zhengye1995 / Tianchi-2019-Guangdong-Intelligent-identification-of-cloth-defects-rank5

天池2019广东工业智造创新大赛 布匹疵点检测 天池水也太深了 季军解决方案
399 stars 142 forks source link

请问作者当时网络到达多少代数才收敛的? #19

Closed Alex-afka closed 3 years ago

zhengye1995 commented 3 years ago

70 epochs, because: (i) I changed the frist conv layer in backbone, so I need train all backbone(not same as the original cascade rcnn which freezes the frist block in backbone) (ii) I did not add more anchor to adapt the long strip target so it needs more training iters to converge