AdvancedEAST
AdvancedEAST copied to clipboard
您好,请问act.jpg有把文字分割,但是画线却画不出这种情况怎么解决
You can try increasing trunc_threshold in cfg.py
@hcnhatnam thanks, i will try it.
@Owenscu Can you tell me what training you have on the dataset and what is the valid loss value after the training? my model is not converging and valid loss always >=0.7
@hcnhatnam on my dataset, val_loss is 0.35
what is the valid loss value on epoch 1 and where epoch drops to 0.35? Do you think 0.35 is still high?
@hcnhatnam it seems that is something wrong.
From epoch number 18 and after, this model is overfit. what is the valid loss value on epoch 1 ?
@hcnhatnam 0.59
oh...Your loss has decreased insignificantly....Did you train on any size [256, 384, 512, 640, 736]?
@hcnhatnam no. I will try when i am free. thanks
Ok @Owenscu .I still don't understand why the valid loss is still so high? I wonder if it's because the code is mistaken somewhere @huoyijie
@hcnhatnam enen, i am checking.
我成功运行了程序、想咨询一下,分阶段训练[256, 384, 512, 640, 736]的epoch问题。在比如在尺寸256,epoch=16的情况下,保存了最优模型。我在下一尺寸384,该怎么设置起始的epoch?我在后面也看到了回答,但是没看太懂,我的理解是下一阶段的epoch=17.请问是这样吗?期待回复。
It is not important to set the next epoch to 17 or not.
@hcnhatnam enen, i am checking.
华为书法识别?留个联系方式交流下吧
老哥、初赛马上就结束了啊。。。我Q:690866816。可讨论下
| | 张东阳 学生 | | 江西财经大学 17607086525 邮箱:[email protected] 江西省南昌市江西财经大学 690866816 海阔凭鱼跃 天高任鸟飞 |
签名由 网易邮箱大师 定制
在2019年03月24日 00:44,Henry 写道:
@hcnhatnam enen, i am checking.
华为书法识别?留个联系方式交流下吧
— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or mute the thread.
@zdyshine 我不是搞cv的,而且做这个比赛也比较晚,成绩不是很好,你还需要交流嘛,需要的话我就加你。
学习第一、比赛第二~~
| | 张东阳 学生 | | 江西财经大学 17607086525 邮箱:[email protected] 江西省南昌市江西财经大学 690866816 海阔凭鱼跃 天高任鸟飞 |
签名由 网易邮箱大师 定制
在2019年03月24日 11:34,Owenscu 写道:
@zdyshine 我不是搞cv的,而且做这个比赛也比较晚,成绩不是很好,你还需要交流嘛,需要的话我就加你。
— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub, or mute the thread.
@zdyshine 690866816这是你qq嘛
你好,我想问一下,Trunc_threshold这个阈值是啥意思呀?没搞清楚
我成功运行了程序、想咨询一下,分阶段训练[256, 384, 512, 640, 736]的epoch问题。在比如在尺寸256,epoch=16的情况下,保存了最优模型。我在下一尺寸384,该怎么设置起始的epoch?我在后面也看到了回答,但是没看太懂,我的理解是下一阶段的epoch=17.请问是这样吗?期待回复。
@zdyshine Set initial_epoch to 16, it will automatically increment this by one.
我检测长文本遇到同样问题,好多激活图(act.jpg)没有绿色和黄色边界点 @Owenscu
@cuimiao187561 Firstly try to increase trunc_threshold paramter. Secondly try finetune net by author provide method. finally you can modify nms strategy such as supplement by manual.
我检测长文本遇到同样问题,好多激活图(act.jpg)没有绿色和黄色边界点 @Owenscu
你好,想问问你有头没尾或者有尾无头的问题是怎么解决的?或者有什么参考的方法吗?
+1