniform
EXP_NAME=lif12 # /0.07
EXP_NAME=lif13 # train 10k
# ok, the fiery code is quite good... let's try that. once i get coffee anyway.
EXP_NAME=lif14 # fiery 100
# only weird thing is that i'm forced to sum over bev myself
# < ah but they just had one giant bin, that's why
EXP_NAME=lif15 # fewer prints
EXP_NAME=lif16 # 10k like this
EXP_NAME=lif17 # 100/10, where all Y coords are zero
EXP_NAME=lif18 # 4gpu, acc1, 10k like this
EXP_NAME=lif19 # 1gpu acc4
EXP_NAME=lif20 # 4gpu acc1, dset=trainval, 25k
EXP_NAME=lif24 # 4gpu acc1, dset=trainval, 25k < i don't have trainval
EXP_NAME=lif25 # mini, 100,10
# what is the problem?
# why are we getting overfitting, assuming the rest is correct?
EXP_NAME=lif26 # random depth
EXP_NAME=lif27 # 0.9 of hypoteneuse
EXP_NAME=lif28 # pred
EXP_NAME=lif21 # don't divide < smoother
EXP_NAME=lif22 # show me the middle slice
EXP_NAME=lif23 # and beg and end
EXP_NAME=lif24 # hm... middle slice again
EXP_NAME=lif25 # unproj ones
EXP_NAME=lif26 # divide by ones
EXP_NAME=lif27 # three ones
# why isn't the camera setup looking more consistent?
EXP_NAME=lif28 # no crop
EXP_NAME=lif29 # don't shuffle cams
# ok looks solid
EXP_NAME=lif30 # 0.8 hyp
EXP_NAME=lif31 # show me the actual feat_bev
EXP_NAME=lif32 # 0.7
EXP_NAME=lif33 # 0.9 again
EXP_NAME=lif34 # rand
EXP_NAME=lif35 # softmax
EXP_NAME=lif36 # pred
EXP_NAME=lif37 # don't set y to 0
EXP_NAME=lif38 # feat=ones
EXP_NAME=lif39 # 8 instead of 4 < it's already 8
EXP_NAME=lif40 # show me max
EXP_NAME=lif41 # pred
EXP_NAME=lif42 # show me mean again
EXP_NAME=lif43 # 1k
EXP_NAME=lif44 # ones, 100,10, better averaging
# feat_bev (float32) min = -12.72, mean = 0.00, max = 13.82
EXP_NAME=lif45 # ones, beg middle end
# min = 0.00, mean = 0.01, max = 1.00
# but not lining up
EXP_NAME=lif46 # don't clone
# still broken
EXP_NAME=lif47 # don't div by ones
EXP_NAME=lif48 # set y to 0
EXP_NAME=lif49 # set y to 1
# ok interesting:
# when i allow y to spread out, it goes everywhere
EXP_NAME=lif50 # don't set y
EXP_NAME=lif51 # ZYX when creating inds
EXP_NAME=lif52 # YX, X for coords
EXP_NAME=lif53 # div by output_ones
EXP_NAME=lif54 # pred
# let's compare cleanly
EXP_NAME=lif55 # don't print stats; go 1k
EXP_NAME=lif56 # don't div by ones, just to se
EXP_NAME=lif57 # div x_b also
EXP_NAME=lif58 # aws
# python train_nuscenes_bevseg.py \
# --exp_name=${EXP_NAME} \
# --max_iters=1000 \
# --log_freq=100 \
# --shuffle=True \
# --dset='mini' \
# --do_val=True \
# --val_freq=100 \
# --save_freq=9999999 \
# --batch_size=1 \
# --grad_acc=1 \
# --lr=5e-4 \
# --use_scheduler=True \
# --weight_decay=1e-6 \
# --nworkers=12 \
# --data_dir=$DATA_DIR \
# --log_dir='logs_nuscenes_bevseg' \
# --ckpt_dir='checkpoints/' \
# --do_rgbcompress=False \
# --res_scale=1 \
# --rand_flip=False \
# --rand_crop_and_resize=False \
# --do_shuffle_cams=False \
# --ncams=6 \
# --nsweeps=3 \
# --encoder_type='res101' \
# --device_ids=[0]
python train_nuscenes_bevseg.py \
--exp_name=${EXP_NAME} \
--max_iters=25000 \
--log_freq=1000 \
--shuffle=True \
--dset='trainval' \
--do_val=True \
--val_freq=100 \
--save_freq=1000 \
--batch_size=8 \
--grad_acc=5 \
--lr=5e-4 \
--use_scheduler=True \
--weight_decay=1e-6 \
--nworkers=12 \
--data_dir=$DATA_DIR \
--log_dir='/mnt/fsx1/bev_baseline/logs_nuscenes_bevseg' \
--ckpt_dir='/mnt/fsx1/bev_baseline/checkpoints/' \
--do_rgbcompress=False \
--res_scale=2 \
--rand_flip=True \
--rand_crop_and_resize=True \
--do_shuffle_cams=True \
--ncams=6 \
--nsweeps=3 \
--encoder_type='res101' \
--device_ids=[0,1,2,3,4,5,6,7]