def __str__(self):
strs = ['#p\t#q\t#f\t#dp\t#y']
for i, Dy in enumerate(self.support_):
strs.append('{0}\t{1}\t{2}\t{3}\t{4}'.format(
self.p(i), self.q(i), npvec2str(self.f(i)),
npvec2str(self.dtheta(i)), Dy.projection))
strs.append('#\t#\t#<f>\t#\t#')
strs.append(' \t \t%s\t \t ' % npvec2str(self.uf()))
return '\n'.join(strs)
def __str__(self):
strs = ['#p\t#q\t#f\t#dp\t#y']
for i, Dy in enumerate(self.support_):
strs.append('{0}\t{1}\t{2}\t{3}\t{4}'.format(self.p(i),
self.q(i),
npvec2str(self.f(i)),
npvec2str(self.dtheta(i)),
Dy.projection))
strs.append('#\t#\t#<f>\t#\t#')
strs.append(' \t \t%s\t \t ' % npvec2str(self.uf()))
return '\n'.join(strs)
def f(w):
self.iteration += 1
# the joint update is no longer supported: self.wmap.update(w)
self.history.append(w)
self.wmap.proxy.update(w)
logging.info('[%d/%d] lambda=%s', self.parent_iteration,
self.iteration, npvec2str(w))
#logging.info('[%d/%d] sampling ...', self.parent_iteration, self.iteration)
self.sample()
#logging.info('[%d/%d] KL ...', self.parent_iteration, self.iteration)
divergences, derivatives = self.KL()
kl = divergences.mean(0)
dkl = derivatives.mean(0)
# r_weight
r_weight = 0
# regularised
regulariser = LA.norm(w, 2)
r_obj = kl + r_weight * regulariser
r_jac = dkl + 2 * r_weight * w
logging.info('[%d/%d] KL=%s regularised=%f', self.parent_iteration,
self.iteration, kl, r_obj)
return r_obj, r_jac
def callback(theta):
logging.info('[%d] new theta: %s', self.iteration,
npvec2str(theta))
# register the theta
self.wmap.target.update(theta)
self.selected_thetas.append(theta)
self.last_callback = self.iteration
def f(theta):
run.next_proxysgd()
self.wmap.proxy.update(theta)
logging.info('[%s] lambda=%s', run, npvec2str(theta))
OBJ = []
JAC = []
H = []
dH = []
for i, seg in enumerate(devset):
derivations = S[i]
lmap = L[i]
local_obj, local_jac, local_H, local_dH = get_obj_and_jac(
derivations,
q_wmap=self.wmap.proxy,
p_wmap=self.wmap.target,
lmap=lmap,
empirical_q=False)
OBJ.append(local_obj)
JAC.append(local_jac)
H.append(local_H)
dH.append(local_dH)
obj, jac = np.mean(OBJ, 0), np.mean(JAC, 0)
if polarity != 1:
obj *= polarity
jac *= polarity
if Tq == 0.0 and l2_weight == 0.0:
logging.info('[%s] O=%f', run, obj)
return obj, jac
else:
r_obj = obj
r_jac = jac.copy()
if Tq != 0.0:
r_obj -= Tq * np.mean(H, 0)
r_jac -= Tq * np.mean(dH, 0)
logging.info('[%s] O=%f H-regularised=%f', run, obj, r_obj)
if l2_weight != 0.0: # regularised
regulariser = LA.norm(theta, 2)**2
r_obj += l2_weight * regulariser
r_jac += 2 * l2_weight * theta
logging.info('[%s] O=%f L2-regularised=%f', run, obj,
r_obj)
return r_obj, r_jac
def f(theta):
run.next_targetsgd()
self.wmap.target.update(theta)
logging.info('[%s] theta=%s', run, npvec2str(theta))
R = []
dR = []
H = []
dH = []
for i, seg in enumerate(devset):
derivations = S[i]
lmap = L[i]
local_R, local_dR, local_H, local_dH = optalg.minrisk(
derivations,
q_wmap=self.wmap.proxy,
p_wmap=self.wmap.target,
lmap=lmap,
empirical_q=True)
R.append(local_R)
dR.append(local_dR)
H.append(local_H)
dH.append(local_dH)
obj, jac = np.mean(R, 0), np.mean(dR, 0)
if Tp == 0.0 and l2_weight == 0.0:
logging.info('[%s] Risk=%f', run, obj)
return obj, jac
else:
r_obj = obj
r_jac = jac.copy()
if Tp != 0.0: # H-regularised
r_obj -= Tp * np.mean(H, 0)
r_jac -= Tp * np.mean(dH, 0)
logging.info('[%s] Risk=%f H-regularised=%f', run, obj,
