Switch branches/tags
Nothing to show
Clone or download
Latest commit 065e498 Mar 16, 2016
Failed to load latest commit information.
cifar-float-extra sorry about compiling Oct 23, 2014
data/ilsvrc12 code-no-polish Sep 17, 2014
docs code-no-polish Sep 17, 2014
examples commit Mar 16, 2016
include/caffe code-no-polish Sep 17, 2014
matlab/caffe code-no-polish Sep 17, 2014
models code-no-polish Sep 17, 2014
python code-no-polish Sep 17, 2014
scripts code-no-polish Sep 17, 2014
src code-no-polish Sep 17, 2014
tools float Oct 23, 2014
INSTALL.md code-no-polish Sep 17, 2014
LICENSE Update LICENSE Jul 27, 2015
Makefile fix makefile bug, okay to compile Oct 23, 2014
Makefile.config code-no-polish Sep 17, 2014
Makefile.config.example code-no-polish Sep 17, 2014
README.md Update README.md May 15, 2015
caffe.cloc code-no-polish Sep 17, 2014


Deeply-supervised Nets

Update For experiments on cifar-100 and SVHN, please use the same architecture as CIFAR-10 provided. This architecture/hyper-parameter settings can generalize pretty well and can achieve the numbers reported in the paper.

Also for Cifar-100 and SVHN, we used softmax losses instead of hinge losses for the output supervision and deep supervision. There are known convergence issues of hinge loss on > 10 classes in Caffe.

Please cite DSN in your publications if it helps your research:

[1] Deeply-Supervised Nets Chen-Yu Lee*, Saining Xie*, Patrick Gallagher, Zhengyou Zhang, Zhuowen Tu (* indicates equal contributions) In Proceedings of AISTATS 2015

If you have problems reproducing the experiments feel free to contact the authors.

Deeply Supervised Nets This DSN code is based on an older version of CAFFE framework. This is for reproducing the results reported in our paper. Though with minimum engineering efforts you can apply this idea to your own code, as well as new network architectures.

We preprocess the data following the methods used in maxout networks and network in network paper. Please find the details here. Basically we only did GCN (global contrast normalization) on the benchmark datasets. Note that the scale of the data is [0,1] instead of [0, 255]. This is a tricky part when you use your own data: You should tune the learning rate accordingly. Also tools/cifar-float-data might be useful if you want to generate your own leveldb database from the gcn processed data.

To reproduce the results more easily, you can also download our processed Leveldb files here

The configuration files are in the examples folder. You can run train_full.sh script and it will automatically complete the training. The final result may vary from machine to machine. For cifar-10 you should be able to get at least 9.65% error, please contact me if you have problem in reproducing this. Thank you.