def testGradient(self, X, y):
'''
Tests the analytical gradient computation by comparing it with the numerical gradients
Arguments
X : data matrix the form [input dim., number of samples]
y : labels in the form [1, number of samples].
Returns
result : 0 if passed, -1 if failed
'''
assert self.isInitialized, 'ERROR:DeepNetwork:testGradient: The instance is not properly initialized'
if self.debug:
print 'DEBUG:DeepNetwork:testGradient: Testing gradient computation...'
result = 0
theta_list = self.getNetworkParameters()
theta = self.unstackParameters(theta_list)
grad = self.computeGradient(theta, X, y)
numGrad = AuxFunctions.computeNumericalGradient(func=self.computeCost,
params=theta,
args=((X, y)))
errorGrad = np.sqrt(np.sum((grad - numGrad)**2))
if errorGrad < 1e-4:
if self.debug:
print 'DEBUG:DeepNetwork:testGradient:Gradient error: ', errorGrad
print 'DEBUG:DeepNetwork:testGradient:Gradient check PASSED!'
print
result = 0
else:
if self.debug:
print 'DEBUG:DeepNetwork:testGradient:Gradient error: ', errorGrad
print 'DEBUG:DeepNetwork:testGradient:Gradient check FAILED!'
print
result = -1
return result
def testGradient(self, X, y):
'''
Tests the analytical gradient computation by comparing it with the numerical gradients
Arguments
X : data matrix the form [number of parameters, number of samples]
y : labels in the form [1, number of samples]
Returns
result : 0 if passed, -1 if failed
'''
assert self.isInitialized, 'ERROR:Linreg:testGradient: The instance is not properly initialized'
assert X.shape[
0] == self.nParams, 'ERROR:Linreg:testGradient: Dimensions of given data do not match with the number of parameters'
if self.debug:
print 'DEBUG:Linreg:testGradient: Testing gradient computation... '
result = 0
grad = self.computeGradient(self.theta, X, y)
numGrad = AuxFunctions.computeNumericalGradient(func=self.computeCost,
params=self.theta,
args=(X, y))
errorGrad = np.sqrt(np.sum((grad - numGrad)**2))
if errorGrad < 1e-4:
if self.debug:
print 'DEBUG:Linreg:testGradient: Gradient error: ', errorGrad
print 'DEBUG:Linreg:testGradient: Gradient check PASSED!'
print
result = 0
else:
if self.debug:
print 'DEBUG:Linreg:testGradient: Gradient error: ', errorGrad
print 'DEBUG:Linreg:testGradient: Gradient check FAILED!'
print
result = -1
return result
def testGradient(self, X, y): ''' Tests the analytical gradient computation by comparing it with the numerical gradients Arguments X : data matrix the form [input dim., number of samples] y : labels in the form [1, number of samples] Returns result : 0 if passed, -1 if failed ''' assert self.isInitialized, 'ERROR:CNN:testGradient: The instance is not properly initialized' if self.debug: print 'DEBUG:CNN:testGradient: Testing gradient computation...' result = 0; [weights, biases] = self.getNetworkParameters(); params = self.rollParameters(weights, biases); grad = self.computeGradient(params, X, y); numGrad = AuxFunctions.computeNumericalGradient( func=self.computeCost, params=params, args=((X, y)) ); errorGrad = np.sqrt(np.sum((grad - numGrad)**2)); if errorGrad<1e-4: if self.debug: print 'DEBUG:CNN:testGradient:Gradient error: ', errorGrad print 'DEBUG:CNN:testGradient:Gradient check PASSED!' print result = 0; else: if self.debug: print 'DEBUG:CNN:testGradient:Gradient error: ', errorGrad print 'DEBUG:CNN:testGradient:Gradient check FAILED!' print result = -1; return result
def testGradient(self, X):
'''
Tests the analytical gradient computation by comparing it with the numerical gradients
Arguments
X : data matrix the form [input dim., number of samples]
Returns
result : 0 if passed, -1 if failed
'''
assert self.isInitialized, 'ERROR:SparseCoding:testGradient: The instance is not properly initialized'
if self.debug:
print 'DEBUG:SparseCoding:testGradient:Checking weight gradient...'
result = 0
grad = self.computeWeightGradient(self.weights_vec, self.features_vec,
X)
numGrad = AuxFunctions.computeNumericalGradient(
func=self.computeWeightCost,
params=self.weights_vec,
args=(self.features_vec, X))
errorGrad = np.sqrt(np.sum((grad - numGrad)**2))
if errorGrad < 1e-4:
if self.debug:
print 'DEBUG:SparseCoding:testGradient:Gradient error: ', errorGrad
print 'DEBUG:SparseCoding:testGradient:Gradient check PASSED!'
print
else:
if self.debug:
print 'DEBUG:SparseCoding:testGradient:Gradient error: ', errorGrad
print 'DEBUG:SparseCoding:testGradient:Gradient check FAILED!'
print
result = -1
if self.debug:
print 'DEBUG:SparseCoding:testGradient:Checking feature gradient...'
grad = self.computeFeatureGradient(self.features_vec, self.weights_vec,
X)
numGrad = AuxFunctions.computeNumericalGradient(
func=self.computeFeatureCost,
params=self.features_vec,
args=(self.weights_vec, X))
errorGrad = np.sqrt(np.sum((grad - numGrad)**2))
if errorGrad < 1e-4:
if self.debug:
print 'DEBUG:SparseCoding:testGradient:Gradient error: ', errorGrad
print 'DEBUG:SparseCoding:testGradient:Gradient check PASSED!'
print
else:
if self.debug:
print 'DEBUG:SparseCoding:testGradient:Gradient error: ', errorGrad
print 'DEBUG:SparseCoding:testGradient:Gradient check FAILED!'
print
result = -1
return result