directly normalize accumulated gradients
authorEvan Shelhamer <shelhamer@imaginarynumber.net>
Thu, 28 May 2015 19:43:29 +0000 (12:43 -0700)
committerEvan Shelhamer <shelhamer@imaginarynumber.net>
Thu, 28 May 2015 19:46:13 +0000 (12:46 -0700)
commit0e7a0785db224aa7cf2bd925d8b7910bdc3f7a98
treed59f1e917dd767d1fef622c99b2e94e6f40420f6
parent92ab737adad6d686ac75cdf934472f6a97b52fe7
directly normalize accumulated gradients

`SGDSolver::Normalize()` normalizes accumulated gradients by scaling
inversely to the accumulation as `1 / iter_size`.

This fixes accumulation for AdaGrad and is more obvious than fooling
with rates and decays in 55585f5.
include/caffe/solver.hpp
src/caffe/solver.cpp