search

lehaifeng / DASNet

135 stars 41 forks source link

About the thresh, f-score, and label data #19

Closed Luffy03 closed 3 years ago

Luffy03 commented 4 years ago

hello, I run your code and CDD dataset under the instruction but get bad results. My 'Best Thresh' is always 1, and F-score has no promotion. Could you please tell me why? Thank you very much! Whats' more, I am curious about the label. The range should be set to [0,1] or [0,255]. The initial data is [0,255], without normalizing to [0,1] I get weird loss just like hundreds. 截屏2020-07-10下午2 51 45

lehaifeng commented 4 years ago

HI Luffy. Would you please give us more about your setting on hyperparameters?

lehaifeng commented 4 years ago

当然

Luffy03 commented 4 years ago

HI Luffy. Would you please give us more about your setting on hyperparameters?

非常感谢您的工作！通过训练我已经用CDDdata得到您在论文中展示的结果。还有几个问题想向您请教 请问loss.py中所有的函数都是您在论文中展示的结构吗，使用某些loss时没法得到效果 另外请问您是否在较大尺度的图像上测试过？

zyp19 commented 4 years ago

@Luffy03，您好，请问您是否方便分享一下您的测试代码呢？十分感谢！！

Luffy03 commented 4 years ago

@Luffy03，您好，请问您是否方便分享一下您的测试代码呢？十分感谢！！

测试代码就是validate函数，我是在single_layer_similar_heatmap_visual里加几行代码将输出结果保存到路径

dsagsdfgfdh commented 4 years ago

label的值应该是0和1，我们测试了很多的loss函数，但是只有我们提出的函数得到了一个最优的结果，受限于实验设施，我们并没有在大尺寸上进行实验。

B-Xi commented 4 years ago

@Luffy03 I run this project followed the instructions, but the F-score is always low as about 0.003 during the training process. Additionally, when I directly use the 'CDD_model_best.pth' downloaded from 'https://pan.baidu.com/s/1GFkBXvVKgD1IqLYYeioX_w#list/path=%2F', Missing key(s) occur in the FC layers and embedding layers. Moreover, I don't know how to test the model. Have you gone through this project, would you please share your code with me? My email is 'xi_201401@163.com', thank you.

Luffy03 commented 4 years ago

@Luffy03 I run this project followed the instructions, but the F-score is always low as about 0.003 during the training process. Additionally, when I directly use the 'CDD_model_best.pth' downloaded from 'https://pan.baidu.com/s/1GFkBXvVKgD1IqLYYeioX_w#list/path=%2F', Missing key(s) occur in the FC layers and embedding layers. Moreover, I don't know how to test the model. Have you gone through this project, would you please share your code with me? My email is 'xi_201401@163.com', thank you.

I have not changed the code, you may just follow the author's command.

B-Xi commented 4 years ago

@Luffy03 I run this project followed the instructions, but the F-score is always low as about 0.003 during the training process. Additionally, when I directly use the 'CDD_model_best.pth' downloaded from 'https://pan.baidu.com/s/1GFkBXvVKgD1IqLYYeioX_w#list/path=%2F', Missing key(s) occur in the FC layers and embedding layers. Moreover, I don't know how to test the model. Have you gone through this project, would you please share your code with me? My email is 'xi_201401@163.com', thank you.

I have not changed the code, you may just follow the author's command.

I run the code following the author's instructions, but the loss is concussive, even negative, and the fscore is so low. As shown below: image Could you please tell me why did it happen and how to correct it?

ClarissaDHL commented 4 years ago

hello, I run your code and CDD dataset under the instruction but get bad results. My 'Best Thresh' is always 1, and F-score has no promotion. Could you please tell me why? Thank you very much! Whats' more, I am curious about the label. The range should be set to [0,1] or [0,255]. The initial data is [0,255], without normalizing to [0,1] I get weird loss just like hundreds. 截屏2020-07-10下午2 51 45

Hello, I encountered the same problem as you, how did you solve it at that time. And did u normalize the imgs?

Luffy03 commented 4 years ago

hello, I run your code and CDD dataset under the instruction but get bad results. My 'Best Thresh' is always 1, and F-score has no promotion. Could you please tell me why? Thank you very much! Whats' more, I am curious about the label. The range should be set to [0,1] or [0,255]. The initial data is [0,255], without normalizing to [0,1] I get weird loss just like hundreds. 截屏2020-07-10下午2 51 45

Hello, I encountered the same problem as you, how did you solve it at that time. And did u normalize the imgs?

no i don't use normalize

ClarissaDHL commented 4 years ago

thank u for ur reply. I have already solved the problem. There is no need to normalize the input imgs but need to normalize the label from [0,255] to [0,1].

------------------ 原始邮件 ------------------ 发件人: "lehaifeng/DASNet" <notifications@github.com>; 发送时间: 2020年9月17日(星期四) 上午10:19 收件人: "lehaifeng/DASNet"<DASNet@noreply.github.com>; 抄送: "185221369"<185221369@qq.com>;"Comment"<comment@noreply.github.com>; 主题: Re: [lehaifeng/DASNet] About the thresh, f-score, and label data (#19)

hello, I run your code and CDD dataset under the instruction but get bad results. My 'Best Thresh' is always 1, and F-score has no promotion. Could you please tell me why? Thank you very much! Whats' more, I am curious about the label. The range should be set to [0,1] or [0,255]. The initial data is [0,255], without normalizing to [0,1] I get weird loss just like hundreds.

Hello, I encountered the same problem as you, how did you solve it at that time. And did u normalize the imgs?

no i don't use normalize

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