def __init__(self, nstates=None, nsymbols=None, transition=None, emission=None, pi=None):
assert not ((nstates is None and transition is None and pi is None) or (nsymbols is None and emission is None))
if not transition is None:
assert transition.ndim == 2
if not nstates is None:
assert nstates == transition.shape[0]
else:
nstates = transition.shape[0]
if not emission is None:
assert transition.shape[0] == emission.shape[0]
if not pi is None:
assert transition.shape[0] == pi.size
if not emission is None:
assert emission.ndim == 2
if not nsymbols is None:
assert nsymbols == emission.shape[1]
else:
nsymbols = emission.shape[1]
if not nstates is None:
assert nstates == emission.shape[0]
else:
nstates == emission.shape[0]
if not pi is None:
assert emission.shape[0] == pi.size
if not pi is None:
assert pi.ndim == 1
if not nstates is None:
assert nstates == pi.size
else:
nstates = pi.size
if transition is None:
transition = almost_uniform_matrix(nstates)
if emission is None:
emission = almost_uniform_matrix(nstates, nsymbols)
if pi is None:
pi = almost_uniform_vector(nstates)
self.set_transition(transition)
self.set_emission(emission)
self.set_pi(pi)
def entropic_reestimate(omega, theta=None, Z=1, maxiter=100, tol=1e-7, verbose=False):
"""
Re-estimates a statistic parameter vector entropically [1]_.
Parameters
----------
omega : array_like
Evidence vector
theta : array_like, optional
Parameter vector to be re-estimated under given evidence and learning rate (default None)
Z : {-1, 0, +1}, optional
-1: Algorithm reduces to traditional MLE (e.g the Baum-Welch)
0: ?
+1: Algorithm will seek maximum structure
maxiter : int, optional
Maximum number of iterations of Fixed-point loop (default 100)
verbose : bool, optional
Display verbose output (default off)
Returns
-------
theta_hat : array_like
Learned parameter vector
Z : float
Final Learning rate
_lambda : float
Limiting value of Lagrange multiplier
Examples
--------
>>> from entropy_map import entropic_reestimate
>>> omega = [1, 2]
>>> theta = [0.50023755, 0.49976245]
>>> theta_hat, final_Z, _lambda = entropic_reestimate(omega, theta, Z=1, tol=1e-6)
>>> theta_hat
array([ 0.33116253, 0.66883747])
>>> final_Z
0.041828014112488016
>>> _lambda
-3.0152672618320637
References
----------
.. [1] Matthiew Brand, "Pattern learning via entropy maximization"
"""
def _debug(msg=''):
if verbose:
print msg
# XXX TODO: handle Z = 0 case
assert Z != 0
# if no initial theta specified, start with uniform candidate
if theta is None:
theta = almost_uniform_vector(len(omega))
# all arrays must be numpy-like
omega = array(omega, dtype='float64')
theta = array(theta, dtype='float64')
# XXX TODO: trim-off any evidence which 'relatively close to 0' (since such evidence can't justify anything!)
informative_indices = nonzero(minimum(omega, theta) > _EPSILON)
_omega = omega[informative_indices]
_theta = theta[informative_indices]
# prepare initial _lambda which will ensure that Lambert's W is real-valued
if Z > 0:
critical_lambda = min(-Z*(2 + log(_omega/Z)))
_lambda = critical_lambda - 1 # or anything less than the critical value above
elif Z < 0:
# make an educated guess
_lambda = -mean(Z*(log(_theta) + 1) + _omega/_theta)
assert all(-_omega*exp(1+_lambda/Z)/Z > -1/e), -_omega*exp(1+_lambda/Z)/Z
# Fixed-point loop
_theta_hat = _theta
iteration = 0
converged = False
_debug("entropy_map: starting Fixed-point loop ..\n")
_debug("Initial model: %s"%_theta)
_debug("Initial lambda: %s"%_lambda)
_debug("Initila learning rate (Z): %s"%Z)
while not converged:
# exhausted ?
if maxiter <= iteration:
break
# if necessary, re-scale learning rate (Z) so that exp(1 + _lambda/Z) is not 'too small'
if _lambda < 0:
if Z > 0:
new_Z = -_lambda/_BEAR
elif Z < 0:
new_Z = _lambda/_BEAR
if new_Z != Z:
Z = new_Z
_debug("N.B:- We'll re-scale learning rate (Z) to %s to prevent Lambert's W function from vanishing."%(Z))
# prepare argument (vector) for Lambert's W function
z = -_omega*exp(1 + _lambda/Z)/Z
assert all(isreal(z))
if any(z < -1/e):
_debug("Lambert's W: argument z = %s out of range (-1/e, +inf)"%z)
break
# compute Lambert's W function at z
if Z <= 0:
g = W(z, k=0)
else:
g = W(z, k=-1)
assert all(isreal(g))
g = real(g)
# check against division by zero (btw we re-scaled Z to prevent this)
# assert all(g != 0)
assert all(abs(g) > _EPSILON)
# re-estimate _theta
_theta_hat = (-_omega/Z)/g
assert all(_theta_hat >= 0)
# normalize the approximated _theta_hat parameter
_theta_hat = normalize_probabilities(_theta_hat)
# re-estimate _lambda
_lambda_hat = -(Z*(log(_theta_hat[0]) + 1) + _omega[0]/_theta_hat[0]) # [0] or any other index [i]
# check whether _lambda values have convergede
converged, _, relative_error = check_converged(_lambda, _lambda_hat, tol=tol)
# verbose for debugging, etc.
_debug("Iteration: %d"%iteration)
_debug('Current parameter estimate:\n%s'%_theta)
_debug('lambda: %s'%_lambda)
_debug("Relative error in lambda over last iteration: %s"%relative_error)
_debug("Learning rate (Z): %s"%Z)
# update _lambda and _theta
_lambda = _lambda_hat
_theta = _theta_hat
# goto next iteration
iteration += 1
_debug('\n')
_debug("Done.")
_debug('Final parameter estimate:\n%s'%_theta)
_debug('lambda: %s'%_lambda)
_debug("Relative error in lambda over last iteration: %s"%relative_error)
_debug("Learning rate (Z): %s"%Z)
# converged ?
if converged:
_debug("entropic_reestimate: loop converged after %d iterations (tolerance was set to %s)"%(iteration,tol))
else:
_debug("entropic_reestimate: loop did not converge after %d iterations (tolerance was set to %s)"\
%(maxiter,tol))
# render results
theta_hat = 0*theta
theta_hat[informative_indices] = _theta_hat
return theta_hat, Z, _lambda
