怎么训练自己的数据呢

[chunleiml]

No description provided.

[Tianxiaomo]

目前还不支持，后续会加上的

可以大家一起实现

[17868380981]

我也在研究这个，我的微信17868380981，可以一起搞嘛

[abhigoku10]

@Tianxiaomo when can we expect the support code to train ??

[weidaolee]

我最近需要用在自己的data上，希望能參與開發。qq: 2379467558

[Tianxiaomo]

@chunleiml 应该可以训练

[ckcraig01]

@Tianxiaomo
Dear Author, thank you very much for the great work. I suppose you are working on training and reproducing the performance of original repo. When you are done, may you release the training flow? btw, Pytorch is great~ Thanks again

[17868380981]

sorry, I'm not the auther.

…

------------------&nbsp;原始邮件&nbsp;------------------ 发件人: "Craig Hsin"<notifications@github.com&gt;; 发送时间: 2020年5月14日(星期四) 上午7:18 收件人: "Tianxiaomo/pytorch-YOLOv4"<pytorch-YOLOv4@noreply.github.com&gt;; 抄送: "1109720378"<1109720378@qq.com&gt;; "Comment"<comment@noreply.github.com&gt;; 主题: Re: [Tianxiaomo/pytorch-YOLOv4] 怎么训练自己的数据呢 (#3) @Tianxiaomo Dear Author, thank you very much for the great work. I suppose you are working on training and reproducing the performance of original repo. When you are done, may you release the training flow? btw, Pytorch is great~ Thanks again — You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.

[Weifeng-Chen]

@chunleiml 应该可以训练

可以写一下怎么训练吗？ 有哪些需要改动的？

[okideal]

我魔改了一下可以支持自己的数据了，训练了几十个epoch后输出的框不太正确，作者大大能帮忙看下么 @Tianxiaomo

[Tianxiaomo]

什么问题，都不正确，还是有不正确的 @okideal

[okideal]

什么问题，都不正确，还是有不正确的 @okideal

都不正确，而且相邻epoch保存的checkpoint输出的结果都相差很大，框都是散布在图片上的，没有规律。

[okideal]

2020-05-19 09:41:24,363 train.py[line:317] DEBUG: Train step_252232: loss : 1.4532806873321533,loss xy : 12.427532196044922,loss wh : 1.475596308708191,loss obj : 6.721678256988525，loss cls : 2.627683162689209,loss l2 : 5.853010177612305
2020-05-19 09:41:28,438 train.py[line:317] DEBUG: Train step_252252: loss : 0.5830423831939697,loss xy : 4.975176811218262,loss wh : 0.122866690158844,loss obj : 4.1869707107543945，loss cls : 0.043664395809173584,loss l2 : 1.4445288181304932
2020-05-19 09:41:32,533 train.py[line:317] DEBUG: Train step_252272: loss : 1.1059558391571045,loss xy : 9.491128921508789,loss wh : 0.018527325242757797,loss obj : 1.4653393030166626，loss cls : 6.720297813415527,loss l2 : 2.798774480819702
2020-05-19 09:41:36,753 train.py[line:317] DEBUG: Train step_252292: loss : 0.7335749268531799,loss xy : 6.689797878265381,loss wh : 0.04984629526734352,loss obj : 4.740152359008789，loss cls : 0.2574022710323334,loss l2 : 1.7148456573486328
2020-05-19 09:41:40,661 train.py[line:317] DEBUG: Train step_252312: loss : 1.0617437362670898,loss xy : 14.04965877532959,loss wh : 0.08109669387340546,loss obj : 2.62939453125，loss cls : 0.22774934768676758,loss l2 : 0.8850184082984924
2020-05-19 09:41:44,627 train.py[line:317] DEBUG: Train step_252332: loss : 1.2092291116714478,loss xy : 8.662763595581055,loss wh : 0.019287779927253723,loss obj : 2.9575741291046143，loss cls : 7.708040237426758,loss l2 : 2.6127541065216064
2020-05-19 09:41:48,755 train.py[line:317] DEBUG: Train step_252352: loss : 0.8405617475509644,loss xy : 8.541584014892578,loss wh : 0.032777439802885056,loss obj : 4.75701904296875，loss cls : 0.11760733276605606,loss l2 : 1.3595856428146362
2020-05-19 09:41:52,804 train.py[line:317] DEBUG: Train step_252372: loss : 0.3742732107639313,loss xy : 4.4370269775390625,loss wh : 0.017715638503432274,loss obj : 1.478102445602417，loss cls : 0.055526264011859894,loss l2 : 0.43905097246170044
2020-05-19 09:41:56,729 train.py[line:317] DEBUG: Train step_252392: loss : 1.0849196910858154,loss xy : 11.690910339355469,loss wh : 0.05718810483813286,loss obj : 3.792856216430664，loss cls : 1.8177597522735596,loss l2 : 1.5996325016021729
训练的时候loss震荡也很大，这算正常的么？

