update recurrent branch with latest BVLC/master#9

Open

dribnet wants to merge 27 commits intojeffdonahue:recurrentfrom dribnet:recurrent-net-fix

+4,168-134

Commits on Mar 13, 2015

zero-init param diffs and accumulate gradients

longjon
authored and
jeffdonahue
committed
zero-init param diffs in gradient checker

longjon
authored and
jeffdonahue
committed
accumulate gradients in inner product layer

sguada
authored and
jeffdonahue
committed
accumulate gradients in (de)conv layers

longjon
authored and
jeffdonahue
committed
accumulate gradients in cudnn conv layer

longjon
authored and
jeffdonahue
committed
Add gpu_util.cuh, with caffe_gpu_atomic_add
jeffdonahue
committed
test_gradient_check_util: check_bottom < -1 only checks params
jeffdonahue
committed
Add EmbedLayer for inner products with sparse input (one-hot vectors),
jeffdonahue
committed
EmbedBackward with no loops -- use caffe_gpu_atomic_add instead
jeffdonahue
committed

Commits on Mar 25, 2015

Commits on Mar 26, 2015

RecurrentLayer bugfix: params still need backprop
jeffdonahue
committed

Commits on Apr 27, 2015

Resolved net.cpp conflict in deleted code
dribnet
committed