thuml/san

Code release for "Partial Transfer Learning with Selective Adversarial Networks" (CVPR 2018)

SAN

SAN Library

This is the code release for "Partial Transfer Learning with Selective Adversarial Networks" (CVPR 2018)

The caffe version is in directory "caffe". Details of the codes are described in the README.md in "caffe" directory. Notes that the performance of SAN in the paper are achieved by the codes of Caffe framework.

The pytorch version is in directory "pytorch". We have released the version test on PyTorch Version 0.3.1. Details of the codes are described in the README.md in "pytorch" directory.

Citation

If you use this code for your research, please consider citing:

    @InProceedings{Cao_2018_CVPR,
      author = {Cao, Zhangjie and Long, Mingsheng and Wang, Jianmin and Jordan, Michael I.},
      title = {Partial Transfer Learning With Selective Adversarial Networks},
      booktitle = {The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
      month = {June},
      year = {2018}
    }

Contact

If you have any problem about our code, feel free to contact

• [email protected]
• [email protected]

or describe your problem in Issues.

san's Issues

question about the partial transfer in DAN

Hey, I am running the partial transfer codes based on DAN ,and the dataset is Office31. The src domain is amazon and the target domain is webcam. But I got the acc is nearly to 0 instead of the acc result showed in the 'Partial Transfer Learning with Selective Adversarial Networks' paper which is 56.52. I feel quite confuse about the amazing result. Can you help me ?

the SAN branch consume a large amount of memory

Hi,
When I training san on imagenet, which has 1000 class numbers of source target . And the base network before san has 256 outputs. The 'ad_net' branch was LittleAdversarialNetwork. However, the ad_net branch would consume a large amount of GPU memory, which has 1000 fully connected branch with 256 inputs and 1 outputs. How do you cope with this problem?

Best Regards.

The VGG16 used in experiment

Hello. I wonder the VGG16 you used in the experiment of office 31-10. As I know, the pretrained VGG16 models are different in Caffe and PyTorch and also have different accuracies. Could you tell me which VGG16 is used as your base model? Thx!

Implementations Entropy-Minimization loss

Hi, Thanks for you code sharing.
Which one is about the Entropy-Minimization loss in loss.py ? EntropyLoss or SAN loss?

about the class of the mission on Caltech256→Office10

I'm trying to do the experiment on the mission of the Caltech256→Office10 ,however,after my first few experiments on C256→W10, the result is far away from the report result.And then I checked the classes of the C256,I realize some of the classes in C256 is a little bit confused,such as the "mountain-bike"and "touring-bike",they are all belong to the "bike" class in office31 dataset,the same situation happened in "caffee mug" and "beer mug",How do your guys deal with it? Can you provide the txt file of the mission of the Caltech256→Office10?