‹ projects

cluster-rnn

a distributed Torch7 RNN cluster over MPI
Log | Files | Refs | README

commit 7084b04ac1f4ea5603c116d665e6c4e48a8b4165
parent 4b160086c9780103cea871284305e4ece7cc117c
Author: umhau <umhau@users.noreply.github.com>
Date:   Wed, 15 Feb 2017 01:44:30 -0500

p

Diffstat:
Mmlaunch.lua | 8++++----
1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/mlaunch.lua b/mlaunch.lua @@ -24,13 +24,13 @@ local iterations = 10 -- i.e., epochs. don't need that many for testing. -- advanced parameters local communicationPeriod = 64 -local movingRateAlpha = .001 --0.9/6 -local learningRate = 1e-2 +local movingRateAlpha = 0.9/6 --.001 --0.9/6 +local learningRate = 5e-2 local momentum = 0.99 -- very advanced parameters -local learningRateDecay = 1e-2 -local learningRateDecayPower = .5 -- see pg. 23, fig 9 of associated paper +local learningRateDecay = 1e-4 +local learningRateDecayPower = 0.5 -- see pg. 23, fig 9 of associated paper -- there's other EAMSGD variables that can be tuned below. I'll do that later.