Skip to content

Commit

Permalink
use functions instead of classes for encore cython files
Browse files Browse the repository at this point in the history 
This fixes a pickle error we have seen with recent versions of cython
and our mocked imports. cython now introduced special reduce and setstate
functions to enable cross python pickling for simple classes. It also tried to
do this for some encore classes that only had a run method and would overwrite
the `__call__` special method. So lets make them function right away.
  • Loading branch information
@kain88-de
kain88-de committed Sep 25, 2017
1 parent 76cd70c commit 2fe4bb8
Show file tree
Hide file tree
Showing 4 changed files with 121 additions and 138 deletions.
2 changes: 1 addition & 1 deletion package/MDAnalysis/analysis/encore/clustering/ClusteringMethod.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -148,7 +148,7 @@ def __call__(self, distance_matrix):
numpy.array
list of cluster indices
"""
clusters = affinityprop.AffinityPropagation().run(
clusters = affinityprop.AffinityPropagation(
s=distance_matrix * -1., # invert sign
preference=self.preference,
lam=self.damping,
Expand Down
169 changes: 81 additions & 88 deletions package/MDAnalysis/analysis/encore/clustering/affinityprop.pyx
View file
Original file line number Diff line number Diff line change
Expand Up @@ -34,108 +34,101 @@ cimport cython

@cython.boundscheck(False)
@cython.wraparound(False)

cdef class AffinityPropagation(object):
def AffinityPropagation(s, preference, float lam, int max_iterations, int convergence, int noise=1):
"""
Affinity propagation clustering algorithm. This class is a Cython wrapper around the Affinity propagation algorithm, which is implement as a C library (see ap.c). The implemented algorithm is described in the paper:
Clustering by Passing Messages Between Data Points.
Brendan J. Frey and Delbert Dueck, University of Toronto
Science 315, 972–976, February 2007
"""

def run(self, s, preference, float lam, int max_iterations, int convergence, int noise=1):
"""
Run the clustering algorithm.
Parameters
----------
s : encore.utils.TriangularMatrix object
Triangular matrix containing the similarity values for each pair of clustering elements. Notice that the current implementation does not allow for asymmetric values (i.e. similarity(a,b) is assumed to be equal to similarity(b,a))
preference : numpy.array of floats or float
Preference values, which the determine the number of clusters. If a single value is given, all the preference values are set to that. Otherwise, the list is used to set the preference values (one value per element, so the list must be of the same size as the number of elements)
lam : float
Floating point value that defines how much damping is applied to the solution at each iteration. Must be ]0,1]
max_iterations : int
Maximum number of iterations
convergence : int
Number of iterations in which the cluster centers must remain the same in order to reach convergence
noise : int
Whether to apply noise to the input s matrix, such there are no equal values. 1 is for yes, 0 is for no.
Parameters
----------
s : encore.utils.TriangularMatrix object
Triangular matrix containing the similarity values for each pair of clustering elements. Notice that the current implementation does not allow for asymmetric values (i.e. similarity(a,b) is assumed to be equal to similarity(b,a))
Returns
-------
preference : numpy.array of floats or float
Preference values, which the determine the number of clusters. If a single value is given, all the preference values are set to that. Otherwise, the list is used to set the preference values (one value per element, so the list must be of the same size as the number of elements)
elements : list of int or None
List of cluster-assigned elements, which can be used by encore.utils.ClustersCollection to generate Cluster objects. See these classes for more details.
lam : float
Floating point value that defines how much damping is applied to the solution at each iteration. Must be ]0,1]
"""
cdef int cn = s.size
cdef float cpreference = preference
max_iterations : int
Maximum number of iterations
# Assign preference values to diagonal
try:
for i in xrange(s.size):
s[i,i] = <float>preference[i]
except:
pass
convergence : int
Number of iterations in which the cluster centers must remain the same in order to reach convergence
if type(preference) == float:
for i in xrange(s.size):
s[i,i] = <float>preference
else:
raise TypeError ("Preference should be of type float")
noise : int
Whether to apply noise to the input s matrix, such there are no equal values. 1 is for yes, 0 is for no.
logging.info("Preference %3.2f: starting Affinity Propagation" % (preference))
# Prepare input and ouput arrays
cdef numpy.ndarray[numpy.float32_t, ndim=1] matndarray = numpy.ascontiguousarray(s._elements, dtype=numpy.float32)
cdef numpy.ndarray[long, ndim=1] clusters = numpy.zeros((s.size),dtype=long)
Returns
-------
# run C module Affinity Propagation
iterations = caffinityprop.CAffinityPropagation( <float*>matndarray.data, cn, lam, max_iterations, convergence, noise, <long*>clusters.data)
elements : list of int or None
List of cluster-assigned elements, which can be used by encore.utils.ClustersCollection to generate Cluster objects. See these classes for more details.
# Provide warning in case of lack of convergence
if iterations == 0:
logging.info("Preference %3.2f: could not converge in %d iterations" % (preference, -iterations))
import warnings
warnings.warn("Clustering with preference {0:3.2f} did not fully converge in {1:d} iterations".format(preference, -iterations))

