apunkt/GPy
1
0
Fork 0
mirror of https://github.com/SheffieldML/GPy.git synced 2026-05-06 18:42:39 +02:00
Code Issues Projects Releases Packages Wiki Activity Actions

simplified the checkgrad logic somewhat

Browse source
This commit is contained in:
James Hensman 2013-02-04 12:36:08 +00:00
parent 9b69b04933
commit dacdaa1b41
1 changed files with 55 additions and 55 deletions
Download patch file Download diff file Expand all files Collapse all files

110
GPy/core/model.py
View file

@ -304,54 +304,62 @@ class model(parameterised):
return '\n'.join(s) return '\n'.join(s)
def checkgrad(self, verbose=False, include_priors=False, step=1e-6, tolerance = 1e-3, return_ratio=False, *args): def checkgrad(self, verbose=False, include_priors=False, step=1e-6, tolerance = 1e-3):
""" """
Check the gradient of the model by comparing to a numerical estimate. Check the gradient of the model by comparing to a numerical estimate.
If the overall gradient fails, invividual components are tested. If the verbose flag is passed, invividual components are tested (and printed)
:param verbose: If True, print a "full" checking of each parameter
:type verbose: bool
:param step: The size of the step around which to linearise the objective
:type step: float (defaul 1e-6)
:param tolerance: the tolerance allowed (see note)
:type tolerance: float (default 1e-3)
Note:-
The gradient is considered correct if the ratio of the analytical
and numerical gradients is within <tolerance> of unity.
""" """
x = self._get_params_transformed().copy() x = self._get_params_transformed().copy()
#choose a random direction to step in: if not verbose:
dx = step*np.sign(np.random.uniform(-1,1,x.size)) #just check the global ratio
dx = step*np.sign(np.random.uniform(-1,1,x.size))
#evaulate around the point x #evaulate around the point x
self._set_params_transformed(x+dx) self._set_params_transformed(x+dx)
f1,g1 = self.log_likelihood() + self.log_prior(), self._log_likelihood_gradients_transformed() f1,g1 = self.log_likelihood() + self.log_prior(), self._log_likelihood_gradients_transformed()
self._set_params_transformed(x-dx) self._set_params_transformed(x-dx)
f2,g2 = self.log_likelihood() + self.log_prior(), self._log_likelihood_gradients_transformed() f2,g2 = self.log_likelihood() + self.log_prior(), self._log_likelihood_gradients_transformed()
self._set_params_transformed(x) self._set_params_transformed(x)
gradient = self._log_likelihood_gradients_transformed() gradient = self._log_likelihood_gradients_transformed()
numerical_gradient = (f1-f2)/(2*dx) numerical_gradient = (f1-f2)/(2*dx)
global_ratio = (f1-f2)/(2*np.dot(dx,gradient)) global_ratio = (f1-f2)/(2*np.dot(dx,gradient))
if verbose:
print "Gradient ratio = ", global_ratio, '\n'
sys.stdout.flush()
if (np.abs(1.-global_ratio)<tolerance) and not np.isnan(global_ratio): if (np.abs(1.-global_ratio)<tolerance) and not np.isnan(global_ratio):
if verbose: return True
print 'Gradcheck passed' else:
return False
else: else:
if verbose: #check the gradient of each parameter individually, and do some pretty printing
print "Global check failed. Testing individual gradients\n" try:
names = self._get_param_names_transformed()
except NotImplementedError:
names = ['Variable %i'%i for i in range(len(x))]
try: # Prepare for pretty-printing
names = self._get_param_names_transformed() header = ['Name', 'Ratio', 'Difference', 'Analytical', 'Numerical']
except NotImplementedError: max_names = max([len(names[i]) for i in range(len(names))] + [len(header[0])])
names = ['Variable %i'%i for i in range(len(x))] float_len = 10
cols = [max_names]
# Prepare for pretty-printing cols.extend([max(float_len, len(header[i])) for i in range(1, len(header))])
header = ['Name', 'Ratio', 'Difference', 'Analytical', 'Numerical'] cols = np.array(cols) + 5
max_names = max([len(names[i]) for i in range(len(names))] + [len(header[0])]) header_string = ["{h:^{col}}".format(h = header[i], col = cols[i]) for i in range(len(cols))]
float_len = 10 header_string = map(lambda x: '|'.join(x), [header_string])
cols = [max_names] separator = '-'*len(header_string[0])
cols.extend([max(float_len, len(header[i])) for i in range(1, len(header))]) print '\n'.join([header_string[0], separator])
cols = np.array(cols) + 5
header_string = ["{h:^{col}}".format(h = header[i], col = cols[i]) for i in range(len(cols))]
header_string = map(lambda x: '|'.join(x), [header_string])
separator = '-'*len(header_string[0])
print '\n'.join([header_string[0], separator])
for i in range(len(x)): for i in range(len(x)):
xx = x.copy() xx = x.copy()
@ -369,24 +377,16 @@ class model(parameterised):
ratio = (f1-f2)/(2*step*gradient) ratio = (f1-f2)/(2*step*gradient)
difference = np.abs((f1-f2)/2/step - gradient) difference = np.abs((f1-f2)/2/step - gradient)
if verbose: if (np.abs(ratio-1)<tolerance):
if (np.abs(ratio-1)<tolerance): formatted_name = "\033[92m {0} \033[0m".format(names[i])
formatted_name = "\033[92m {0} \033[0m".format(names[i]) else:
else: formatted_name = "\033[91m {0} \033[0m".format(names[i])
formatted_name = "\033[91m {0} \033[0m".format(names[i]) r = '%.6f' % float(ratio)
r = '%.6f' % float(ratio) d = '%.6f' % float(difference)
d = '%.6f' % float(difference) g = '%.6f' % gradient
g = '%.6f' % gradient ng = '%.6f' % float(numerical_gradient)
ng = '%.6f' % float(numerical_gradient) grad_string = "{0:^{c0}}|{1:^{c1}}|{2:^{c2}}|{3:^{c3}}|{4:^{c4}}".format(formatted_name,r,d,g, ng, c0 = cols[0]+9, c1 = cols[1], c2 = cols[2], c3 = cols[3], c4 = cols[4])
grad_string = "{0:^{c0}}|{1:^{c1}}|{2:^{c2}}|{3:^{c3}}|{4:^{c4}}".format(formatted_name,r,d,g, ng, c0 = cols[0]+9, c1 = cols[1], c2 = cols[2], c3 = cols[3], c4 = cols[4]) print grad_string
print grad_string
if verbose:
print ''
if return_ratio:
return global_ratio
else:
return False
def EPEM(self,epsilon=.1,**kwargs): def EPEM(self,epsilon=.1,**kwargs):
""" """