def sw_sums(a, b):
abw = apply_scale(w, a, b)
np.divide(abw, 1 + abw, out = abw)
abw[np.isnan(abw)] = 1
swr = abw.sum(1, keepdims = True)
swc = abw.sum(0, keepdims = True)
return swr, swc
def sw_sums(a, b):
abw = apply_scale(w, a, b)
np.divide(abw, 1 + abw, out=abw)
abw[np.isnan(abw)] = 1
swr = abw.sum(1, keepdims=True)
swc = abw.sum(0, keepdims=True)
return swr, swc
def approximate_conditional_nll(A, w, sort_by_wopt_var = True):
"""Return approximate row/column-conditional NLL of binary matrix.
Return the approximate nll of an observed binary matrix given
specified Bernoulli weights, conditioned on having the observed
margins.
Inputs:
A: observed data, (m x n) binary matrix
w: weight matrix, (m x n) matrix with values in (0, +infty)
Output:
ncll: negative conditional log-likelihood
"""
assert(A.shape == w.shape)
r = A.sum(1, dtype=np.int)
c = A.sum(0, dtype=np.int)
r, c, arrays, _ = _prune(r, c, A, w)
A, w = arrays
# Sizing
m, n = len(r), len(c)
if (m == 0) or (n == 0):
return 0.0
# Sort the row margins (descending)
rndx = np.argsort(-r)
rsort = r[rndx]
# Balance the weights
a_scale, b_scale = canonical_scalings(w, r, c)
wopt = apply_scale(w, a_scale, b_scale)
if np.isnan(wopt).any():
wopt = w
# Reorder the columns
if sort_by_wopt_var:
cndx = np.lexsort((-wopt.var(0), c))
else:
cndx = np.argsort(c)
csort = c[cndx]
wopt = wopt[:,cndx]
# Compute G
G = _compute_G(r, m, n, wopt)
return _compute_cnll(A, r, rsort, rndx, csort, cndx, m, n, G)
def approximate_conditional_nll(A, w, sort_by_wopt_var=True):
"""Return approximate row/column-conditional NLL of binary matrix.
Return the approximate nll of an observed binary matrix given
specified Bernoulli weights, conditioned on having the observed
margins.
Inputs:
A: observed data, (m x n) binary matrix
w: weight matrix, (m x n) matrix with values in (0, +infty)
Output:
ncll: negative conditional log-likelihood
"""
assert (A.shape == w.shape)
r = A.sum(1, dtype=np.int)
c = A.sum(0, dtype=np.int)
r, c, arrays, _ = _prune(r, c, A, w)
A, w = arrays
# Sizing
m, n = len(r), len(c)
if (m == 0) or (n == 0):
return 0.0
# Sort the row margins (descending)
rndx = np.argsort(-r)
rsort = r[rndx]
# Balance the weights
a_scale, b_scale = canonical_scalings(w, r, c)
wopt = apply_scale(w, a_scale, b_scale)
if np.isnan(wopt).any():
wopt = w
# Reorder the columns
if sort_by_wopt_var:
cndx = np.lexsort((-wopt.var(0), c))
else:
cndx = np.argsort(c)
csort = c[cndx]
wopt = wopt[:, cndx]
# Compute G
G = _compute_G(r, m, n, wopt)
return _compute_cnll(A, r, rsort, rndx, csort, cndx, m, n, G)
def approximate_from_margins_weights(r, c, w, T = None,
sort_by_wopt_var = True):
"""Return approximate samples from row/column-conditional binary matrices.
Return a binary matrix (or a list of binary matrices) sampled
approximately according to the specified Bernoulli weights,
conditioned on having the specified margins.
Inputs:
r: row margins, length m
c: column margins, length n
w: weight matrix, (m x n) matrix with values in (0, +infty)
T: number of matrices to sample
sort_by_wopt_var: when enabled, column ordering depends on w
Output:
B_sample_sparse: (T default) sparse representation of (m x n) binary matrix
(T >= 1) list of (sparse binary matrices, logQ, logP)
More explicitly, consider independent Bernoulli random variables
B(i,j) arranged as an m x n matrix B given the m-vector of row sums
r and the n-vector of column sums c of the sample, i.e., given that
sum(B_sample, 1) = r and sum(B_sample, 0) = c.