r_obj)
if l2_weight != 0.0: # L2-regularised
regulariser = LA.norm(theta, 2)**2
r_obj += l2_weight * regulariser
r_jac += 2 * l2_weight * theta
logging.info('[%s] Risk=%f L2-regularised=%f', run, obj,
r_obj)
return r_obj, r_jac
def save_estimates(sid, empdist, losses, posterior, risk, dR, stems):
"""
"""
with open('{0}/{1}'.format(stems[0], sid), 'w') as fo:
print >> fo, empdist
with open('{0}/{1}'.format(stems[1], sid), 'w') as fo:
print >> fo, '#loss'
print >> fo, '\n'.join('%.4f' % l for l in losses)
with open('{0}/{1}'.format(stems[2], sid), 'w') as fo:
print >> fo, '#risk\t#dR'
print >> fo, '{0}\t{1}'.format(risk, npvec2str(dR))
def f(theta):
self.iteration += 1
# the joint update is no longer supported: self.wmap.update(w)
self.wmap.target.update(theta)
logging.info('[%d] theta=%s', self.iteration, npvec2str(theta))
# KL here?
# optimises KL if necessary
if self.kl_current > self.kl_min * self.kl_coeff or self.iteration > self.kl_lastit + 10:
lambdas, self.kl_min = self.optimise_proxy(target=theta)
self.kl_lastit = self.iteration
# register the lambda
self.wmap.proxy.update(lambdas)
self.selected_lambdas.append(lambdas)
#logging.info('[%d] sampling ...', self.iteration)
self.sample()
# logging.info('[%d] decision rules ...', self.iteration)
# self.decide()
# compute loss
#logging.info('[%d] assessing risk ...', self.iteration)
risks, jacs, kls = self.risk()
risk = risks.mean(0)
jac = jacs.mean(0)
kl = kls.mean(0)
if kl < self.kl_min:
self.kl_min = kl
self.kl_current = kl
# r_weight
r_weight = 0
# regularised
regulariser = LA.norm(theta, 2)
r_obj = risk + r_weight * regulariser
r_jac = jac + 2 * r_weight * theta
logging.info('[%d] RISK=%f regularised=%f KL=%f', self.iteration,
risk, r_obj, kl)
#logging.info('[%d] jac=%s', self.iteration, jac)
return r_obj, r_jac
def callback(w):
self.selected.append(w)
logging.info('[%d/%d] new lambda: %s', self.parent_iteration,
self.iteration, npvec2str(w))
def optimise(self):
self.iteration = -1
self.history = []
self.selected = []
def f(w):
self.iteration += 1
# the joint update is no longer supported: self.wmap.update(w)
self.history.append(w)
self.wmap.proxy.update(w)
logging.info('[%d/%d] lambda=%s', self.parent_iteration,
self.iteration, npvec2str(w))
#logging.info('[%d/%d] sampling ...', self.parent_iteration, self.iteration)
self.sample()
#logging.info('[%d/%d] KL ...', self.parent_iteration, self.iteration)
divergences, derivatives = self.KL()
kl = divergences.mean(0)
dkl = derivatives.mean(0)
# r_weight
r_weight = 0
# regularised
regulariser = LA.norm(w, 2)
r_obj = kl + r_weight * regulariser
r_jac = dkl + 2 * r_weight * w
logging.info('[%d/%d] KL=%s regularised=%f', self.parent_iteration,
self.iteration, kl, r_obj)
return r_obj, r_jac
def callback(w):
self.selected.append(w)
logging.info('[%d/%d] new lambda: %s', self.parent_iteration,
self.iteration, npvec2str(w))
self.result_ = minimize(
f,
self.wmap.proxy.asarray,
method='L-BFGS-B',
jac=True,
callback=callback,
options={
'maxfun': 10,
'ftol': 1e-6,
'gtol': 1e-3,
'disp': False
}) # TODO find a way to stop the search earlier 'maxfun'
logging.info('[%d/%d] final KL=%s lambda=%s',
self.parent_iteration, self.iteration, self.result_.fun,
npvec2str(self.result_.x))
return self.result_.x, self.result_.fun
def callback(theta):
logging.info('[%s] new lambda: %s', run, npvec2str(theta))