Document Caffe by a tour of key subjects

[shelhamer]

Lead a tour through the main points of how Caffe works and how to work with Caffe.

The Story

• nets, layers, and blobs
• forward / backward passes
• solving
• loss
• interfaces: cli, python, matlab
• figures

Details

• layer catalogue skeleton
• data, IO
• later: convolution
• later: add Net::Init to nets, layers, blobs
• later: DAGs
• later: weight sharing

Comments

• Doxygen generation
• add brief + TODO stubs for includes
• later: comment crusade

@Yangqing @jeffdonahue let's collaborate on my fork to write this all up and edit this description as you like. BVLC folk, feel free to help!

[kloudkl]

@borisgin, your Caffe summer workshop has a lot of good stuffs that fit here.

[borisgin]

I would be glad to help with tutorial. How I can contribute?

[kloudkl]

Maybe a new feature branch should be created and this PR tracks it so that more people can contribute.

[shelhamer]

@borisgin @kloudkl and all:

The first version of this documentation needs to be written by BVLC members. Once the first version is merged to dev everyone will of course be welcome to contribute.

Thank you for your understanding, and certainly for your enthusiasm!

[jeffdonahue]

I pushed a Doxygen config file (should be buildable with make docs if you're using Ubuntu and do sudo apt-get install doxygen; outputs to ./doxygen/) and changed much of the existing comment documentation to Doxygen format, and added a good amount of new documentation (especially for the loss layers and neuron layers, which I think all now have pretty latex forward/backward formulas). There's still quite a bit of work to do -- I think I only put in single-sentence documentation for all of {common,data,vision}_layers.hpp. @shelhamer @longjon @sergeyk it would be great if you guys could help with some of this.

Currently I have the documentation being served at http://durian3.banatao.berkeley.edu:8000/html/annotated.html if you'd like a preview (but I don't guarantee this will remain up at or for any particular amount of time), see http://durian3.banatao.berkeley.edu:8000/html/classcaffe_1_1SigmoidCrossEntropyLossLayer.html for what I think is reasonably good documentation -- ideally we'd pretty much use the same notation everywhere (I tried to do this but not sure I succeeded), so please try to follow the existing notation where applicable (or change it everywhere if it's really that bad) in any new documentation.

[bhack]

@jeffdonahue please take a look also to link in comments #934.

[kloudkl]

Pretty comprehensive and precise. Best practical document of deep learning ever seen. Can't wait to see all of them written up.

[kloudkl]

Hope you'll be able to publish the documentation before the ECCV'14 tutorial. Good luck!

[shelhamer]

@jeffdonahue the docs target and docs script are now configured to link doxygen's html output to docs/doxygen for publishing on the site.

[shelhamer]

Merging -- further documentation will follow. @Yangqing @longjon @sergeyk please join the doxygen comment crusade and use this commenting style in future code.