ZhangLeUestc / transfer-caffe

caffe for transfer learning

transfer-caffe

This is a caffe repository for transfer learning. 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".
• Add entropy layer and outerproduct layer described in paper "Unsupervised Domain Adaptation with Residual Transfer Networks".
• Copy grl layer and messenger.hpp from repository Caffe.
• Emit SOLVER_ITER_CHANGE message in solver.cpp when iter_ changes.

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 is 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"

Parameter Tuning

In mmd-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}
}