Next: 0.9999

[shelhamer]

The next release packages up 500+ commits by 17 authors. Thanks all!

• (Specify net params in solver; log {Net,Solver} parameters; multiple test nets #404 from @jeffdonahue) the solver configuration can define the net inline; see lenet_consolidated_solver.prototxt.
• (Argmax layer #421, Threshold layer to binarize features #422 from @sguada) argmax and threshold layers
• (Changes in Matlab compiling behavior #428 from @shelhamer) make the detection example an R-CNN example
• (add option for lmdb #431 from @mavenlin) add support for lmdb storage; this allows multiple reads unlike leveldb.
• (Padding for Max Pooling #473 from @shelhamer) Padded max pooling
• (Improve Backward method interface #497 from @jeffdonahue) propagate_down in backward method is a vector and is correctly set for each input (fixes long-standing bug in EuclideanLossLayer)
• (Weight sharing #500 from @jeffdonahue) Parameters can be shared between layers / Weight sharing for RNNs, Siamese Nets, etc.
• (Non-square Filters and Separated Stride and Padding #505 from @shelhamer) ConvolutionLayer kernels, padding, and stride can be non-square
• (Add From/To options for Forward and Backward #518 from @longjon) Compute partial forward / backward passes
• (Top-k accuracy #531 from @robwhess) AccuracyLayer can compute the top k accuracy
• (CPU-only build #561 from @shelhamer) CPU-only Caffe / "Caffe coldbrew"
• (Rectangular pooling #614 from @ronghanghu) PoolingLayer kernels, padding, and stride can be non-square
• (Add split dim layer #680 from @bhack) SliceLayer to split a blob in two (or more) across channels or samples
• (Leaky ReLU #740 from @qipeng) Leaky ReLU: non-zero coefficient for negative ReLU inputs
• (All-in-one nets #734 from @jeffdonahue) Consolidated network definitions -- train and test nets can now be specified in a single prototxt
• (Print blob debug info during training if SolverParameter "debug_info" field is set #796 from @jeffdonahue) Print blob statistics during forward/backward passes and updates with SolverParameter setting: debug_info: true
• (Make training iterations 0-indexed to display training loss in 0th iter #800 from @jeffdonahue) Training iterations are 0-indexed
• (Consolidate tools. #789, Improve caffe tool #872 from @Yangqing, @shelhamer) Consolidated caffe tool for training, finetuning, test, etc. (adds a new dependency on gflags)
• (Small improvements to pycaffe #733, Improve and polish pycaffe #816 from @longjon, @shelhamer) Python interface improvements and bug fix for input preprocessing errors.
• (Announce caffe-users mailing list #825 from @shelhamer) introduce caffe-users mailing list

LICENSE: We have clarified the project license and copyright. While the terms are the same as before, they are now more clear.

DEPENDENCIES / PLATFORM: lmdb and gflags are now required. Instructions for Ubuntu 14.04 are included. CUDA 6.0 with driver version 340.* or later is recommended -- driver version 331.* cannot be used due to a memory bug in the driver identified by NVIDIA.

DEPRECATION: the new caffe tool includes commands for model training and testing, querying devices, and timing models. This tool produces output by default for more informative execution (and no need for GLOG_logtostderr environment variable). The corresponding train_net.bin, finetune_net.bin, test_net.bin, device_query.bin, and net_speed_benchmark.bin are deprecated.