An error is generated if no binary matrix agrees with r and c.
B(i,j) is Bernoulli(p(i,j)) where p(i,j) = w(i,j)/(1+w(i,j)), i.e.,
w(i,j) = p(i,j)/(1-p(i,j)). [The case p(i,j) = 1 must be handled by
the user in a preprocessing step, by converting to p(i,j) = 0 and
decrementing the row and column sums appropriately.]
The sparse representation used for output is a matrix giving the
locations of the ones in the sample. If d = sum(r) = sum(c), then
B_sample_sparse has dimensions (d x 2). If something goes wrong (due
to undetected improper input), some of the rows of B_sample_sparse
may [-1,-1], indicating no entry of B_sample.
B_sample can be recovered from B_sample_sparse via:
B_sample = np.zeros((m,n), dtype=np.bool)
for i, j in B_sample_sparse:
if i == -1: break
B_sample[i,j] = 1
"""
r_prune, c_prune, arrays_prune, unprune = _prune(r, c, w)
w_prune = arrays_prune[0]
_check_margins(r_prune, c_prune)
### Preprocessing
# Sizing (making copies of m and n, as they are mutated during sampling)
r_init = r_prune.copy()
m, n = len(r_prune), len(c_prune)
if (m == 0) or (n == 0):
if T:
return [unprune([np.empty((0,2)), 0, 0]) for t in xrange(T)]
else:
return np.empty((0,0))
m_init, n_init = m, n
assert((m,n) == w_prune.shape)
# Sort the row margins (descending)
rndx_init = np.argsort(-r_prune)
rsort = r_prune[rndx_init]
# Balance the weights
a_scale, b_scale = canonical_scalings(w_prune, r_prune, c_prune)
wopt = apply_scale(w_prune, a_scale, b_scale)
# Reorder the columns
if sort_by_wopt_var:
cndx = np.lexsort((-wopt.var(0), c_prune))
else:
cndx = np.argsort(c_prune)
csort = c_prune[cndx]
wopt = wopt[:,cndx]
# Precompute log weights
logw = np.log(w_prune)
# Compute G
G = _compute_G(r_prune, m, n, wopt)
# Generate the inverse index for the row orders to facilitate fast
# sorting during the updating
irndx_init = np.argsort(rndx_init)
# Compute the conjugate of c
cconj_init = conjugate(csort, m)
# Get the running total of number of ones to assign
count_init = np.sum(rsort)
def do_sample():
sample_prune = _compute_sample(logw,
count_init, m_init, n_init,
r_init, rndx_init, irndx_init,
csort, cndx, cconj_init,
G)
return unprune(sample_prune)
if T:
return [do_sample() for t in xrange(T)]
else:
return do_sample()[0]
return logkappa, logcvsq
if __name__ == '__main__':
# Test of binary matrix generation code
m = np.random.random(size=(12,10)) < 0.3
r, c = np.sum(m, axis = 1), np.sum(m, axis = 0)
print r, c
A = arbitrary_from_margins(r, c)
print np.sum(A, axis = 1), np.sum(A, axis = 0)
# Test of "rc" balancing
m = np.random.normal(10, 1, size = (6,5))
r, c = np.ones(6), np.ones(5)
c[0] = 2
a, b = canonical_scalings(m, r, c)
m_canonical = apply_scale(m, a, b)
print m_canonical.sum(1)
print m_canonical.sum(0)
# Test of conjugate
print conjugate([1,1,1,1,2,8], 10)
# Test of approximate margins-conditional sampling
N = 50
a_out = np.random.normal(0, 1, N)
a_in = np.random.normal(0, 1, N)
x = np.random.normal(0, 1, (N,N))
theta = 0.8
logit_P = np.zeros((N,N))
for i, a in enumerate(a_out):
logit_P[i,:] += a
def approximate_from_margins_weights(r, c, w, T=None, sort_by_wopt_var=True):
"""Return approximate samples from row/column-conditional binary matrices.
Return a binary matrix (or a list of binary matrices) sampled
approximately according to the specified Bernoulli weights,
conditioned on having the specified margins.
