How to train the models

[Wangman1]

Hi, thank you for your code, and I want to know how to train the model in your code? thank you!

[lhwcv]

Sorry for my late reply. (A little busy at work)
My result is lower than in the paper, about ~2-3 percent lower than it. Maybe next week I'll push my training code.

[Wangman1]

Thank you for your reply. Looking forward to your open source！

…

------------------ 原始邮件 ------------------ 发件人: "lhwcv/mlsd_pytorch" ***@***.***>; 发送时间: 2021年7月13日(星期二) 中午11:57 ***@***.***>; ***@***.******@***.***>; 主题: Re: [lhwcv/mlsd_pytorch] How to train the models (#2) Sorry for my late reply. (A little busy at work) My result is lower than in the paper, about ~2-3 percent lower than it. Maybe next week I'll push my training code. — You are receiving this because you authored the thread. Reply to this email directly, view it on GitHub, or unsubscribe.

[lhwcv]

@Wangman1 Here is my result in mlsd_pytorch/README.md

[lhwcv]

model	img_size	sAP10
mlsd_tiny (this repo)	512	56.4
mlsd_tiny (in the paper)	512	58.0
mlsd_large (this repo)	512	59.6
mlsd_large (in the paper)	512	62.1

[Wangman1]

Wow, these results look great, and I can't wait to see how you train the model, so great job, thank you!!!

[Wangman1]

Hi, when I train the model in your mode, raise the follow error:

[Wangman1]

And the Google link to JSON file cannot download the content. It does not have relevant download permission.

[lhwcv]

@Wangman1

1. pip install yacs
2. use https://github.com/lhwcv/mlsd_pytorch/blob/main/data/wireframe_anno.zip
3. google cloud disk I'll check the issue later

[Wangman1]

@lhwcv
Thank you so much for your help. My code has been successfully run!!!

[ForestWang]

all loss will be nan after about 15 epoch, anyone knows why?

[YangTao516]

@lhwcv Thank you so much for your help. My code has been successfully run!!!

hi, could you please share the datasets with me ? It seems that the url for the datasets is unusable.

[harolddu]

@lhwcv

Thanks for your code and I can run the demo successfully. But loss values are both "nan" on wireframe and custom datasets during training, do you have any idea?

[michelebechini]

Hi, how long does it take (on average obviously) to train this model? I am trying using Colab GPUs and the process seems extremely slow...