Next: release candidate

[shelhamer]

The next release packages up 400+ commits by 18 authors. Thanks all!

• Document Caffe by a tour of key subjects #973 by @shelhamer, @jeffdonahue, @Yangqing, @longjon: documentation in both tutorial and developer API form.
• Standard format for Caffe models to enable easy distribution with forthcoming "model zoo" #917, [model zoo] download gist script #1110 by @sergeyk standard model format and model zoo clearinghouse for sharing models, with the NIN (network-in-network) and VGG Devil models coming soon.
• cuDNN acceleration #1046 by @shelhamer: cuDNN acceleration for 1.2-3x layer-wise speedups in many cases.
• Solver switching support & implementation of Nesterov's accelerated grad... #741 by @qipeng: add Adaptive Gradient (AdaGrad) and Nesterov's Accelerated Gradient (NAG) solvers while SGDSolver tests #855 by @jeffdonahue prepared tests for the standard SGD solver already included
• CMake build system (feature development tracking PR) #623 by @akosiorek and @bhack: CMake build system
• Loss generalization #686 by @jeffdonahue: generalization of loss allowing for multiple loss functions with different weights
• On-the-fly net resizing, without reallocation (where possible) #594 by @longjon: on-the-fly net input resizing that lets nets change their input dimensions in-place.
• Channel softmax #940 by @ronghanghu: channel-wise softmax for multiple outputs.
• mean-variance normalization layer #846 by @qipeng: mean-variance normalization layer.
• Elementwise max layer #688, rebase and fixup #688 from @to3i: elementwise max #1053 by @to3i, @jeffdonahue: elementwise max by EltwiseLayer operation.
• Refactor data layers to avoid duplication of data transformation code #954, Further simplify the data layers and extend the MemoryDataLayer #995 by @GeenuX, @kloudkl: simplify and de-duplicate data layers
• Replace pthread with boost::thread #1000 by @kloudkl: migrate to boost::thread
• various documentation and example improvements from the community -- thanks!

DOCUMENTATION: there is tutorial documentation and developer API documentation courtesy of Doxygen (thanks to Jeff!). The documentation, in particular the developer API, is still in-progress so come help and join the comment crusade!

DEPENDENCIES: CUDA 6.5 is the suggested version. cuDNN is an acceleration library for deep network operations with drop-in integration to Caffe. It is not required but suggested for best performance. See #1046.

DEPRECATION: transformation parameters now have their own configuration message to reduce duplication across the data layers. For instance

layers {
  name: "mnist"
  type: DATA
  top: "data"
  top: "label"
  data_param {
    source: "examples/mnist/mnist_train_lmdb"
    backend: LMDB
    batch_size: 64
  }
  transform_param {
    scale: 0.00390625
  }
}

is now the proper format with the transform_param block. Old models are currently automagically upgraded but you should upgrade with the included tools upgrade_net_proto_{text,binary}.