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

[opt] unified printing of status of optimization

Browse source
This commit is contained in:
Max Zwiessele 2015-01-15 09:02:45 +00:00
parent 4deac0103c
commit 57e941140c
4 changed files with 17 additions and 8 deletions
Download patch file Download diff file Expand all files Collapse all files

2
GPy/inference/optimization/optimization.py
View file

@ -37,7 +37,7 @@ class Optimizer():
self.x_opt = None
self.funct_eval = None
self.status = None
self.max_f_eval = int(max_f_eval)
self.max_f_eval = int(max_iters)
self.max_iters = int(max_iters)
self.bfgs_factor = bfgs_factor
self.trace = None

11
GPy/inference/optimization/scg.py
View file

@ -61,6 +61,7 @@ def SCG(f, gradf, x, optargs=(), maxiters=500, max_f_eval=np.inf, display=True,
function_eval = 1
fnow = fold
gradnew = gradf(x, *optargs) # Initial gradient.
function_eval += 1
#if any(np.isnan(gradnew)):
# raise UnexpectedInfOrNan, "Gradient contribution resulted in a NaN value"
current_grad = np.dot(gradnew, gradnew)
@ -96,6 +97,7 @@ def SCG(f, gradf, x, optargs=(), maxiters=500, max_f_eval=np.inf, display=True,
sigma = sigma0 / np.sqrt(kappa)
xplus = x + sigma * d
gplus = gradf(xplus, *optargs)
function_eval += 1
theta = np.dot(d, (gplus - gradnew)) / sigma
# Increase effective curvature and evaluate step size alpha.
@ -111,10 +113,10 @@ def SCG(f, gradf, x, optargs=(), maxiters=500, max_f_eval=np.inf, display=True,
fnew = f(xnew, *optargs)
function_eval += 1
# if function_eval >= max_f_eval:
# status = "maximum number of function evaluations exceeded"
# break
# return x, flog, function_eval, status
if function_eval >= max_f_eval:
status = "maximum number of function evaluations exceeded"
break
return x, flog, function_eval, status
Delta = 2.*(fnew - fold) / (alpha * mu)
if Delta >= 0.:
@ -156,6 +158,7 @@ def SCG(f, gradf, x, optargs=(), maxiters=500, max_f_eval=np.inf, display=True,
# Update variables for new position
gradold = gradnew
gradnew = gradf(x, *optargs)
function_eval += 1
current_grad = np.dot(gradnew, gradnew)
fold = fnew
# If the gradient is zero then we are done.