def fprime(y):
logLik, gw = eval(y)
#print '==================='
#print '=>', joint, np.min(model.logpxz(_w, x, _z)), np.min(model.mclik(_w, x))
#print '==================='
gw = ndict.flatten(gw)
log['loglik'].append(logLik)
return - gw
def lbfgs(posterior, w, hook=None, hook_wavelength=2, m=100, maxiter=15000):
print 'L-BFGS', m, maxiter
cache = [1234,0,0]
def eval(y):
# TODO: solution for random seed in ipcluster
# maybe just execute a remote np.random.seed(0) there?
np.random.seed(0)
_w = ndict.unflatten(y, w)
logpw, gw = cache[1], cache[2]
if np.linalg.norm(y) != cache[0]:
logpw, gw = posterior.grad(_w)
#print logpw, np.linalg.norm(y)
cache[0], cache[1], cache[2] = [np.linalg.norm(y), logpw, gw]
return logpw, gw
def f(y):
logLik, gw = eval(y)
return - logLik.mean()
def fprime(y):
logLik, gw = eval(y)
#print '==================='
#print '=>', joint, np.min(model.logpxz(_w, x, _z)), np.min(model.mclik(_w, x))
#print '==================='
gw = ndict.flatten(gw)
log['loglik'].append(logLik)
return - gw
t = [0, 0, time.time()]
def callback(wz):
if hook is None: return
t[1] += 1
if time.time() - t[2] > hook_wavelength:
_w = ndict.unflatten(wz, w)
hook(t[1], _w, cache[1]) # num_its, w, logpw
t[2] = time.time()
x0 = ndict.flatten(w)
xn, f, d = scipy.optimize.fmin_l_bfgs_b(func=f, x0=x0, fprime=fprime, m=m, iprint=0, callback=callback, maxiter=maxiter)
#scipy.optimize.fmin_cg(f=f, x0=x0, fprime=fprime, full_output=True, callback=hook)
#scipy.optimize.fmin_ncg(f=f, x0=x0, fprime=fprime, full_output=True, callback=hook)
w = ndict.unflatten(xn, w)
#print 'd: ', d
if d['warnflag'] is 2:
print 'warnflag:', d['warnflag']
print d['task']
info = d
return w, info