thuml / DAN

Code release of "Learning Transferable Features with Deep Adaptation Networks" (ICML 2015)

Deep Adaptation Network (DAN)

This is a caffe repository for deep adaptation network (DAN). We fork the repository with version ID 29cdee7 from Caffe and make our modifications. The main modifications are listed as follow:

• Add mmd layer described in paper "Learning Transferable Features with Deep Adaptation Networks".
• Emit SOLVER_ITER_CHANGE message in solver.cpp when iter_ changes.

The value of the mmd loss could be negative since we used the linear-time unbiased estimate of the mmd, which lends us an O(n) time cost but may cause negative mmd values for some mini-batch. This negative mmd value however will not influence the final classification performance. Please refer to our paper for the technical details.

If you have any problem about this code, feel free to concact us with the following email:

• [e-mail]
• [e-mail]
• [e-mail]

Change Note

We have moved the implementations of Residual Transfer Network (NIPS '16) and Joint Adaptation Network (ICML '17) to the Xlearn library, which is our actively-maintained library for deep transfer learning.

Data Preparation

In data/office/*.txt, we give the lists of three domains in Office dataset.

Training Model

In models/DAN/amazon_to_webcam, we give an example model based on Alexnet to show how to transfer from amazon to webcam. In this model, we insert mmd layers after fc7 and fc8 individually.

The bvlc_reference_caffenet is used as the pre-trained model. If the Office dataset and pre-trained caffemodel are prepared, the example can be run with the following command:

"./build/tools/caffe train -solver models/DAN/amazon_to_webcam/solver.prototxt -weights models/bvlc_reference_caffenet/bvlc_reference_caffenet.caffemodel"

Resnet pre-trainded model is here. We use Resnet-50.

Parameter Tuning

In mmd-layer and jmmd-layer, parameter loss_weight can be tuned to give mmd loss different weights.

Citation

@inproceedings{DBLP:conf/icml/LongC0J15,
  author    = {Mingsheng Long and
               Yue Cao and
               Jianmin Wang and
               Michael I. Jordan},
  title     = {Learning Transferable Features with Deep Adaptation Networks},
  booktitle = {Proceedings of the 32nd International Conference on Machine Learning,
               {ICML} 2015, Lille, France, 6-11 July 2015},
  pages     = {97--105},
  year      = {2015},
  crossref  = {DBLP:conf/icml/2015},
  url       = {http://jmlr.org/proceedings/papers/v37/long15.html},
  timestamp = {Tue, 12 Jul 2016 21:51:15 +0200},
  biburl    = {http://dblp2.uni-trier.de/rec/bib/conf/icml/LongC0J15},
  bibsource = {dblp computer science bibliography, http://dblp.org}
}