question about the DAN and DDC result about transferlearning HOT 11 CLOSED

@yuzehui1996 When you download the office-31 dataset, the images are at the root webcam/images/. (for webcam dataset).
For running of the above codes, you'd better place the images at the root webcam/. Otherwise, you'd make the following changes in DDC.py:
source_name = "amazon"
target_name = "webcam"
--> source_name = "amazon/images"
target_name = "webcam/images"

from transferlearning.

Yeah, the max acc is always 100% in task webcam to dslr, and the max acc in the task with most methods such as DDC is also 100%. Hence, here I record the acc of the last epoch.

from transferlearning.

but when I change the src and target, the max acc is always 100%. And during the training , all the epochs show the acc is 100%.Why?

from transferlearning.

I don't know the reason... Might be the environment? I have run the code in several machines, and all of them are ok.

from transferlearning.

And during the training epochs, the soft loss (classification loss) is always 0. Is there any questions about the calculating the soft loss? I mean the log_softloss function in the DDC.py train(). Thanks a lot!

from transferlearning.

I have run the code again, and I still don't find any mistakes with it.

from transferlearning.

Thanks!I have run the codes successfully.But I have tried to remove the mmd loss from the final loss, and it get the higher acc than the loss contain mmd loss. I was confused with for quite a long time. Do you have any explanation?

from transferlearning.

First, you tried to remove the mmd loss from the final loss. This is not same as the ResNet setting in DAN paper, and the detail you could ask Mingsheng Long. Second, I have run as you say. The acc is higher than ResNet in DAN paper, however it is not higher than DAN. I don't konw the reason for your result, too.

from transferlearning.

Thanks a lot! And I am trying to remove 10 classes from the target domain and the num of classes are reduce to 21 classes. The num of the source domain class is 31. And I didn't change any more.With the mmd , I got the acc result is nearly to 0%. I feel quite confused about the result.

from transferlearning.

I don't know. This is a partial transfer proble, and you could ask the author of Partial Transfer Learning with Selective Adversarial for the detail.

from transferlearning.

Thanks any way!

from transferlearning.