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.

==== Below is original caffe README, please follow the installation instructions to compile this code.

Created by Yangqing Jia, UC Berkeley EECS department. In active development by the Berkeley Vision and Learning Center (BVLC).

Caffe aims to provide computer vision scientists with a clean, modifiable implementation of state-of-the-art deep learning algorithms. Network structure is easily specified in separate config files, with no mess of hard-coded parameters in the code. Python and Matlab wrappers are provided.

At the same time, Caffe fits industry needs, with blazing fast C++/Cuda code for GPU computation. Caffe is currently the fastest GPU CNN implementation publicly available, and is able to process more than 20 million images per day on a single Tesla K20 machine *.

Caffe also provides seamless switching between CPU and GPU, which allows one to train models with fast GPUs and then deploy them on non-GPU clusters with one line of code: Caffe::set_mode(Caffe::CPU).

Even in CPU mode, computing predictions on an image takes only 20 ms when images are processed in batch mode.

• Installation instructions
• Caffe presentation at the Berkeley Vision Group meeting

* When measured with the SuperVision model that won the ImageNet Large Scale Visual Recognition Challenge 2012.

Caffe is BSD 2-Clause licensed (refer to the LICENSE for details).

The pretrained models published by the BVLC, such as the Caffe reference ImageNet model are licensed for academic research / non-commercial use only. However, Caffe is a full toolkit for model training, so start brewing your own Caffe model today!

Please kindly cite Caffe in your publications if it helps your research:

@misc{Jia13caffe,
  Author = {Yangqing Jia},
  Title = { {Caffe}: An Open Source Convolutional Architecture for Fast Feature Embedding},
  Year  = {2013},
  Howpublished = {\url{http://caffe.berkeleyvision.org/}
}

Tutorials and general documentation are written in Markdown format in the docs/ folder. While the format is quite easy to read directly, you may prefer to view the whole thing as a website. To do so, simply run jekyll serve -s docs and view the documentation website at http://0.0.0.0:4000 (to get jekyll, you must have ruby and do gem install jekyll).

We strive to provide provide lots of usage examples, and to document all code in docstrings. We'd appreciate your contribution to this effort!

Caffe is developed with active participation of the community by the Berkeley Vision and Learning Center. We welcome all contributions!

Our workflow is this:

• The dev branch is for new development, community contributions, and testing.
• The master branch is handled by BVLC, which will integrate changes from dev on a roughly monthly schedule.
• Do new development in feature branches with decriptive names.
• Bring your work up-to-date by rebasing onto the latest dev. (Polish your changes by interactive rebase, if you'd like.)
• Pull request your contribution to BVLC/caffe's dev branch for discussion and review.
  • PRs should live fast, die young, and leave a beautiful merge. Pull request sooner than later so that discussion can guide development.
  • Code must be accompanied by documentation and tests at all times.
  • Only fast-forward merges will be accepted.

See our development guidelines for further details–the more closely these are followed, the sooner your work will be merged.

Make the feature branch off of the latest bvlc/dev

git checkout dev
git pull upstream dev
git checkout -b feature
# do your work, make commits

Prepare to merge by rebasing your branch on the latest bvlc/dev

# make sure dev is fresh
git checkout dev
git pull upstream dev
# rebase your branch on the tip of dev
git checkout feature
git rebase dev

Push your branch to pull request it into dev

git push origin feature
# ...make pull request to dev...

Now make a pull request! You can do this from the command line (git pull-request -b dev) if you install hub.

The pull request of feature into dev will be a clean merge. Applause.

experiment-no-polish