apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-18 13:55:14 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

[scg] minor adjustements based on original publication

Browse source
This commit is contained in:
mzwiessele 2014-06-18 08:52:27 -07:00
parent fdf8f984c1
commit 47d10600d1
1 changed files with 5 additions and 5 deletions
Download patch file Download diff file Expand all files Collapse all files

10
GPy/inference/optimization/scg.py
View file

@ -56,13 +56,13 @@ def SCG(f, gradf, x, optargs=(), maxiters=500, max_f_eval=np.inf, display=True,
if gtol is None: if gtol is None:
gtol = 1e-5 gtol = 1e-5
sigma0 = 1.0e-8 sigma0 = 1.0e-7
fold = f(x, *optargs) # Initial function value. fold = f(x, *optargs) # Initial function value.
function_eval = 1 function_eval = 1
fnow = fold fnow = fold
gradnew = gradf(x, *optargs) # Initial gradient. gradnew = gradf(x, *optargs) # Initial gradient.
if any(np.isnan(gradnew)): #if any(np.isnan(gradnew)):
raise UnexpectedInfOrNan, "Gradient contribution resulted in a NaN value" # raise UnexpectedInfOrNan, "Gradient contribution resulted in a NaN value"
current_grad = np.dot(gradnew, gradnew) current_grad = np.dot(gradnew, gradnew)
gradold = gradnew.copy() gradold = gradnew.copy()
d = -gradnew # Initial search direction. d = -gradnew # Initial search direction.
@ -168,13 +168,13 @@ def SCG(f, gradf, x, optargs=(), maxiters=500, max_f_eval=np.inf, display=True,
if Delta < 0.25: if Delta < 0.25:
beta = min(4.0 * beta, betamax) beta = min(4.0 * beta, betamax)
if Delta > 0.75: if Delta > 0.75:
beta = max(0.5 * beta, betamin) beta = max(0.25 * beta, betamin)
# Update search direction using Polak-Ribiere formula, or re-start # Update search direction using Polak-Ribiere formula, or re-start
# in direction of negative gradient after nparams steps. # in direction of negative gradient after nparams steps.
if nsuccess == x.size: if nsuccess == x.size:
d = -gradnew d = -gradnew
# beta = 1. # TODO: betareset!! beta = 1. # This is not in the original paper
nsuccess = 0 nsuccess = 0
elif success: elif success:
Gamma = np.dot(gradold - gradnew, gradnew) / (mu) Gamma = np.dot(gradold - gradnew, gradnew) / (mu)