def test_knn_isomap():
"""
Test of the isomap algorithm on knn graphs
"""
X = randn(10,3)
M = knn_Isomap(X, rdim=1)
u = M.train(k=2)
x = X[:2,:]
a = M.test(x)
eps = 1.e12
test = np.sum(a-u[:2,:])**2<eps
print np.sum(a-u[:2,:])**2
assert test
def test_knn_isomap():
"""
Test of the isomap algorithm on knn graphs
this one fails if the graph as more than one connected component
to avoid this we use a non-random example
"""
prng = np.random.RandomState(seed=2)
X = prng.randn(10,3)
M = knn_Isomap(X, rdim=1)
u = M.train(k=2)
x = X[:2,:]
a = M.test(x)
eps = 1.e-12
test = np.sum(a-u[:2,:])**2<eps
assert test
Bertrand Thirion, 2008-2010
"""
print __doc__
import numpy as np
from nipy.neurospin.eda.dimension_reduction import swiss_roll, isomap, CCA,\
mds, MDS, knn_Isomap , PCA
import nipy.neurospin.graph.graph as fg
import matplotlib.pylab as mp
# Generate a swiss roll datasets
# y are the 3D coordinates
# x are he 2D latent coordinates
nbsamp = 1000
y, x = swiss_roll(nbsamp)
M = knn_Isomap(y, rdim=3)
u = M.train(k=7)
# cheack that u and x span the same space,
# i.e. their two canonical coorelations are close to 1
sv = CCA(x-x.mean(0),u[:,:2])
print 'the canonical correlations between true and estimated parameters are %f, %f' %(sv[0],sv[1])
ax = M.G.show(u[:,:2])
ax.set_title('embedding of the data graph')
ax.axis('off')
mp.show()
