Fixed optimization tolerances

psdr/opt/gn.py

@@ -78,7 +78,7 @@ def linesearch_armijo(f, g, p, x0, bt_factor=0.5, ftol=1e-4, maxiter=40, traject
 	return x, alpha, fx
 
 
-def gauss_newton(f, F, x0, tol=1e-5, tol_normdx=1e-12, 
+def gauss_newton(f, F, x0, tol=1e-10, tol_normdx=1e-12, 
 	maxiter=100, fdcheck=False, linesearch=None, verbose=0, trajectory=None, gnsolver = None):
 	r"""A Gauss-Newton solver for unconstrained nonlinear least squares problems.
 

psdr/polyridge.py

@@ -209,7 +209,7 @@ def __init__(self, degree, subspace_dimension, basis = 'legendre',
 
 
 	def fit(self, X, fX, U0 = None, **kwargs):
-		r""" Given samples, fit the polynomial ridge approximation
+		r""" Given samples, fit the polynomial ridge approximation.
 
 		Parameters
 		----------
@@ -220,9 +220,19 @@ def fit(self, X, fX, U0 = None, **kwargs):
 		
 		"""
 
-		self.m = X.shape[1]
-		fX = fX.flatten()
+		X = np.array(X)
+		fX = np.array(fX).flatten()	
 
+		assert X.shape[0] == fX.shape[0], "Dimensions of input do not match"
+
+		if U0 is not None:
+			U0 = np.array(U0)
+			assert U0.shape[0] == X.shape[1], "U0 has %d rows, expected %d based on X" % (U0.shape[0], X.shape[1])
+			assert U0.shape[1] == self.subspace_dimension, "U0 has %d columns; expected %d" % (U0.shape[1], self.subspace_dimension)
+			U0 = orth(U0)
+
+
+		# TODO Implement multiple initializations
 		if self.norm == 2:
 			return self._fit_2_norm(X, fX, U0, **kwargs)
 		else:

tests/test_polyridge.py

@@ -88,8 +88,9 @@ def test_exact():
 
 	# Random point
 	U, _ = np.linalg.qr(np.random.randn(m,n))
-	U, _ = np.linalg.qr(np.vstack([a,b]))	
-
+	# Actual ridge subspace
+	#U, _ = np.linalg.qr(np.vstack([a,b]).T)	
+
 	pra = PolynomialRidgeApproximation(degree = p, subspace_dimension = n, scale = False)
 	pra.fit(X, fX, U0 = U, verbose = 1)
 	# Because the data is an exact ridge function, we should (I think) converge to the global solution