apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-21 14:05:14 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

non verbose checkgrad adjusted to new system

Browse source
This commit is contained in:
Max Zwiessele 2014-02-14 09:57:18 +00:00
parent ffd09c7820
commit 8d7273b8ff
1 changed files with 15 additions and 13 deletions
Download patch file Download diff file Expand all files Collapse all files

28
GPy/core/model.py
View file

@ -403,32 +403,34 @@ class Model(Parameterized):
x = self._get_params_transformed().copy() x = self._get_params_transformed().copy()
if not verbose: if not verbose:
# just check the global ratio # make sure only to test the selected parameters
dx = step * np.sign(np.random.uniform(-1, 1, x.size))
# evaulate around the point x
f1 = self.objective_function(x + dx)
f2 = self.objective_function(x - dx)
gradient = self.objective_function_gradients(x)
if target_param is None: if target_param is None:
transformed_index = range(len(x)) transformed_index = range(len(x))
else: else:
transformed_index = self._raveled_index_for(target_param) transformed_index = self._raveled_index_for(target_param)
if self._has_fixes(): if self._has_fixes():
indices = np.r_[:self.size] indices = np.r_[:self.size]
which = (transformed_index[:,None]==indices[self._fixes_]).nonzero()[0] which = (transformed_index[:,None]==indices[self._fixes_][None,:]).nonzero()
indices -= (~self._fixes_).cumsum() transformed_index = (indices-(~self._fixes_).cumsum())[transformed_index[which[0]]]
transformed_index = indices[which]
if transformed_index.size == 0: if transformed_index.size == 0:
print "No free parameters to check" print "No free parameters to check"
return return
# just check the global ratio
dx = np.zeros_like(x)
dx[transformed_index] = step * np.sign(np.random.uniform(-1, 1, transformed_index.size))
# evaulate around the point x
f1 = self.objective_function(x + dx)
f2 = self.objective_function(x - dx)
gradient = self.objective_function_gradients(x)
dx = dx[transformed_index]
gradient = gradient[transformed_index] gradient = gradient[transformed_index]
numerical_gradient = (f1 - f2) / (2 * dx) numerical_gradient = (f1 - f2) / (2 * dx)
global_ratio = (f1 - f2) / (2 * np.dot(dx, np.where(gradient == 0, 1e-32, gradient))) global_ratio = (f1 - f2) / (2 * np.dot(dx, np.where(gradient == 0, 1e-32, gradient)))
return (np.abs(1. - global_ratio) < tolerance) or (np.abs(gradient - numerical_gradient).mean() < tolerance) return (np.abs(1. - global_ratio) < tolerance) or (np.abs(gradient - numerical_gradient).mean() < tolerance)
else: else:
# check the gradient of each parameter individually, and do some pretty printing # check the gradient of each parameter individually, and do some pretty printing
@ -457,7 +459,7 @@ class Model(Parameterized):
which = (param_index[:,None]==indices[self._fixes_][None,:]).nonzero() which = (param_index[:,None]==indices[self._fixes_][None,:]).nonzero()
param_index = param_index[which[0]] param_index = param_index[which[0]]
transformed_index = (indices-(~self._fixes_).cumsum())[param_index] transformed_index = (indices-(~self._fixes_).cumsum())[param_index]
print param_index, transformed_index #print param_index, transformed_index
else: else:
transformed_index = param_index transformed_index = param_index