chengy12 / apnet Goto Github PK

Hi,
Thanks for sharing the code!!!
Why did I test with MSMT17 and only get map data, but not the generated images.
Generate a test image that corresponds to that part of the code.

Hi,
Thanks for sharing the code！
I`m curious about the code of spatial attention(APNet-S). Would you mind sharing it, and this is my email address if it's convenient，[email protected]

作者您好，感谢分享，我有几个问题想请教一下：

1、在SAM中，splite 切分的时候是切了通道数，splited = torch.split(x, channel//self.radix, dim=1)，而图2画的是将上下切分，或者切4份？

2、切分完，做注意力，2层的时候，level=2注意力只是进行了softmax处理，atten = F.softmax(atten, dim=1).view(batch, -1, 1, 1)，level=1 注意力用了SEnet？

3、按照图2，先全局，后逐步细粒度，但是代码里为啥顺序是相反的？
x = self.base_1(x)
if self.level > 2:
x = self.att_ss1(x)
x = self.BN_1(x)
if self.level > 1:
x = self.att_s1(x)
x = self.BN1(x)
if self.level > 0:
y = self.att1(x)
x = x*y.expand_as(x)

麻烦了~

There is only one SAMS layer when ap_level = 2 accoding to baseline.py. Is that consistent with paper?

作者您好！请问如果我想测试APlevel=3时的效果 除了将APNET： LEVEL参数改成了3之外还要进行什么其他的改动吗？
我在下载了您的代码之后 将configs/msmt.yml中的APNET： LEVEL参数改成了3 但是程序却报错：RuntimeError: running_mean should contain 63 elements not 64

cfg.TEST.WEIGHT 请教一下，这个路径怎么填，怎么挑选最好的模型呢？

Hello. Is the accuracy reported in the article the best accuracy? Because every time the network is trained, the accuracies are slightly different

Hello, how can you pick out your "attention pyramid module" and add it to other networks？ Is it the SAM part of the code？

Hi,
Thanks for sharing the code!!!
I run the experiments on Matket and Duke, and the results are similar to that in your paper. But the experiment result of MSMT is different from that in your paper. The map and rank1 of MSMT are 58.7% and 80.4% in my experiment. But, in your paper, the map and rank1 of MSMT are 63.5% and 83.7%. Maybe the experiment setup on MSMT is different from other datasets？

First of all thank you very much for your outstanding contribution！
In the forword of APNet/modeling/baseline.py, I found that the code is first split into two trainings, and finally after a se, which does not match the description in Figure 2 in the text (first horizontal training as a whole, and then cut into 2 training , Then cut into 4 points for training), which one prevails?

针对cuhk03这个数据集，我一直跑不出您论文给出的结果，想请问下当时您对于该数据集的参数是怎么设置的。举个例子：论文中cuhk03(l):mAP 85.3%; R-1 87.4% ，我测试的结果是 mAP 77.8%; R-1 80.9%。