forked from aharley/simple_bev
-
Notifications
You must be signed in to change notification settings - Fork 0
/
eval_runner.sh
131 lines (119 loc) · 4.94 KB
/
eval_runner.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
#!/bin/bash
DATA_DIR="/mnt/fsx/nuscenes"
# there should be ${DATA_DIR}/trainval/v1.0-trainval
# rgb00: default, on gpus 0-3
# rgb01: flip=True, on gpus 4-7
# python eval_nuscenes_bevseg.py \
# --exp_name="rgb00" \
# --data_dir=$DATA_DIR \
# --log_dir='logs_eval_nuscenes_bevseg' \
# --init_dir='checkpoints/40_3e-4_rgb00_00:52:38' \
# --device='cuda:0' \
# --device_ids=[0,1,2,3] \
# # --device_ids=[0,1,2,3] \
# 20k: 0.4644063812303654
# 19k: 0.4686688397908696
# 40_3e-5_rgb02_00:51:48 23k: 39.0 ? what the hell
# --init_dir='checkpoints/40_3e-5_rgb02_00:51:48' \
# --init_dir='checkpoints/40_3e-4_rgb01_01:13:43' \
# oh oh
# it's because i updated the cropping
# 40_3e-5_rgb02_00:51:48 23k and feed new resize lim stuff:
# --init_dir='checkpoints/40_3e-4_rgb01_01:13:43' \
# --dset='mini' \
# after updating the settings slightly:
# rgb02: 44.7
# ok garbage. let's wait and see the result of the next nets.
# rgb04: 25k: 45.79609990531495
# rgb05: 25k: 47.1787797626123
# 8x5_3e-4_rgb06_19:35:12: 46.463983336823304
# 8x5_3e-4_rgb07_20:26:00: 46.316773315601644
# 8x5_4e-4_rgb08_20:27:12: 46.4973733482569
# 8x5_5e-4_rgb09_19:15:31: 46.236855964350525
# 8x5_3e-4_rgb10_06:24:04: 46.540322982356514
# 8x5_3e-4_rgb11_02:54:25: 46.388311827282486
# gosh maybe none of these used the scheduler
# ...rewrote the eval
# rgb11 re-eval: 46.40237103956216 < ok match
# 8x5_5e-4_rgb12_22:43:46: 47.55780943879203
# 8x5_4e-4_rgb13_22:45:52: 47.40533491090929
# alt eval, where we return the iou every time
# rgb13 re-eval: iou_ev 47.4, iou_ev2 46.6
# so the global iou is helping
# but let's check again sanja's repo
# yup. https://github.com/nv-tlabs/lift-splat-shoot/blob/master/src/tools.py#L269
# 'iou': total_intersect / total_union,
# so let's stick with first eval
# 8x5_4e-4s_rgb14_21:14:05: 47.437589519538776
# 8x5_3e-4s_rgb15_21:20:26: 47.33251224323072
# now for ablations
# resolution
# 8x5_5e-4s_rgb16_04:10:41, scale 1: 42.36708138655802
# 8x5_5e-4s_rgb17_04:11:43, scale 3 (400):
# 8x5_5e-4s_rgb18_19:23:55, scale 0.5, also set to 416 instead of 400: 36.57758048736164
# batchsize
# 8_5e-4s_rgb19_19:26:58: 44.480124184301005
# 8x2_5e-4s_rgb20_19:30:31: 46.371658090584
# 4_5e-4s_rgb21_04:03:00
# 2_5e-4s_rgb22_04:06:15
# 1_5e-4s_rgb23_04:06:46
# 1_5e-4s_rgb24_05:57:24
# 1_5e-4s_rgb25_05:58:15
# cameras
# 8x5_5e-4s_rgb26_17:52:44
# 8x5_5e-4s_rgb27_17:15:55
python eval_nuscenes_bevseg.py \
--exp_name="rgb27" \
--data_dir=$DATA_DIR \
--log_dir='/mnt/fsx1/bev_baseline/logs_eval_nuscenes_bevseg' \
--init_dir='/mnt/fsx1/bev_baseline/checkpoints/8x5_5e-4s_rgb27_17:15:55' \
--resolution_scale=2 \
--device='cuda:4' \
--device_ids=[4,5,6,7]
# rad04 40k: 53.36171725828961
# rad09 30k: 53.189459455580256
# 40_3e-4_rad18_17:41:36 25k: 55.38595757475545
# 40_3e-4_rad16_03:12:44 at 122k: 55.13208535120829
# 40_3e-4_rad19_19:09:57 at 50k: 55.42852276637131
# 40_3e-4_rad20_19:13:17 at 50k: 55.55791382662683
# 40_3e-4_rad20_19:13:17 at 49k: 55.55368349387456
# note rgb05 had resize_lim 0.9-1.1; here it's 0.8-1.2
# 40_3e-4_rad21_21:16:12 25k: 55.00816918777085
# this was weight decay 1e-4
# but there was no effect on train compared to rad18, so i think this was just bad luck in the ckpt
# 8x5_3e-4_rad22_04:04:53 22k (killed by accident, use_metaradar=False by accident): 53.986605660376966 < not bad for a no meta model
# 8x5_3e-4_rad23_04:04:53 25k: 55.14818582340928
# 8x5_4e-4_rad24_22:39:20 25k: 55.5924135087522
# ok cool. let's increase one more time and go
# --init_dir='/mnt/fsx1/bev_baseline/checkpoints/8x5_4e-4_rad24_22:39:20' \
# 8x5_5e-4_rad25_18:55:34: 55.79075749280007
# 8x5_6e-4_rad25_18:56:01: 55.620367287757944
# 8x5_5e-4_rad27_19:26:17: 55.71121026326078
# python eval_nuscenes_bevseg.py \
# --exp_name="rad27" \
# --data_dir=$DATA_DIR \
# --log_dir='/mnt/fsx1/bev_baseline/logs_eval_nuscenes_bevseg' \
# --init_dir='/mnt/fsx1/bev_baseline/checkpoints/8x5_5e-4_rad27_19:26:17' \
# --use_radar=True \
# --use_metaradar=True \
# --device='cuda:4' \
# --device_ids=[4,5,6,7]
# # --device='cuda:0' \
# # --device_ids=[0,1,2,3]
# # --init_dir='/mnt/fsx1/bev_baseline/checkpoints/40_3e-4_rad18_17:41:36' \
# 8x5_5e-4_lid00_17:21:49: 63.89539955852593
# 8x5_4e-4_lid01_20:47:23: 63.95550985705564
# 8x5_3e-4_lid02_19:13:50: 63.468334914611006
# 8x5_6e-4_lid03_06:20:19: 64.17614962980704
# 8x5_6e-4_lid04_02:55:17: 63.87449500151892
# python eval_nuscenes_bevseg.py \
# --exp_name="lid04" \
# --data_dir=$DATA_DIR \
# --log_dir='/mnt/fsx1/bev_baseline/logs_eval_nuscenes_bevseg' \
# --init_dir='/mnt/fsx1/bev_baseline/checkpoints/8x5_6e-4_lid04_02:55:17' \
# --use_lidar=True \
# --device='cuda:0' \
# --device_ids=[0,1,2,3]
# # --device='cuda:0' \
# # --device_ids=[0,1,2,3]
# # --init_dir='/mnt/fsx1/bev_baseline/checkpoints/40_3e-4_rad18_17:41:36' \