训练结果记录

训练结果记录

train_dataset:day&night

Time	step	learning_rate	Loss	mAP_car	mAP_car_night	mAP_person
2017.12.4	100000	0.008	-	0.5790	0.5478	0.1130
2017.12.4	200000	0.008	-	0.6044	0.6196	0.1382
2017.12.4	300000	0.008	-	0.5938	0.6191	0.1329
2017.12.4	330000	0.008	-	0.6056	0.6376	0.1329
2017.12.4	350000	0.008	-	0.6040	0.6529	0.1599
2017.12.4	370000	0.008	-	0.5975	0.6197	0.1043
2017.12.4	400000	0.008	-	0.6061	0.6097	0.1017

---

valid_dataset:day&night

Time	step	learning_rate	Loss	mAP_car	mAP_car_night	mAP_person
2017.12.4	100000	0.008	-	0.6376	0.5769	0.1378
2017.12.4	200000	0.008	-	0.6712	0.5662	0.1814
2017.12.4	300000	0.008	-	0.6708	0.5918	0.2132
2017.12.4	330000	0.008	-	0.6679	0.6268	0.1661
2017.12.4	350000	0.008	-	0.6558	0.6369	0.1728
2017.12.4	370000	0.008	-	0.6684	0.6069	0.2190
2017.12.4	400000	0.008	-	0.6511	0.5870	0.1977

---

valid_dataset:night

Time	step	learning_rate	mAP_car_night
2017.12.4	100000	0.008	0.5770
2017.12.4	200000	0.008	0.5662
2017.12.4	300000	0.008	0.5918
2017.12.4	330000	0.008	0.6307
2017.12.4	350000	0.008	0.6369
2017.12.4	370000	0.008	0.6069
2017.12.4	400000	0.008	0.5893

---

valid_dataset:day

Time	step	learning_rate	mAP_car	mAP_person
2017.12.4	100000	0.008	0.6403	0.1613
2017.12.4	200000	0.008	0.6796	0.2073
2017.12.4	300000	0.008	0.6764	0.2510
2017.12.4	330000	0.008	0.6696	0.2009
2017.12.4	350000	0.008	0.6604	0.2039
2017.12.4	370000	0.008	0.6767	0.2545
2017.12.4	400000	0.008	0.6596	0.2330

---

imagenet_anngic5:基于Imagenet160万次继续训练anngic5类

Time	step	learning_rate	Loss	mAP_person	car	car_night	elec	bicycle
2017.12.11	100000	0.008	-	0.0623	0.4971	0.5874	0.2098	0.0531
2017.12.11	120000	0.008	-	0.0761	0.5172	0.5946	0.2789	0.0810
2017.12.11	150000	0.008	-	0.0950	0.5325	0.5893	0.2707	0.0873
2017.12.11	180000	0.008	-	0.0853	0.5076	0.5960	0.2633	0.0589
2017.12.11	200000	0.008	-	0.0799	0.5191	0.6133	0.2639	0.0509
2017.12.11	220000	0.008	-	0.0975	0.5239	0.6120	0.2848	0.0797
2017.12.11	250000	0.008	-	0.0833	0.5225	0.6166	0.2716	0.1077
2017.12.11	280000	0.008	-	0.0960	0.5207	0.6263	0.2311	0.0729
2017.12.11	300000	0.008	-	0.1073	0.5191	0.5955	0.2821	0.0433
2017.12.11	320000	0.008	-	0.0890	0.5424	0.6204	0.2791	0.0817

---

imagenet_coco_anngic5:基于Imagenet160万次继续训练coco4类

写入面积>0.008的person的box；面积>0.0025的其他类的box；
从coco随机挑选5000张验证集，验证集未过滤小目标

Time	step	learning_rate	Loss	mAP_person	car	elec	bicycle
2017.12.11	120000	0.008	-	0.2434	0.1266	0.2656	0.0830
2017.12.11	150000	0.008	-	0.2523	0.1302	0.2638	0.1054
2017.12.11	200000	0.008	-	0.2551	0.1361	0.2709	0.1141
2017.12.11	230000	0.008	-	0.2602	0.1288	0.2785	0.1186
2017.12.11	250000	0.008	-	0.2591	0.1351	0.1073	0.1567

若使用23800张验证集：

Time	step	learning_rate	Loss	mAP_person	car	elec	bicycle
2017.12.12	250000	0.008	-	0.2173	0.1153	0.2316	0.0793
2017.12.12	300000	0.008	-	0.2227	0.1212	0.2369	0.0876

---

imagenet_tiny_coco_anngic5:基于Imagenet160万次继续训练tiny-coco4类

写入面积>0.1的person的box；面积>0.0025的其他类的box；
从coco随机挑选5000张验证集，验证集未过滤小目标
删减人样本数量，提高样本均衡性

Time	step	learning_rate	Loss	mAP_person	car	elec	bicycle
2017.12.11	120000	0.008	-	0.1492	0.1439	0.2949	0.1072
2017.12.11	150000	0.008	-	0.1437	0.1548	0.2848	0.1087
2017.12.11	200000	0.008	-	0.1579	0.1552	0.2847	0.1244
2017.12.11	230000	0.008	-	0.1468	0.1483	0.2927	0.1158
2017.12.11	250000	0.008	-	0.1669	0.1490	0.2952	0.1347

