出现更换新的backbone后loss不下降的问题

[mcmingchang]

我参考dlav的结构把vovnet和resnest加入进了训练，发现新的两个backbone在各种深度下Loss都不能很好地下降，训练coco数据集总的loss在8左右

[xingyizhou]

I haven't tried this yet. Here are two places need care: how do you initialize the new weights or load pretrained weights; how do you place BN layers.

[mcmingchang]

I built mobilenetv2 v3 in the same way, they all work well, I can provide the codes of resnest and vovnet to you.

[mcmingchang]

I tried to use dlav0_34 to train the birds, dogs, and cats of the coco dataset. The first training was very successful, but the second time Loss got stuck at 2.5. batch_size=12, lr=6.25e-5, input_shape=384*512.

[usergxx]

@mcmingchang if you use mobilenet v2,can you get a high map? i trained use resnet18 without pretrained, but i can not reach the result as high as show . until i used pretrained weights , i can get the result mAP， so i thiink the pretrained weight effect; is must used

[enhenghengheng]

我以相同的方式构建了mobilenetv2 v3，它们都运行良好，我可以为您提供resnest和vovnet的代码。

HI,How did you add Mobilenetv2,Can you share your backbone code?I'm a new bee.

[mcmingchang]

https://github.com/CaoWGG/Mobilenetv2-CenterNet

[usergxx]

@enhenghengheng i just trained resdcn_18 without pretrained, you can try it, i got only about 62%mAP without pretrained weights

[enhenghengheng]

@enhenghengheng i just trained resdcn_18 without pretrained, you can try it, i got only about 62%mAP without pretrained weights

OK，i will try it, and thanks.

[usergxx]

@enhenghengheng if you have result, please informe me , thanks

[enhenghengheng]

@enhenghengheng if you have result, please informe me , thanks

Ok,no problem,if I have a better result I will show it for you.

[JoJoliking]

@mcmingchang 请问你的问题解决了吗我准备也将换上vov。我之前尝试换过Re2Net的bottlenet以及用上一些小的tricker。效果良好。不足就是hmloss1点多以及idloss比较高。导致了整体损失偏高

[usergxx]

我测试了您发的  mobilenetv2版本 您把上采样替换了，在pascal voc的精度更低啊， map只有0.33

------------------ 原始邮件 ------------------ 发件人: "xingyizhou/CenterNet" <[email protected]>; 发送时间: 2020年7月27日(星期一) 下午4:06 收件人: "xingyizhou/CenterNet"<[email protected]>; 抄送: "葛栢林"<[email protected]>;"Comment"<[email protected]>; 主题: Re: [xingyizhou/CenterNet] 出现更换新的backbone后loss不下降的问题 (#776)

我以相同的方式构建了mobilenetv2 v3，它们都运行良好，我可以为您提供resnest和vovnet的代码。

HI,How did you add Mobilenetv2,Can you share your backbone code?I'm a new bee.

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe.