search

svishwa / crowdcount-mcnn

Single Image Crowd Counting via MCNN (Unofficial Implementation)
MIT License
499 stars 179 forks source link

关于学习率 #57

Open EillotY opened 4 years ago

EillotY commented 4 years ago

你好，我现在想利用这个模型提优，加入了trick之后，是有点提高，但是我觉得应该还有进步，现在想询问你的是：代码中设置的学习率很低，你是根据经验设置的嘛？还是自己寻找到的这个学习率，我想找到最优学习率，我现在能想到的是，慢慢的提高学习率这样，但是感觉还是会有问题，因为我还是不能知道什么时候才能得到一个好的学习率

ft-z commented 4 years ago

Hello, I am also looking at this code, can you please implement the detection after inserting a new picture? 你好，我也在看这个代码，请问你能实现新的图片插入后检测吗？

EillotY commented 4 years ago

Hello, I am also looking at this code, can you please implement the detection after inserting a new picture? 你好，我也在看这个代码，请问你能实现新的图片插入后检测吗？

you mean a interface to detect one image from reality? i think you should use the last.pt to create a model, this code does not have the function.

ft-z commented 4 years ago

Unfortunately, I didn't find the last.pt information you mentioned, then I will try to add the code to generate density map. Thank you for the reply!

很遗憾我没有找到你说的last.pt相关资料，那我将尝试加入生成密度图的代码。谢谢你的答复！

------------------ 原始邮件 ------------------ 发件人: "M.Yuan"<notifications@github.com>; 发送时间: 2020年4月15日(星期三) 晚上11:58 收件人: "svishwa/crowdcount-mcnn"<crowdcount-mcnn@noreply.github.com>; 抄送: "郑兴文"<546080688@qq.com>;"Comment"<comment@noreply.github.com>; 主题: Re: [svishwa/crowdcount-mcnn] 关于学习率 (#57)

Hello, I am also looking at this code, can you please implement the detection after inserting a new picture? 你好，我也在看这个代码，请问你能实现新的图片插入后检测吗？

you mean a interface to detect one image from reality? i think you should use the last.pt to create a model, this code does not have the function.

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

EillotY commented 4 years ago

soy, i find out it is the best.h5 file to create model. my bad

ft-z commented 4 years ago

thanks！

------------------ 原始邮件 ------------------ 发件人: "M.Yuan"<notifications@github.com>; 发送时间: 2020年4月16日(星期四) 凌晨0:11 收件人: "svishwa/crowdcount-mcnn"<crowdcount-mcnn@noreply.github.com>; 抄送: "郑兴文"<546080688@qq.com>;"Comment"<comment@noreply.github.com>; 主题: Re: [svishwa/crowdcount-mcnn] 关于学习率 (#57)

soy, i find out it is the best.h5 file to create model. my bad

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