much improved NN graph

.travis.yml

@@ -19,7 +19,8 @@ matrix:
 addons:
   apt:
     packages:
-    - libqt5gui5  # pyqt5>5.11 fails to load the xcb platform plugin without it
+      - libqt5gui5  # pyqt5>5.11 otherwise cannot load the xcb platform plugin
+      - libflann-dev
 
 install:
   - pip install -U --upgrade-strategy eager .[alldeps,test,doc]

README.rst

@@ -37,7 +37,7 @@ The documentation is available on
 `Read the Docs <https://pygsp.readthedocs.io>`_
 and development takes place on
 `GitHub <https://github.com/epfl-lts2/pygsp>`_.
-A (mostly unmaintained) `Matlab version <https://lts2.epfl.ch/gsp>`_ exists.
+A (mostly unmaintained) `Matlab version <https://epfl-lts2.github.io/gspbox-html>`_ exists.
 
 The PyGSP facilitates a wide variety of operations on graphs, like computing
 their Fourier basis, filtering or interpolating signals, plotting graphs,

doc/history.rst

@@ -24,6 +24,9 @@ History
 * New implementation of the Sensor graph that is simpler and scales better.
 * A new learning module with three functions to solve standard semi-supervised
   classification and regression problems.
+* A much improved, fixed, documented, and tested NNGraph. The user can now
+  select the backend and similarity kernel. The radius can be estimated and
+  features standardized. (PR #43)
 
 Experimental filter API (to be tested and validated):
 

doc/tutorials/optimization.rst

@@ -85,7 +85,7 @@ We start with the graph TV regularization. We will use the :class:`pyunlocbox.so
     >>> prob1 = pyunlocbox.solvers.solve([d, r, f], solver=solver,
     ...                                  x0=x0, rtol=0, maxit=1000)
     Solution found after 1000 iterations:
-        objective function f(sol) = 2.250584e+02
+        objective function f(sol) = 2.256055e+02
         stopping criterion: MAXIT
     >>>
     >>> fig, ax = G.plot(prob1['sol'])
@@ -107,7 +107,7 @@ This figure shows the label signal recovered by graph total variation regulariza
     >>> prob2 = pyunlocbox.solvers.solve([r, f], solver=solver,
     ...                                  x0=x0, rtol=0, maxit=1000)
     Solution found after 1000 iterations:
-        objective function f(sol) = 6.504290e+01
+        objective function f(sol) = 4.376481e+01
         stopping criterion: MAXIT
     >>>
     >>> fig, ax = G.plot(prob2['sol'])

pygsp/filters/filter.py

@@ -254,7 +254,7 @@ def filter(self, s, method='chebyshev', order=30):
         >>> _ = G.plot(s1, ax=axes[0])
         >>> _ = G.plot(s2, ax=axes[1])
         >>> print('{:.5f}'.format(np.linalg.norm(s1 - s2)))
-        0.26808
+        0.26995
 
         Perfect reconstruction with Itersine, a tight frame:
 
@@ -449,7 +449,7 @@ def estimate_frame_bounds(self, x=None):
         A=1.708, B=2.359
         >>> A, B = g.estimate_frame_bounds(G.e)
         >>> print('A={:.3f}, B={:.3f}'.format(A, B))
-        A=1.723, B=2.359
+        A=1.839, B=2.359
 
         The frame bounds can be seen in the plot of the filter bank as the
         minimum and maximum of their squared sum (the black curve):

pygsp/graphs/fourier.py

@@ -78,7 +78,7 @@ def coherence(self):
         >>> graph.compute_fourier_basis()
         >>> minimum = 1 / np.sqrt(graph.n_vertices)
         >>> print('{:.2f} in [{:.2f}, 1]'.format(graph.coherence, minimum))
-        0.88 in [0.12, 1]
+        0.91 in [0.12, 1]
         >>>
         >>> # Plot the most localized eigenvector.
         >>> import matplotlib.pyplot as plt

pygsp/graphs/nngraphs/bunny.py

@@ -34,7 +34,5 @@ def __init__(self, **kwargs):
             'distance': 8,
         }
 
