TinyPerson_v2

• Big table

GPU: 3080 x2
train set: train + val; eval set: test
cut size: (640, 640)

detector	batch	lr			script
Faster RCNN	4x2	0.04	67.36 73.44	54.51 60.78	exp/sh/Baseline_TinyPersonV2.sh:1.3
RetinaNet	4x2	0.04	54.03 72.75	42.11 59.45	exp/sh/Baseline_TinyPersonV2.sh:2.1
Adap RetinaNet	3x2	0.04	74.19	63.78	exp/sh/Baseline_TinyPersonV2.sh:2.2
FCOS	4x2	0.04	48.09 65.08	41.39 54,37	exp/sh/Baseline_TinyPersonV2.sh:3.1
Adap FCOS	4x2	0.04	65.32 71.59	54.02 60.98	exp/sh/Baseline_TinyPersonV2.sh:3.2
RepPoint	4x2	0.04	54.71 73.47	43.18 60.79	exp/sh/Baseline_TinyPersonV2.sh:4.1
Adap RepPoint	4x2	0.04	68.92 75.27	56.17 64.56	exp/sh/Baseline_TinyPersonV2.sh:4.2

• 目录结构

数据集位置：~/dataset/tiny_set_v2
数据集来源：17suo(基础) + tiny_set_v1
规则：模态/视频号/图片
划分：10: 1 : 10（视频号级别）

• TinyPerson_v2（RGB）

Tab. 1. 目标size ratio统计.

dataset_v3.0.1	absolute size	relative size	aspect ratio
TinyPerson_v2 (rgb)	23.197±9.815	0.016±0.007	0.731±0.422
TinyPerson_v2 (infra)	12.614±4.817	0.017±0.006	0.827±0.422

dataset_v2.0.1	absolute size	relative size	aspect ratio
TinyPerson_v2 (rgb)	22.692±12.117	0.016±0.008	0.746±0.509
TinyPerson_v2 (rgb without ignore)	22.619±10.849	0.016±0.007	0.723±0.424
TinyPerson_v2 (xray)	13.551±5.078	0.018±0.007	0.874±0.495
TinyPerson_v2 (xray without ignore)	13.615±5.099	0.018±0.007	0.862±0.481

TinyPerson

Tab. 2. 数据集图片与标注数量.

TinyPerson_v3.0.1	Train set	Valid set	Test set	sum
#image (rgb)	4614	375	3559	8548
#image (infra)	4614	375	3559	8548
#annotation (rgb)	170383	16408	275486	462277
#annotation (infra)	172190	16447	228775	417412

TinyPerson_v2.0.1	Train set	Valid set	Test set	sum
#image (rgb)	5711	568	5753 (5754?)	12032
#image (xray)	8938	906	8950	18794
#annotation (rgb)	262063	42399	315165	619627
#annotation (xray)	390356	72525	291968	754849

TinyPerson

Fig. 分布.

TinyPerson

[Issue]

• issue: 无法算均值方差，存在annitation的bbox为nan

• issue: 复杂bug（7.9）>> [继续train]

• issue: >> [继续train]
  --resume-from + {dir/} + epoch_9.pth

--resume-from ../TOV_mmdetection_cache/work_dir/${CONFIG}/trainval${WH[0]}x${WH[1]}_lr${LR}_clipg10_1x_b${B}${GPU}g/epoch_9.pth

• issue: >> [re-test]
  evaluation.do_final_eval=True \
--resume-from + {dir/} + epoch_9.pth

evaluation.do_final_eval=True \
--resume-from ../TOV_mmdetection_cache/work_dir/${CONFIG}/trainval${WH[0]}x${WH[1]}_lr${LR}_1x_b${B}${GPU}g/epoch_12.pth

• issue: >> [删除无用Epoch]
  $python exp/tools/clear_tmp_pth.py ../TOV_mmdetection_cache/work_dir/

• issue: 发散 >> [clip_g]
  梯度过大，loss发散，但仍需训练。可改进网路，增加梯度裁剪。

  
  issue 1: loss不收敛

issue 2: loss NaN，程序退出

FCOS

"""
"configs2/TinyPersonV2/base/fcos_r50_caffe_fpn_gn-head_1x_TinyPersonV2_640.py"
"""
optimizer_config = dict(    # wd: 改小grad_clip['max_norm']
    _delete_=True, grad_clip=dict(max_norm=15, norm_type=2))
# optimizer_config = dict(
#     _delete_=True, grad_clip=dict(max_norm=35, norm_type=2))

# learning policy
lr_config = dict(
    warmup_iters=500,)    # wd: 观察500 iters

RetinaNet

"""
"configs2/TinyPersonV2/base/retinanet_r50_fpn_1x_TinyPersonV2_640.py"
"""
optimizer_config = dict(    # wd: 改小grad_clip['max_norm']
    _delete_=True, grad_clip=dict(max_norm=15, norm_type=2))
# optimizer_config = dict(
#     _delete_=True, grad_clip=dict(max_norm=35, norm_type=2))

# learning policy
lr_config = dict(
    warmup_iters=500,)    # wd: 观察500 iters

---

# xb exp
python exp/tools/semi_coarse_point_manager.py
python exp/tools/coarse_point_manager.py