Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

关于ACFPN中上菜样和CxAM模块的几个问题 #3

Open
littletomatodonkey opened this issue Jun 1, 2020 · 4 comments
Open

关于ACFPN中上菜样和CxAM模块的几个问题 #3

littletomatodonkey opened this issue Jun 1, 2020 · 4 comments

Comments

@littletomatodonkey
Copy link

你好,感谢开源ACFPN,有2个问题想请教下

  1. 下面这个地方is_upsample应该一直设置为True?最开始P5和C4融合时,就应该对P5进行降采样(我看backbone部分只修改了dilation,应该没有对C5的分辨率进行修改?)

is_upsample = True

2.论文中的CxAM和CnAM模块中的1x1卷积部分,是否对通道数进行了缩减了呢?nonlocal中是对计算通道数减半,减少了计算量,不知道你是否有意愿开源这两个模块呢?

image

期待你的回复,多谢!

@shaohailin
Copy link

请问作者图中的F2 F3 F4 F5是对应Resnet的后面几个stage的输出结果吧?之前在别的论文里面看标注的是C2 C3 C4 C5,所以有点不确定,希望得到你的回复!感谢! @littletomatodonkey

@littletomatodonkey
Copy link
Author

littletomatodonkey commented Jun 3, 2020

是的,作者在这里是把C4->C5的dilation修改为了2(默认为1),但是我自己做实验的时候发现设置为1的话,效果会好一些。
这里是我的一个实现,不过是基于paddle的,不过暂时没有添加CxAM和CnAM模块,你可以看下 @shaohailin
https://github.com/PaddlePaddle/PaddleDetection/blob/master/configs/acfpn/README.md
https://github.com/PaddlePaddle/PaddleDetection/blob/master/ppdet/modeling/backbones/acfpn.py

@xiaoshuyun
Copy link

您好,打扰一下。我有几个问题想请教您一下。
1.请问论文中最终扩展卷积的输出与输入的上采样进行了concat,请问代码中是如何体现的,以及此处是对什么进行了上采样?
![批注 2020-06-15 235701](https://user-ima
ges.githubusercontent.com/56627406/84682297-05622880-af68-11ea-9feb-caff92a9d0d0.png)
2.我对caffe和paddle都不是很熟。我不太明白扩展卷积里这几个参数是什么意思。input_num, num1, num2,以及drop_out。【def _DenseAsppBlock(model, _input, prefix, input_num, num1, num2, dilation_rate, drop_out):】
期待您的回复。非常感谢。

@anonymoussss
Copy link

是的,作者在这里是把C4->C5的dilation修改为了2(默认为1),但是我自己做实验的时候发现设置为1的话,效果会好一些。
这里是我的一个实现,不过是基于paddle的,不过暂时没有添加CxAM和CnAM模块,你可以看下 @shaohailin
https://github.com/PaddlePaddle/PaddleDetection/blob/master/configs/acfpn/README.md
https://github.com/PaddlePaddle/PaddleDetection/blob/master/ppdet/modeling/backbones/acfpn.py

Hi,
I have been trying to reproduce this paper using mmdetection, I refer to this repository and your version of paddlepaddle . but when using the 1x learning schedule, the AP performance of Faster R-CNN w/Res50 has not been able to reach 38.5 (baseline is 37.4). I would like to ask, in the implementation of the CEM, what are the differences between it and DenseASPP besides Group Normalization?(I’m sorry I didn’t see the deformable convolution mentioned in the paper. Could you please point me?)

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment
Labels
None yet
Projects
None yet
Development

No branches or pull requests

4 participants