Inputs:
r: row margins, length m
c: column margins, length n
w: weight matrix, (m x n) matrix with values in (0, +infty)
T: number of matrices to sample
sort_by_wopt_var: when enabled, column ordering depends on w
Output:
B_sample_sparse: (T default) sparse representation of (m x n) binary matrix
(T >= 1) list of (sparse binary matrices, logQ, logP)
More explicitly, consider independent Bernoulli random variables
B(i,j) arranged as an m x n matrix B given the m-vector of row sums
r and the n-vector of column sums c of the sample, i.e., given that
sum(B_sample, 1) = r and sum(B_sample, 0) = c.
An error is generated if no binary matrix agrees with r and c.
B(i,j) is Bernoulli(p(i,j)) where p(i,j) = w(i,j)/(1+w(i,j)), i.e.,
w(i,j) = p(i,j)/(1-p(i,j)). [The case p(i,j) = 1 must be handled by
the user in a preprocessing step, by converting to p(i,j) = 0 and
decrementing the row and column sums appropriately.]
The sparse representation used for output is a matrix giving the
locations of the ones in the sample. If d = sum(r) = sum(c), then
B_sample_sparse has dimensions (d x 2). If something goes wrong (due
to undetected improper input), some of the rows of B_sample_sparse
may [-1,-1], indicating no entry of B_sample.
B_sample can be recovered from B_sample_sparse via:
B_sample = np.zeros((m,n), dtype=np.bool)
for i, j in B_sample_sparse:
if i == -1: break
B_sample[i,j] = 1
"""
r_prune, c_prune, arrays_prune, unprune = _prune(r, c, w)
w_prune = arrays_prune[0]
_check_margins(r_prune, c_prune)
### Preprocessing
# Sizing (making copies of m and n, as they are mutated during sampling)
r_init = r_prune.copy()
m, n = len(r_prune), len(c_prune)
if (m == 0) or (n == 0):
if T:
return [unprune([np.empty((0, 2)), 0, 0]) for t in xrange(T)]
else:
return np.empty((0, 0))
m_init, n_init = m, n
assert ((m, n) == w_prune.shape)
# Sort the row margins (descending)
rndx_init = np.argsort(-r_prune)
rsort = r_prune[rndx_init]
# Balance the weights
a_scale, b_scale = canonical_scalings(w_prune, r_prune, c_prune)
wopt = apply_scale(w_prune, a_scale, b_scale)
# Reorder the columns
if sort_by_wopt_var:
cndx = np.lexsort((-wopt.var(0), c_prune))
else:
cndx = np.argsort(c_prune)
csort = c_prune[cndx]
wopt = wopt[:, cndx]
# Precompute log weights
logw = np.log(w_prune)
# Compute G
G = _compute_G(r_prune, m, n, wopt)
# Generate the inverse index for the row orders to facilitate fast
# sorting during the updating
irndx_init = np.argsort(rndx_init)
# Compute the conjugate of c
cconj_init = conjugate(csort, m)
# Get the running total of number of ones to assign
count_init = np.sum(rsort)
def do_sample():
sample_prune = _compute_sample(logw, count_init, m_init, n_init,
r_init, rndx_init, irndx_init, csort,
cndx, cconj_init, G)
return unprune(sample_prune)
if T:
return [do_sample() for t in xrange(T)]
else:
return do_sample()[0]
if __name__ == '__main__':
# Test of binary matrix generation code
m = np.random.random(size=(12, 10)) < 0.3
r, c = np.sum(m, axis=1), np.sum(m, axis=0)
print r, c
A = arbitrary_from_margins(r, c)
print np.sum(A, axis=1), np.sum(A, axis=0)
# Test of "rc" balancing
m = np.random.normal(10, 1, size=(6, 5))
r, c = np.ones((6, 1)), np.ones((1, 5))
c[0] = 2
a, b = canonical_scalings(m, r, c)
m_canonical = apply_scale(m, a, b)
print m_canonical.sum(1)
print m_canonical.sum(0)
# Test of conjugate
print conjugate([1, 1, 1, 1, 2, 8], 10)
# Test of approximate margins-conditional sampling
N = 50
a_out = np.random.normal(0, 1, N)
a_in = np.random.normal(0, 1, N)
x = np.random.normal(0, 1, (N, N))
theta = 0.8
logit_P = np.zeros((N, N))
for i, a in enumerate(a_out):
logit_P[i, :] += a
