raoyongming / cal Goto Github PK

is:open 你好！代码拉取下来之后阅读readme 把数据集放在相应位置之后 就可以运行吗？我尝试这样，并未成功

I quickly checked the model script baseline.py and found that you used the cls_score as output when doing inference. I am wondering if your published results were generated by this instead of features before classifier (which is regularly applied in popular reid framework).

WS-DAN 使用的主干网络是inceptionV3，而CAL使用的是resnet101,后者性能明显比前者好，为什么论文中在不同数据集上对比时，没用也使用inceptionV3呢

Sorry to bother you, I have a question about the fgvc part. Why the normalized feature_matrix and feature_matrix_hat need to multiply 100 before the fc layer?

Hi, I read your paper and got some new experiences; I have a question about counterfactual attention, where you use random attention sampled from a uniform distribution U(0, 2). I wonder why not sampling from the Gaussian distribution from ubiquitous and intuitive considerations? Or is there a reference for this design?

hello~，你们的工作非常好。我看到

我想知道是不是推理阶段不用计算反事实的attention吗，只使用actual attention的那个分支来进行$\hat{y}$的输出吗

I am trying to reproduce the result of the CUB dataset, which is 90.6 acc ( table 1 in the paper). However, I use the same config and startup script as the code repo, but only get 90.03 acc for the last epoch. I notice that the total training epoch for fgvc task is not reported in the paper. So what is the proper epoch to get the 90.6 acc? Are there any other reasons that could affect reproducing the acc?

Please see attachment for my training log. Thanks
train.log
!

你好！请问能否提供FGVC任务的预训练模型呢？打扰了！

最后的损失的函数加了(y_effect,label)的交叉熵损失，不太懂这个原理，想请教一下。

Hello everyone :)
thanks for you nice work! Do you provide pre-trained weights for the person re-identification datasets somewhere?
Thanks in advance! :)

I want to run CAL without WS-DAN. How can I do so?

你好: 在CarDataset中第39行的cars_annos.mat是哪个文件呢？好像和readme中给出的Stanfordcar的结构不一致？

作者您好，我对您的论文进行了改进，train 准确率为 90.43%，但是运行infer.py 准确率为90.28%，想问下这是什么原因？

I've tried to train this code, but I'm still confused about understanding the code in the train_distributed.py file

In this coding, the aug variable uses the crop method, and the aux variable uses the drop method

But why does the aug_images variable use both methods (crop&drop), not just the crop method?

Thank you

Hello, may I know about the WSDAN_CAL architecture and its clearer visualisation? Because when I read the journal I am still quite confused, then I have also run and got the model results, but I am still quite confused to understand it. Thank you for your help, I am very happy to receive help from you.

@raoyongming 你好，请问可以出一个demo吗，检测一张图片的python脚本，谢谢

您好，请问能否分享一下做Visualization部分的代码？我使用了GradCAM来进行可视化，可一直存在BUG，可视化出来的结果很差，所以想请教一下您是怎么做可视化的？十分期待您的回复，谢谢！

作者你好，我阅读了您的文章，很受启发，并在GPU上复现了一下您的代码，发现代码中的batch_size=4，而论文里的batch_size=16，这个对准确率有影响吗

As above

请问车辆再识别的程序还会放上来嘛

When I run infer.py in fgvc, I get this error.

File "infer.py", line 100, in visualize
attention_maps = net.visualize(X)
File "/kaggle/working/CAL/fgvc/models/cal.py", line 180, in visualize
p = self.fc(feature_matrix * 100)
File "/opt/conda/lib/python3.7/site-packages/torch/nn/modules/module.py", line 1051, in _call_impl
return forward_call(*input, **kwargs)
File "/opt/conda/lib/python3.7/site-packages/torch/nn/modules/linear.py", line 96, in forward
return F.linear(input, self.weight, self.bias)
File "/opt/conda/lib/python3.7/site-packages/torch/nn/functional.py", line 1847, in linear
return torch._C._nn.linear(input, weight, bias)
TypeError: linear(): argument 'input' (position 1) must be Tensor, not tuple

Can you help me fix this? Thank you!

I am trying to reproduce the result as shown in the paper for the MSMT which is mAP@64% and [email protected]%; however, I could not do it. May I ask about the backbone you are using to get these results? Is it the same with the code in your repository or do you use different approach? And do you take the best case during training or the result after training for the whole 160 epoch? I am sorry if these questions bother you. Thank in advance!