若使用23800张验证集：

Time	step	learning_rate	Loss	mAP_person	car	elec	bicycle
2017.12.12	250000	0.008	-	0.1405	0.1244	0.2363	0.0961
2017.12.12	300000	0.008	-	0.1157	0.1330	0.2341	0.0933

---

官网yolo.weights测试coco80类完整验证集（只展示需要类别）

Time	mAP_person	car	bus	truck	motorcycle	bicycle
2017.12.11	0.5872	0.2653	0.6051	0.3337	0.4762	0.2913

---

imagenet_coco5:基于Imagenet160万次继续训练coco4类

训练集：写入面积>0.008的person的box；面积>0.0025的其他类的box；
验证集:写入面积>0.008的person的box；面积>0.0025的其他类的box；

Time	step	learning_rate	Loss	mAP_person	car	elec	bicycle
2017.12.14	100000	0.008	-	0.4076	0.1776	0.2616	0.0988
2017.12.14	150000	0.008	-	0.4237	0.1793	0.2442	0.1031
2017.12.14	200000	0.008	-	0.4216	0.1794	0.2748	0.1124

继续基于imagenet_coco5_100000.weights 使用poly学习率训练

Time	step	learning_rate	Loss	mAP_person	car	elec	bicycle
2017.12.14	100000	0.008000	9.495427	0.4076	0.1776	0.2616	0.0988
2017.12.15	150000	0.001221	9.663104	0.4460	0.2031	0.2969	0.1167
2017.12.15	170000	0.000875	8.008812	0.4547	0.2070	0.3045	0.1318
2017.12.15	200000	0.000500	8.623980	0.4762	0.2296	0.3138	0.1430

---

imagenet_tiny_coco5:基于Imagenet160万次继续训练coco4类

训练集：写入面积>0.008的person的box；面积>0.0025的其他类的box；
若一张图片中有10个以上面积<0.008的人 ，则删去这张图
验证集:写入面积>0.008的person的box；面积>0.0025的其他类的box；

Time	step	learning_rate	Loss	mAP_person	car	elec	bicycle
2017.12.14	100000	0.008	-	0.3995	0.1780	0.2453	0.1036
2017.12.14	150000	0.008	-	0.4344	0.1849	0.2904	0.1177
2017.12.14	200000	0.008	-	0.4162	0.1702	0.2713	0.1073

继续基于imagenet_tiny_coco5_100000.weights 使用poly学习率训练

Time	step	learning_rate	Loss	mAP_person	car	elec	bicycle
2017.12.14	100000	0.00800	10.087712	0.3995	0.1780	0.2453	0.1036
2017.12.15	150000	0.001221	7.373747	0.4423	0.1749	0.2853	0.1285
2017.12.15	170000	0.000875	8.408542	0.4756	0.2216	0.3113	0.1420
2017.12.15	200000	0.000500	8.078330	0.4293	0.1714	0.2797	0.1176

---

imagenet_tiny_coco_2:基于Imagenet160万次继续训练coco4类

训练集：写入面积>0.008且<0.52的person的box；面积>0.0025的其他类的box；（与imagenet_tiny_coco_3区别：失误在每个labels里面重复写了一遍box，故2中labels大小是3的两倍）
验证集:写入面积>0.008的person的box；面积>0.0025的其他类的box；

Time	step	learning_rate	Loss	mAP_person	car	elec	bicycle
2017.12.14	100000	0.008	-	0.4141	0.1873	0.2737	0.1052
2017.12.14	130000	0.008	-	0.4214	0.1732	0.2605	0.0996
2017.12.14	140000	0.008	-	0.4165	0.1879	0.2692	0.1037
2017.12.14	150000	0.008	-	0.3857	0.1652	0.2482	0.0981

继续基于imagenet_tiny_coco5_2_100000.weights 使用poly学习率训练

Time	step	learning_rate	Loss	mAP_person	car	elec	bicycle
2017.12.14	100000	0.00800	10.520602	0.4141	0.1873	0.2737	0.1052
2017.12.15	150000	0.001221	8.319970	0.4019	0.1793	0.2709	0.1124
2017.12.15	170000	0.000875	8.257561	0.3695	0.1623	0.2619	0.1023
2017.12.15	200000	0.000500	9.154430	0.4716	0.2341	0.3270	0.1372

---

imagenet_tiny_coco_3:基于Imagenet160万次继续训练coco4类

训练集：写入面积>0.008且<0.52的person的box；面积>0.0025的其他类的box；
验证集:写入面积>0.008的person的box；面积>0.0025的其他类的box；
继续基于imagenet_tiny_coco5_3_100000.weights 使用poly学习率训练

Time	step	learning_rate	Loss	mAP_person	car	elec	bicycle
2017.12.14	100000	0.008000	8.811008	0.3098	0.1217	0.1898	0.0747
2017.12.15	150000	0.001221	7.485109	0.4565	0.2063	0.2957	0.1306
2017.12.15	170000	0.000875	8.071655	0.4638	0.2264	0.3224	0.1418
2017.12.15	200000	0.000500	8.536077	0.4745	0.2373	0.3258	0.1490

基于imagenet_tiny_coco_3训练20万次基础，继续partial前13层训练anngic结果记录如下。

不partial 直接训练