def testElect(self):
result = mpi.elect()
self.assertLess(result, mpi.SIZE)
all_results = mpi.COMM.allgather(result)
self.assertEqual(len(set(all_results)), 1)
num_presidents = mpi.COMM.allreduce(mpi.is_president())
self.assertEqual(num_presidents, 1)
def testElect(self):
result = mpi.elect()
self.assertLess(result, mpi.SIZE)
all_results = mpi.COMM.allgather(result)
self.assertEqual(len(set(all_results)), 1)
num_presidents = mpi.COMM.allreduce(mpi.is_president())
self.assertEqual(num_presidents, 1)
def omp_n_maximize(X, labels, val, k):
"""Maximization of omp_n, with the given labels and vals.
Note that X is the local data hosted in each MPI node.
"""
dim = X.shape[1]
# G is the gram matrix of the activations
AtA_local = np.zeros((k, k))
AtX_local = np.zeros((k, dim))
A = None
for start in range(0, X.shape[0], _MINIBATCH):
end = min(start + _MINIBATCH, X.shape[0])
batchsize = end - start
if A is None:
A = np.zeros((batchsize, k))
else:
A[:] = 0
for i in range(batchsize):
A[i, labels[start + i]] = val[start + i]
AtA_local += mathutil.dot(A.T, A)
AtX_local += mathutil.dot(A[:batchsize].T, X[start:end])
AtA = np.empty_like(AtA_local)
AtX = np.empty_like(AtX_local)
mpi.COMM.Allreduce(AtA_local, AtA)
mpi.COMM.Allreduce(AtX_local, AtX)
# add a regularization term
isempty = np.diag(AtA) == 0
AtA.flat[:: k + 1] += 1e-8
centroids = np.ascontiguousarray(np.linalg.solve(AtA, AtX))
# let's deal with inactive guys
for i in range(k):
if isempty[i]:
# randomly restart one
centroids[i] = X[np.random.randint(X.shape[0])]
mpi.COMM.Bcast(centroids[i], root=mpi.elect())
scale = np.sqrt((centroids ** 2).sum(1)) + np.finfo(np.float64).eps
centroids /= scale[:, np.newaxis]
return centroids
def _m_step(X, z, k):
"""M step of the K-means EM algorithm
Computation of cluster centers/means
Parameters
----------
X: array, shape (n_samples, n_features)
z: array, shape (n_samples)
Current assignment
k: int
Number of desired clusters
Returns
-------
centers: array, shape (k, n_features)
The resulting centers
"""
dim = X.shape[1]
centers_local = np.zeros((k, dim))
counts_local = np.zeros(k, dtype = int)
centers = np.zeros((k, dim))
counts = np.zeros(k, dtype = int)
for q in range(k):
center_mask = np.flatnonzero(z == q)
counts_local[q] = len(center_mask)
if counts_local[q] > 0:
centers_local[q] = X[center_mask].sum(axis=0)
mpi.COMM.Allreduce(counts_local, counts)
mpi.COMM.Allreduce(centers_local, centers)
for q in range(k):
if counts[q] == 0:
centers[q] = X[np.random.randint(X.shape[0])]
mpi.COMM.Bcast(centers[q], root=mpi.elect())
counts[q] = 1
centers /= counts.reshape((centers.shape[0], 1))
return centers
def _m_step(X, z, k):
"""M step of the K-means EM algorithm
Computation of cluster centers/means
Parameters
----------
X: array, shape (n_samples, n_features)
z: array, shape (n_samples)
Current assignment
k: int
Number of desired clusters
Returns
-------
centers: array, shape (k, n_features)
The resulting centers
"""
dim = X.shape[1]
centers_local = np.zeros((k, dim))
counts_local = np.zeros(k, dtype=int)
centers = np.zeros((k, dim))
counts = np.zeros(k, dtype=int)
for q in range(k):
center_mask = np.flatnonzero(z == q)
counts_local[q] = len(center_mask)
if counts_local[q] > 0:
centers_local[q] = X[center_mask].sum(axis=0)
mpi.COMM.Allreduce(counts_local, counts)
mpi.COMM.Allreduce(centers_local, centers)
for q in range(k):
if counts[q] == 0:
centers[q] = X[np.random.randint(X.shape[0])]
mpi.COMM.Bcast(centers[q], root=mpi.elect())
counts[q] = 1
centers /= counts.reshape((centers.shape[0], 1))
return centers
def omp_n_maximize(X, labels, val, k):
'''Maximization of omp_n, with the given labels and vals.
Note that X is the local data hosted in each MPI node.
'''
dim = X.shape[1]
# G is the gram matrix of the activations
AtA_local = np.zeros((k, k))
AtX_local = np.zeros((k, dim))
A = None
for start in range(0, X.shape[0], _MINIBATCH):
end = min(start + _MINIBATCH, X.shape[0])
batchsize = end - start
if A is None:
A = np.zeros((batchsize, k))
else:
A[:] = 0
for i in range(batchsize):
A[i, labels[start + i]] = val[start + i]
AtA_local += mathutil.dot(A.T, A)
AtX_local += mathutil.dot(A[:batchsize].T, X[start:end])
AtA = np.empty_like(AtA_local)
AtX = np.empty_like(AtX_local)
mpi.COMM.Allreduce(AtA_local, AtA)
mpi.COMM.Allreduce(AtX_local, AtX)
# add a regularization term
isempty = (np.diag(AtA) == 0)
AtA.flat[::k + 1] += 1e-8
centroids = np.ascontiguousarray(np.linalg.solve(AtA, AtX))
# let's deal with inactive guys
for i in range(k):
if isempty[i]:
# randomly restart one
centroids[i] = X[np.random.randint(X.shape[0])]
mpi.COMM.Bcast(centroids[i], root=mpi.elect())
scale = np.sqrt((centroids**2).sum(1)) + np.finfo(np.float64).eps
centroids /= scale[:, np.newaxis]
return centroids
