search

lixiaotong97 / DSU

[ICLR 2022] Official pytorch implementation of "Uncertainty Modeling for Out-of-Distribution Generalization" in International Conference on Learning Representations (ICLR) 2022.
139 stars 16 forks source link

The results of instance retrieval #5

Open Baymine opened 1 year ago

Baymine commented 1 year ago

Hi, thanks for sharing your code.

I run instance retrieval on my machine. On the first test, the mAP is 24.9%. But in the following test, the mAP gradually decreased. On the last test, the mAP is only 1.2%.

Do you have any ideas about why this phenomenon happened? Thank you very much!

GoodMan0 commented 1 year ago

@Baymine Hi, did you fix the problem? I have the same issue.

lixiaotong97 commented 1 year ago

Hi, could you please provide more information such as the running environment, training script, and the loss value for convergence. Moreover, did you modify anywhere of the code?

Best regards.

freshman97 commented 1 year ago

@Baymine Hi, did you fix the problem? I have the same issue.

Hi, I also get the same problem...

lixiaotong97 commented 1 year ago

There is a simple way to try to solve this. You can use the original repo MMT (https://github.com/yxgeee/MMT), and comment this line. Then, you can add DSU module into the resnet.py and should get similar results.

I run without problems in my environment, so if you find any reson to cause it, please kindly let me know.

Best regards.

feimo49 commented 1 year ago

@freshman97 Hi, did you fix the problem? I have the same issue.

lixiaotong97 commented 1 year ago

You can try the above solution.

freshman97 commented 1 year ago

@freshman97 Hi, did you fix the problem? I have the same issue.

Nope. After commenting on that line it seems we still have the decreasing mAP problem. But you can try the implementation setup of MixStyle and add the DSU module.