Next: 0.999

[shelhamer]

The next release packages up 400+ commits by 9 authors. Thanks all!

Please upgrade your models! Caffe's proto definition was changed in #208 and #219 for extensibility. The upgrade_net_proto_binary.bin and upgrade_net_proto_text.bin tools are provided to convert current models. Caffe will attempt to automagically upgrade old models when loaded, but doesn't save the changes.

Update your Makefile.config! Caffe has a new Makefile and Makefile.config that learned to auto-configure themselves a bit better. Look at the new Makefile.config.example and update your configuration accordingly.

• Caffe learned to BLAS with ATLAS and OpenBLAS instead of MKL (though MKL is still supported too) MKL/non-MKL Reconciliation #165, Document BLAS install, tailor Makefile and Makefile.config to platform #305. Caffe with OpenBlas #314
• Caffe has a new, modular model definition Split monolithic LayerParameter into per-layer Protobuf Messages #208, Refactor LayerParameter into per-layer Parameter messages #219
• Forward compute loss Compute loss in the forward pass #209
• Improved python interface, solving from python, and initializing empty nets Improve python wrapper #311, Expose SGDSolver to pycaffe #286, Add unary CaffeNet constructor for uninitialized nets #365
• Improved MATLAB interface Improved matcaffe #223
• Makefile improvements Makefile improvements #277 + more
• Caffe RNG for all random numbers and solver seed for reproducibility Eliminate use of "rand" in code; enable deterministic results via solver seed #352
• Copyright assignment and contributors credit Standardize copyright, add root-level CONTRIBUTORS credit #255

Layers

• in-memory data layer Add a layer for in-memory datasets, and expose it to Python #294
• hinge-loss layer HingeLossLayer #303
• within-channel LRN layer Within-channel LRN layer #273
• hdf5 output layer Hdf5 output layer #252

Examples/Improvements/Fixes

• Unsupervised learning! MNIST autoencoder example MNIST autoencoder example #330
• double the protocol buffer size limit for large models Doubled Protobuf Bytes Limit #302
• create proto if needed when writing Write/create/truncate prototxt when saving to fix #341 #417
• fix non-square input processing Fix im2col nonsquare bug (#284) #298
• beginnings of layer documentation Layer definitions and declarations re-organization and documentation #376.
• speed up gradient tests Gradient check speedups #363
• randomized test order for better coverage Randomize order of test execution by make runtest #367
• share blobs/diffs, share parameters between training and test net add Share{Data,Diff} methods to blobs to enable "virtual" copies #318, add ShareTrainedLayersWith method and use for test net in solver #332
• more math functions Add more convenience math functions and all tests pass #201

...and more. Milestones to 1.0 #165, #208, and #311 are done.

DEPENDENCIES: Caffe's matrix and vector computations can be done with ATLAS, OpenBLAS, or MKL. The hard dependency on MKL is no more!

DEPRECATION: V0 model definitions. While Caffe will try to automagically upgrade old models when loaded, see tools/upgrade_net_proto* to make the permanent upgrade since this will be dropped.

[shelhamer]

[kloudkl]

The release notes or change logs need to be kept in the root directory of the project.