[Tianxiaomo]

2020-05-19 09:41:24,363 train.py[line:317] DEBUG: Train step_252232: loss : 1.4532806873321533,loss xy : 12.427532196044922,loss wh : 1.475596308708191,loss obj : 6.721678256988525，loss cls : 2.627683162689209,loss l2 : 5.853010177612305
2020-05-19 09:41:28,438 train.py[line:317] DEBUG: Train step_252252: loss : 0.5830423831939697,loss xy : 4.975176811218262,loss wh : 0.122866690158844,loss obj : 4.1869707107543945，loss cls : 0.043664395809173584,loss l2 : 1.4445288181304932
2020-05-19 09:41:32,533 train.py[line:317] DEBUG: Train step_252272: loss : 1.1059558391571045,loss xy : 9.491128921508789,loss wh : 0.018527325242757797,loss obj : 1.4653393030166626，loss cls : 6.720297813415527,loss l2 : 2.798774480819702
2020-05-19 09:41:36,753 train.py[line:317] DEBUG: Train step_252292: loss : 0.7335749268531799,loss xy : 6.689797878265381,loss wh : 0.04984629526734352,loss obj : 4.740152359008789，loss cls : 0.2574022710323334,loss l2 : 1.7148456573486328
2020-05-19 09:41:40,661 train.py[line:317] DEBUG: Train step_252312: loss : 1.0617437362670898,loss xy : 14.04965877532959,loss wh : 0.08109669387340546,loss obj : 2.62939453125，loss cls : 0.22774934768676758,loss l2 : 0.8850184082984924
2020-05-19 09:41:44,627 train.py[line:317] DEBUG: Train step_252332: loss : 1.2092291116714478,loss xy : 8.662763595581055,loss wh : 0.019287779927253723,loss obj : 2.9575741291046143，loss cls : 7.708040237426758,loss l2 : 2.6127541065216064
2020-05-19 09:41:48,755 train.py[line:317] DEBUG: Train step_252352: loss : 0.8405617475509644,loss xy : 8.541584014892578,loss wh : 0.032777439802885056,loss obj : 4.75701904296875，loss cls : 0.11760733276605606,loss l2 : 1.3595856428146362
2020-05-19 09:41:52,804 train.py[line:317] DEBUG: Train step_252372: loss : 0.3742732107639313,loss xy : 4.4370269775390625,loss wh : 0.017715638503432274,loss obj : 1.478102445602417，loss cls : 0.055526264011859894,loss l2 : 0.43905097246170044
2020-05-19 09:41:56,729 train.py[line:317] DEBUG: Train step_252392: loss : 1.0849196910858154,loss xy : 11.690910339355469,loss wh : 0.05718810483813286,loss obj : 3.792856216430664，loss cls : 1.8177597522735596,loss l2 : 1.5996325016021729
训练的时候loss震荡也很大，这算正常的么？

数据增强原来有bug，另外学习率在训练时有没有调整

[okideal]

数据增强之前我把mosaic关掉了，只用了普通的增强。学习率的话，adam自己也会调整把，只是不会指数衰减

[moonlightian]

我魔改了一下可以支持自己的数据了，训练了几十个epoch后输出的框不太正确，作者大大能帮忙看下么 @Tianxiaomo

大佬可以分享一下魔改方法嘛，最近也想在bdd100k上训练一版pytorch-v4