# Find centroids
centroids = numpy.unique(clusters)
for k in numpy.arange(centroids.shape[0]):
ii = numpy.where(clusters == centroids[k])[0]
small_mat = numpy.zeros((ii.shape[0], ii.shape[0]))
for ii1 in numpy.arange(ii.shape[0]):
for ii2 in numpy.arange(ii.shape[0]):
small_mat[ii1,ii2] = s[ ii[ii1], ii[ii2] ]
j = numpy.argmax(numpy.sum(small_mat, axis=0))

centroids[k] = ii[j]

# Similarity to centroids
S_centroids = numpy.zeros((s.size, centroids.shape[0]))
for line in numpy.arange(s.size):
for c in numpy.arange(centroids.shape[0]):
S_centroids[line,c] = s[line, centroids[c]]

# Center values for each observation
c = numpy.argmax(S_centroids, axis=1)

# Centroids should point to themselves
c[centroids] = numpy.arange(centroids.shape[0])

# Assign centroid indices to all observables
clusters = centroids[c]

logging.info("Preference %3.2f: converged in %d iterations" % (preference, iterations))

return clusters
"""
cdef int cn = s.size
cdef float cpreference = preference

# Assign preference values to diagonal
try:
for i in xrange(s.size):
s[i,i] = <float>preference[i]
except:
pass

if type(preference) == float:
for i in xrange(s.size):
s[i,i] = <float>preference
else:
raise TypeError ("Preference should be of type float")

logging.info("Preference %3.2f: starting Affinity Propagation" % (preference))

# Prepare input and ouput arrays
cdef numpy.ndarray[numpy.float32_t, ndim=1] matndarray = numpy.ascontiguousarray(s._elements, dtype=numpy.float32)
cdef numpy.ndarray[long, ndim=1] clusters = numpy.zeros((s.size),dtype=long)

# run C module Affinity Propagation
iterations = caffinityprop.CAffinityPropagation( <float*>matndarray.data, cn, lam, max_iterations, convergence, noise, <long*>clusters.data)

# Provide warning in case of lack of convergence
if iterations == 0:
logging.info("Preference %3.2f: could not converge in %d iterations" % (preference, -iterations))
import warnings
warnings.warn("Clustering with preference {0:3.2f} did not fully converge in {1:d} iterations".format(preference, -iterations))

# Find centroids
centroids = numpy.unique(clusters)
for k in numpy.arange(centroids.shape[0]):
ii = numpy.where(clusters == centroids[k])[0]
small_mat = numpy.zeros((ii.shape[0], ii.shape[0]))
for ii1 in numpy.arange(ii.shape[0]):
for ii2 in numpy.arange(ii.shape[0]):
small_mat[ii1,ii2] = s[ ii[ii1], ii[ii2] ]
j = numpy.argmax(numpy.sum(small_mat, axis=0))

centroids[k] = ii[j]

# Similarity to centroids
S_centroids = numpy.zeros((s.size, centroids.shape[0]))
for line in numpy.arange(s.size):
for c in numpy.arange(centroids.shape[0]):
S_centroids[line,c] = s[line, centroids[c]]

# Center values for each observation
c = numpy.argmax(S_centroids, axis=1)

# Centroids should point to themselves
c[centroids] = numpy.arange(centroids.shape[0])

# Assign centroid indices to all observables
clusters = centroids[c]

logging.info("Preference %3.2f: converged in %d iterations" % (preference, iterations))

return clusters
2 changes: 1 addition & 1 deletion package/MDAnalysis/analysis/encore/dimensionality_reduction/DimensionalityReductionMethod.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -137,7 +137,7 @@ def __call__(self, distance_matrix):
"""
final_stress, coordinates = \
stochasticproxembed.StochasticProximityEmbedding().run(
stochasticproxembed.StochasticProximityEmbedding(
s=distance_matrix,
rco=self.distance_cutoff,
dim=self.dimension,
Expand Down
86 changes: 38 additions & 48 deletions package/MDAnalysis/analysis/encore/dimensionality_reduction/stochasticproxembed.pyx
View file
Original file line number Diff line number Diff line change
Expand Up @@ -36,78 +36,68 @@ cimport cython


