Next: release candidater

[shelhamer]

A lot has happened since the last release! This packages up 400+ commits by 27 authors. Thanks all! A follow-up release will come in ~3 weeks to cover documentation, cuDNN v2, and as always more new features.

With all releases one should do make clean && make superclean to clear out old materials before compiling the new release to avoid inconsistencies.

• BVLC models for unrestricted use Open release of BVLC model weights #1650: the BVLC models including their weights are licensed for unrestricted use (already announced but included for completeness).
• New Development Workflow New Development Workflow: The One True Branch #1743 Documentation #1887: switch to a "one true branch" model in which all changes stem from master and releases are tagged. This is simpler and frees up maintainer time for real work.
• Layer Registry Layer factory. #1167 Give back to layer what is layer's, and to factory what is factory's #1270: the layer factory is now a registry to decentralize the inclusion of layers in the framework. Now REGISTER_LAYER_CLASS follows INSTANTIATE_CLASS.
• Pythonified Caffe + Python layer Reform the boost::python wrapper, including layers implemented in Python #1703: expose Layer and all Solver to Python and implement Layers in Python via the PythonLayer type. This changes the pycaffe interface.
• Give Phase to Net Give phase to Net #1790: phase is an instantiation argument to Net and immutable. This is the first step to dismantling the Caffe singleton and prevents Phase missteps like data_transformers applies jitter in TEST phase #1430. This changes the Net constructor and pycaffe and matcaffe interfaces.
• Dynamic Linking Dynamic Linking #1842: link the tools, tests, examples, and pycaffe dynamically. This makes the build faster and more incremental while reducing the compiled size (< 500 mb for both release and debug).
• CMake build Improved CMake scripts #1667: the CMake build, while unofficial, has been vastly improved and should work automatically in many cases. Feedback and follow-up improvements are welcome!
• Layer type is a string Layer type is a string #1694: eliminates the use of an enum for the LayerParameter type; adds a message ParamSpec unifying parameter-related proto fields for a single parameter (name, blobs_lr, weight_decay, ...) into a single structure. You should update your model definitions.
• ExpLayer EXP neuron layer #1204 adds a layer to compute the exp of the input
• Reshaping data Reshape single input batches for inputs of varying dimension #1313 Data and ImageData layers now reshape when the batch size == 1 to handle inputs of varying dimensions
• Encoded image inputs Allow using encoded images in Datum, LevelDB, LMDB #1239 Encoded cleanup #1878: save storage by storing + loading encoded bytes instead of decoded uint8 or float data.
• OS X 10.10 / Yosemite Support OS X Yosemite / 10.10 #1740: Makefile configured for Yosemite + CUDA 7 + libc++. No more libstdc++ torture and editing homebrew formulae -- if you have already installed Caffe by the old method then clear out the dependencies and do it the easy way.
• 1x1 conv. optimization Optimize 1x1 convolution for Network-in-Network style operation #1118: speed-up 1x1 convolution with the CAFFE engine by skipping im2col when it's the identity.
• make LMDB default default to lmdb for database storage #1128: concurrent access wins.
• and more!

Fixes

• Test phase jitter data_transformers applies jitter in TEST phase #1430: Random cropping was accidentally done at test time during training for models with cropping configured. Results of caffe test were unaffected. Fixed by Give phase to Net #1790.
• Pycaffe device / mode / phase Python can't set the device / phase for net initialization #1700: Python couldn't set the device, mode, or phase for proper init. since these methods were attached to caffe.Net. Fixed by exposing these methods at the module level in Give phase to Net #1790. Call caffe.set_{mode,device} before making a Net. This also solves issues with pycaffe and cuDNN.
• Image loading failures Die on inputs that fail to load #1837: ImageDataLayer would silently skip images it failed to load, but now it dies loudly.

Documentation: The follow-up release will concentrate on pycaffe + matcaffe examples and interface documentation.

• new installation guides are split up by platform
• the examples reflect the protobuf definition and interface changes in Layer type is a string #1694 and elsewhere

Dependencies: unchanged, except

• CUDA 7 is strongly suggested -- especially on OS X
• cuDNN compatibility is still at v1 but see cuDNN R2 #1731 for v2 release candidate compatibility -- the switch to v2 will come when cuDNN v2 is released

Interface Change / Deprecation

• the new protobuf model definition format is detailed in Layer type is a string #1694. Old models are currently automagically upgraded but you should upgrade with the included tools upgrade_net_proto_{text,binary}.
• layers are registered for layer factory Layer factory. #1167 Give back to layer what is layer's, and to factory what is factory's #1270 -- see the layer development guide
• Phase is now set by NetParameter::set_type or the Net(string file, Phase phase) constructor Give phase to Net #1790.
• Pycaffe phase, mode, and device handling are done by module methods or set at construction. Call caffe.set_{mode,device} before making a Net and include the phase as an argument as net = caffe.Net('def.prototxt', caffe.TRAIN for example.
• Pycaffe pre-processing is now done by caffe.Transformer -- refer to the bundled examples and python/caffe/io.py for details. See How to pycaffe #1774 for a suggestion on caffe.Net usage.

☕

[bhack]

We are interested in #1482 and #1484 for the next release. This PRs were opened in November and already partially reviewed. We are available to reserve working time to finalize that if you have other feedbacks. /cc @mtamburrano

[ducha-aiki]

@shelhamer
Actually, we have done some bets in our office if any layer from @bhack and Co would be ever merged (while some others branches are merged very quickly). Now stacks are 70 to 30 that they would be never merged :) No offence.

[mtamburrano]

@ducha-aiki I would like to bet 10 bucks against us, too ;)