-        super(Bunny, self).__init__(Xin=data['bunny'],
-                                    epsilon=0.02, NNtype='radius',
-                                    center=False, rescale=False,
+        super(Bunny, self).__init__(data['bunny'], kind='radius', radius=0.02,
                                     plotting=plotting, **kwargs)

pygsp/graphs/nngraphs/cube.py

@@ -10,12 +10,12 @@ class Cube(NNGraph):
 
     Parameters
     ----------
-    radius : float
-        Edge lenght (default = 1)
     nb_pts : int
         Number of vertices (default = 300)
     nb_dim : int
         Dimension (default = 3)
+    length : float
+        Edge length (default = 1)
     sampling : string
         Variance of the distance kernel (default = 'random')
         (Can now only be 'random')
@@ -35,20 +35,23 @@ class Cube(NNGraph):
     """
 
     def __init__(self,
-                 radius=1,
                  nb_pts=300,
                  nb_dim=3,
+                 length=1,
                  sampling='random',
                  seed=None,
                  **kwargs):
 
-        self.radius = radius
         self.nb_pts = nb_pts
         self.nb_dim = nb_dim
+        self.length = length
         self.sampling = sampling
         self.seed = seed
         rs = np.random.RandomState(seed)
 
+        if length != 1:
+            raise NotImplementedError('Only length=1 is implemented.')
+
         if self.nb_dim > 3:
             raise NotImplementedError("Dimension > 3 not supported yet!")
 
@@ -89,12 +92,10 @@ def __init__(self,
             'distance': 9,
         }
 
-        super(Cube, self).__init__(Xin=pts, k=10,
-                                   center=False, rescale=False,
-                                   plotting=plotting, **kwargs)
+        super(Cube, self).__init__(pts, k=10, plotting=plotting, **kwargs)
 
     def _get_extra_repr(self):
-        attrs = {'radius': '{:.2f}'.format(self.radius),
+        attrs = {'length': '{:.2e}'.format(self.length),
                  'nb_pts': self.nb_pts,
                  'nb_dim': self.nb_dim,
                  'sampling': self.sampling,

pygsp/graphs/nngraphs/imgpatches.py

@@ -10,7 +10,8 @@ class ImgPatches(NNGraph):
 
     Extract a feature vector in the form of a patch for every pixel of an
     image, then construct a nearest-neighbor graph between these feature
-    vectors. The feature matrix, i.e. the patches, can be found in :attr:`Xin`.
+    vectors. The feature matrix, i.e., the patches, can be found in
+    :attr:`features`.
 
     Parameters
     ----------
@@ -35,9 +36,10 @@ class ImgPatches(NNGraph):
     >>> from skimage import data, img_as_float
     >>> img = img_as_float(data.camera()[::64, ::64])
     >>> G = graphs.ImgPatches(img, patch_shape=(3, 3))
-    >>> print('{} nodes ({} x {} pixels)'.format(G.Xin.shape[0], *img.shape))
+    >>> N, d = G.patches.shape
+    >>> print('{} nodes ({} x {} pixels)'.format(N, *img.shape))
     64 nodes (8 x 8 pixels)
-    >>> print('{} features per node'.format(G.Xin.shape[1]))
+    >>> print('{} features per node'.format(d))
     9 features per node
     >>> G.set_coordinates(kind='spring', seed=42)
     >>> fig, axes = plt.subplots(1, 2)
@@ -83,16 +85,16 @@ def __init__(self, img, patch_shape=(3, 3), **kwargs):
         # Alternative: sklearn.feature_extraction.image.extract_patches_2d.
         #              sklearn has much less dependencies than skimage.
         try:
-            import skimage
+            from skimage.util import view_as_windows
         except Exception as e:
             raise ImportError('Cannot import skimage, which is needed to '
                               'extract patches. Try to install it with '
                               'pip (or conda) install scikit-image. '
                               'Original exception: {}'.format(e))
-        patches = skimage.util.view_as_windows(img, window_shape=window_shape)
-        patches = patches.reshape((h * w, r * c * d))
+        self.patches = view_as_windows(img, window_shape)
+        self.patches = self.patches.reshape((h * w, r * c * d))
 
-        super(ImgPatches, self).__init__(patches, **kwargs)
+        super(ImgPatches, self).__init__(self.patches, **kwargs)
 
     def _get_extra_repr(self):
         attrs = dict(patch_shape=self.patch_shape)