Caffe Files

A lot has happened since the last release! This packages up ~800 commits by 119 authors. Thanks all!

With all releases one should do make clean && make superclean to clear out old materials before compiling the new release. - layers - batch normalization [#3229] [#3299] - scale + bias layers [#3591] - PReLU [#1940] [#2414], ELU [#3388], and log [#2090] non-linearities - tile layer [#2083], reduction layer [#2089] - embed layer [#2032] - spatial pyramid pooling [#2117] - batch reindex layer [#2966] - filter layer [#2054] - solvers: Adam [#2918], RMSProp [#2867], AdaDelta [#2782] - accumulate gradients to decouple computational and learning batch size [#1977] - de-duplicate solver code [#2518] - make solver type a string and split classes [#3166] -- you should update your solver definitions - MSRA [#1946] and bilinear interpolation [#2213] weight fillers - N-D blobs [#1970] and convolution [#2049] for higher dimensional data and filters - tools: - test caffe command line tool execution [#1926] - network summarization tool [#3090] - snapshot on signal / before quit [#2253] - report ignored layers when loading weights [#3305] - caffe command fine-tunes from multiple caffemodels [#1456] - pycaffe: - python net spec [#2086] [#2813] [#2959] - handle python exceptions [#2462] - python layer arguments [#2871] - python layer weights [#2944] - snapshot in pycaffe [#3082] - top + bottom names in pycaffe [#2865] - python3 compatibility improvements - matcaffe: totally new interface with examples and tests [#2505] - cuDNN: switch to v2 [#2038], switch to v3 [#3160], make v4 compatible [#3439] - separate IO dependencies for configurable build [#2523] - large model and solverstate serialization through hdf5 [#2836] - train by multi-GPU data parallelism [#2903] [#2921] [#2924] [#2931] [#2998] - dismantle layer headers so every layer has its own include [#3315] - workflow: adopt build versioning [#3311] [#3593], contributing guide [#2837], and badges for build status and license [#3133] - SoftmaxWithLoss normalization options [#3296] - dilated convolution [#3487] - expose Solver Restore() to C++ and Python [#2037] - set mode once and only once in testing [#2511] - turn off backprop by skip_propagate_down [#2095] - flatten layer learns axis [#2082] - trivial slice and concat [#3014] - hdf5 data layer: loads integer data [#2978], can shuffle [#2118] - cross platform adjustments [#3300] [#3320] [#3321] [#3362] [#3361] [#3378] - speed-ups for GPU solvers [#3519] and CPU im2col [#3536] - make and cmake build improvements - and more!

Fixes - [#2866] fix weight sharing to (1) reduce memory usage and computation (2) correct momentum and other solver computations - [#2972] fix concat (broken in [#1970]) - [#2964] [#3162] fix MVN layer - [#2321] fix contrastive loss layer to match Hadsell et al. 2006 - fix deconv backward [#3095] and conv reshape [#3096] (broken in [#2049]) - [#3393] fix in-place reshape and flatten - [#3152] fix silence layer to not zero bottom on backward - [#3574] disable cuDNN max pooling (incompatible with in-place) - make backward compatible with negative LR [#3007] - [#3332] fix pycaffe forward_backward_all() - [#1922] fix cross-channel LRN for large channel band - [#1457] fix shape of C++ feature extraction demo output

Dependencies: - hdf5 is required - cuDNN compatibility is now at v3 + v4 and cuDNN v1 and v2 are not supported - IO dependencies (lmdb, leveldb, opencv) are now optional [#2523]

:coffee: