def slda_update_log_phi(text, log_phi, log_gamma, log_beta, y_d, eta, sigma_squared):
"""
Same as update_phi_lda_E_step but in log probability space.
"""
(N, K) = log_phi.shape
log_phi_sum = logsumexp(log_phi, axis=0)
Ns = (N * sigma_squared)
ElogTheta = graphlib.dirichlet_expectation(np.exp(log_gamma))
front = (-1.0 / (2 * N * Ns))
pC = (1.0 * y_d / Ns * eta)
eta_dot_eta = front * (eta * eta)
log_const = np.log(ElogTheta + pC + eta_dot_eta)
log_right_eta_times_const = np.log(front * 2 * eta)
ensure(isinstance(text, np.ndarray))
# if text is in array form, do an approximate fast matrix update
log_phi_minus_n = -1 + (logsumexp([log_phi, (-1 + log_phi_sum)]))
log_phi[:,:] = logsumexp([log_beta[:,text].T,
logdotexp(np.matrix(logdotexp(log_phi_minus_n, np.log(eta))).T,
np.matrix(log_right_eta_times_const)),
log_const,], axis=0)
graphlib.log_row_normalize(log_phi)
return log_phi
def initialize_beta(num_topics, num_words):
"""Initializes beta randomly using a random dirichlet.
Accepts integers number of topics, and number of words in vocab.
Returns a TxW matrix which have the probabilities of word
distributions. Each row sums to 1.
"""
log_beta = initialize_log_beta(num_topics, num_words)
return np.exp(log_beta)
def partial_slda_update_phi(text, phi, gamma, beta, y_d, eta, sigma_squared):
"""Same as slda update phi, but eta may be smaller than total number of topics.
So only some of the topics contribute to y.
"""
(N, K) = phi.shape
Ks = len(eta)
phi_sum = np.sum(phi[:,:Ks], axis=0)
Ns = (N * sigma_squared)
ElogTheta = graphlib.dirichlet_expectation(gamma)
front = (-1.0 / (2 * N * Ns))
eta_dot_eta = front * (eta * eta)
pC = ((1.0 * y_d / Ns) * eta) + eta_dot_eta
right_eta_times_const = (front * 2 * eta)
if isinstance(text, np.ndarray):
# if text is in array form, do an approximate fast matrix update
phi_minus_n = -(phi[:,:Ks] - phi_sum)
phi[:,:] = ElogTheta + np.log(beta[:,text].T)
phi[:,:Ks] += pC
phi[:,:Ks] += np.dot(np.matrix(np.dot(phi_minus_n, eta)).T, np.matrix(right_eta_times_const))
graphlib.log_row_normalize(phi)
phi[:,:] = np.exp(phi[:,:])
else:
# otherwise, iterate through each word
for n,word,count in iterwords(text):
phi_sum -= phi[n,:Ks]
pB = np.log(beta[:,word])
pD = (np.dot(eta, phi_sum) * right_eta_times_const)
# must exponentiate and normalize immediately!
phi[n,:] = ElogTheta + pB
phi[n,:] += pC + pD
phi[n,:] -= graphlib.logsumexp(phi[n,:]) # normalize in logspace
phi[n,:] = np.exp(phi[n,:])
# add this back into the sum
# unlike in LDA, this cannot be computed in parallel
phi_sum += phi[n,:Ks]
return phi
def test_exp(self):
import math
from numpypy import array, exp
a = array([-5.0, -0.0, 0.0, 12345678.0, float("inf"), -float("inf"), -12343424.0])
b = exp(a)
for i in range(4):
try:
res = math.exp(a[i])
except OverflowError:
res = float("inf")
assert b[i] == res
def _unoptimized_slda_update_phi(text, phi, gamma, beta, y_d, eta, sigma_squared):
"""
Update phi in LDA.
phi is N x K matrix.
gamma is a K-size vector
update phid:
φd,n ∝ exp{ E[log θ|γ] +
E[log p(wn|β1:K)] +
(y / Nσ2) η —
[2(ηTφd,-n)η + (η∘η)] / (2N2σ2) }
Note that E[log p(wn|β1:K)] = log βTwn
"""
(N, K) = phi.shape
#assert len(eta) == K
#assert len(gamma) == K
#assert beta.shape[0] == K
phi_sum = np.sum(phi, axis=0)
Ns = (N * sigma_squared)
ElogTheta = graphlib.dirichlet_expectation(gamma)
ensure(len(ElogTheta) == K)
pC = (1.0 * y_d / Ns * eta)
eta_dot_eta = (eta * eta)
front = (-1.0 / (2 * N * Ns))
for n,word,count in iterwords(text):
phi_sum -= phi[n]
ensure(len(phi_sum) == K)
pB = np.log(beta[:,word])
pD = (front * (((2 * np.dot(eta, phi_sum) * eta) + eta_dot_eta))
)
ensure(len(pB) == K)
ensure(len(pC) == K)
ensure(len(pD) == K)
# must exponentiate and sum immediately!
#phi[n,:] = np.exp(ElogTheta + pB + pC + pD)
#phi[n,:] /= np.sum(phi[n,:])
# log normalize before exp for numerical stability
phi[n,:] = ElogTheta + pB + pC + pD
phi[n,:] -= graphlib.logsumexp(phi[n,:])
phi[n,:] = np.exp(phi[n,:])
# add this back into the sum
# unlike in LDA, this cannot be computed in parallel
phi_sum += phi[n]
return phi
def lda_update_phi(text, phi, gamma, beta, normalize=True, logspace=False):
"""
Update phi in LDA.
phi is N x K matrix.
gamma is a K-size vector
update phid:
φd,n ∝ exp{ E[log θ|γ] +
E[log p(wn|β1:K)] }
Note that E[log p(wn|β1:K)] = log βTwn
"""
(N, K) = phi.shape
ElogTheta = graphlib.dirichlet_expectation(gamma)
# todo: call a log version of this in slda and others!
ensure(isinstance(text, np.ndarray))
phi[:,:] = ElogTheta + np.log(beta[:,text].T)
if normalize:
graphlib.log_row_normalize(phi)
if not logspace:
phi[:,:] = np.exp(phi[:,:])
return phi
def logistic_sigmoid(v):
"""Returns 1 / (1 + e^(-v))"""
return 1.0 / (1 + np.exp(-v))
do_var = ptr(do_var)
getattr(ff, name_fit)(P.size, ptr(P), x.size, ptr(x), ptr(y),
ptr(ydata), ptr(a), do_var)
return P
return fun, fun_diff, fun_rms, fun_fit
e2, e2_diff, e2_rms, e2_fit = _fun_factory('_e2')
IV3, IV3_diff, IV3_rms, IV3_fit = _fun_factory('_IV3')
IVdbl, IVdbl_diff, IVdbl_rms, IVdbl_fit = _fun_factory('_IVdbl')
IV4, IV4_diff, IV4_rms, IV4_fit = _fun_a_factory('_IV4')
IV5, IV5_diff, IV5_rms, IV5_fit = _fun_a_factory('_IV5')
IV6, IV6_diff, IV6_rms, IV6_fit = _fun_a_factory('_IV6')
IVdbl2, IVdbl2_diff, IVdbl2_rms, IVdbl2_fit = _fun_a_factory('_IVdbl2')
if __name__ == "__main__":
try:
import numpypy as np
except ImportError:
import numpy as np
P = np.array([1., 1.])
x = np.arange(4.)
y = np.empty(x.shape, x.dtype)
print e2(P, x, y)
print P[0] * np.exp(-P[1] * x)