@cython.embedsignature(True)

cdef class StochasticProximityEmbedding:
def StochasticProximityEmbedding(s, double rco, int dim, double maxlam, double minlam, int ncycle, int nstep, int stressfreq):
"""
Stochastic proximity embedding dimensionality reduction algorithm. The
algorithm implemented here is described in this paper:
Dmitrii N. Rassokhin, Dimitris K. Agrafiotis
A modified update rule for stochastic proximity embedding
Journal of Molecular Graphics and Modelling 22 (2003) 133–140
Dmitrii N. Rassokhin, Dimitris K. Agrafiotis
A modified update rule for stochastic proximity embedding
Journal of Molecular Graphics and Modelling 22 (2003) 133–140
This class is a Cython wrapper for a C implementation (see spe.c)
"""

def run(self, s, double rco, int dim, double maxlam, double minlam, int ncycle, int nstep, int stressfreq):
"""
Run stochastic proximity embedding.
Parameters:
----------
Parameters:
----------
s : encore.utils.TriangularMatrix object
Triangular matrix containing the distance values for each pair of
elements in the original space.
s : encore.utils.TriangularMatrix object
Triangular matrix containing the distance values for each pair of
elements in the original space.
rco : float
neighborhood distance cut-off
rco : float
neighborhood distance cut-off
dim : int
number of dimensions for the embedded space
dim : int
number of dimensions for the embedded space
minlam : float
final learning parameter
minlam : float
final learning parameter
maxlam : float
starting learning parameter
maxlam : float
starting learning parameter
ncycle : int
number of cycles. Each cycle is composed of nstep steps. At the end
of each cycle, the lerning parameter lambda is updated.
ncycle : int
number of cycles. Each cycle is composed of nstep steps. At the end
of each cycle, the lerning parameter lambda is updated.
nstep : int
number of coordinate update steps for each cycle
nstep : int
number of coordinate update steps for each cycle
Returns
-------
Returns
-------
space : (float, numpy.array)
float is the final stress obtained; the array are the coordinates of
the elements in the embedded space
space : (float, numpy.array)
float is the final stress obtained; the array are the coordinates of
the elements in the embedded space
stressfreq : int
calculate and report stress value every stressfreq cycle
stressfreq : int
calculate and report stress value every stressfreq cycle
"""

"""

cdef int nelem = s.size
cdef double finalstress = 0.0

logging.info("Starting Stochastic Proximity Embedding")
cdef int nelem = s.size
cdef double finalstress = 0.0

cdef numpy.ndarray[numpy.float64_t, ndim=1] matndarray = numpy.ascontiguousarray(s._elements, dtype=numpy.float64)
cdef numpy.ndarray[numpy.float64_t, ndim=1] d_coords = numpy.zeros((nelem*dim),dtype=numpy.float64)
logging.info("Starting Stochastic Proximity Embedding")

finalstress = cstochasticproxembed.CStochasticProximityEmbedding( <double*>matndarray.data, <double*>d_coords.data, rco, nelem, dim, maxlam, minlam, ncycle, nstep, stressfreq)
cdef numpy.ndarray[numpy.float64_t, ndim=1] matndarray = numpy.ascontiguousarray(s._elements, dtype=numpy.float64)
cdef numpy.ndarray[numpy.float64_t, ndim=1] d_coords = numpy.zeros((nelem*dim),dtype=numpy.float64)

logging.info("Stochastic Proximity Embedding finished. Residual stress: %.3f" % finalstress)
finalstress = cstochasticproxembed.CStochasticProximityEmbedding( <double*>matndarray.data, <double*>d_coords.data, rco, nelem, dim, maxlam, minlam, ncycle, nstep, stressfreq)

return (finalstress, d_coords.reshape((-1,dim)).T)
logging.info("Stochastic Proximity Embedding finished. Residual stress: %.3f" % finalstress)

def __call__(self, *args):
return self.run(*args)
return (finalstress, d_coords.reshape((-1,dim)).T)

0 comments on commit 2fe4bb8

Please sign in to comment.