def __init__(self, func, eps1=1e-4, eps2=1e-8, nthreads=1 ,kw={}):
self.func=func
if nthreads > 1:
self.pool = workerpool.pool(func, nthreads=nthreads, kw=kw)
self.eps1 = eps1
self.eps2 = eps2
def _minimize_neldermead_para(func, x0, callback=None,
xtol=1e-4, ftol=1e-4, maxiter=None, maxfev=None,
disp=False, return_all=False,pool=None, nthreads=1,
nverts=None, nactive=None):
maxfun = maxfev
retall = return_all
allvecs = []
x0 = np.asfarray(x0).flatten()
N = len(x0)
fcalls = [0]
allcalls = []
if nverts is None:
nverts = (N+1)*2
assert(nverts>=(N+1))
if nthreads>=nverts:
nthreads = nverts
pool = workerpool.pool(func, nthreads)
if nactive is None:
nactive = nthreads # number of active points
if nactive >= nverts:
nactive = nverts-1
def applicator(i, x):
#print x
pool.apply_async(i, x)
fcalls[0] +=1
allcalls.append(x.tolist())
rank = len(x0.shape)
if not -1 < rank < 2:
raise ValueError("Initial guess must be a scalar or rank-1 sequence.")
if maxiter is None:
maxiter = N * 200
if maxfun is None:
maxfun = N * 200
rho = 1
chi = 2
psi = 0.5
sigma = 0.5
nonzdelt = 0.05
zdelt = 0.00025
excess = nverts - N -1
one2np1 = list(range(1, nverts))
if rank == 0:
sim = np.zeros((nverts,), dtype=x0.dtype)
else:
sim = np.zeros((nverts, N), dtype=x0.dtype)
fsim = np.zeros((nverts,), float)
applicator(0, x0)
sim[0] = x0
if False:
for k in range(0, N):
y = np.array(x0, copy=True)
if y[k] != 0:
y[k] = (1 + nonzdelt) * y[k]
else:
y[k] = zdelt
sim[k + 1] = y
applicator(k + 1, y)
nover = int(np.ceil(excess *1. / N))
for k in range(N, nverts-1):
pos1 = (k-N) / nover
pos2 = (k-N) % nover
#print k,pos1,pos2
y = (sim[1+pos1]-x0) * (pos2+1)*1./(nover+1)+x0
#y = np.array(x0, copy=True)
#fac= np.random.uniform(0,zdelt,size=N)
#y = (x0*(1+fac))*(x0!=0).astype(int)+(fac)*(x0==0).astype(int)
sim[k + 1] = y
applicator(k + 1, y)
state=np.random.get_state()
np.random.seed(1)
for k in range(0, nverts-1):
y = np.array(x0, copy=True)
rand=np.random.normal(0,1, len(x0))
rand=rand/(rand**2).sum()**.5+1./len(x0)**.5
y = (x0)*(1+rand*nonzdelt)*(x0!=0).astype(int)+(x0==0).astype(int)*(rand)*nonzdelt
sim[k + 1] = y
applicator(k + 1, y)
np.random.set_state(state)
for k in range(0, nverts):
fsim[k] = pool.get(k)
ind = np.argsort(fsim)
fsim = np.take(fsim, ind, 0)
sim = np.take(sim, ind, 0)
iterations = 1
simR = sim.copy() # reflection
simE = sim.copy() # expansion
simC = sim.copy() # contraction
simFinal = sim.copy() # final
fsimR = fsim.copy()
fsimE = fsim.copy()
fsimC = fsim.copy()
fsimFinal = fsim.copy()
while (fcalls[0] < maxfun and iterations < maxiter):
if (np.max(np.ravel(np.abs(sim[1:] - sim[0]))) <= xtol and
np.max(np.abs(fsim[0] - fsim[1:])) <= ftol):
break
xbar = np.add.reduce(sim[:-nactive], 0) / (nverts - nactive)
state = np.zeros(nactive, dtype=int)
actives = set(range(nactive))
shrinks = []
for i in range(nactive):
# reflect bad points
#xbar = (np.add.reduce(sim[:], 0)-sim[-i-1,:]) / N
simR[-i - 1, :] = (1 + rho) * xbar - rho * sim[-i-1,:]
applicator(i, simR[-i-1, :])
state[i] = 1
#print 'A'
while len(actives)>0:
#print 'ST', state
i, curfsim = pool.get_any()
#xbar = (np.add.reduce(sim[:], 0)-sim[-i-1,:]) / N
if state[i] == 1: # point just got evaled after first reflection
fsimR[-i-1] = curfsim
if fsimR[-i-1] < fsim[0]: # expand
simE[-i-1, :] = (1 + rho * chi) * xbar - rho * chi* sim[-i -1, :]
applicator(i, simE[-i - 1, :])
state[i] = 2
else:
if fsimR[-i-1] < fsim[-i-2]: # better than the next worst
simFinal[-i-1, :] = simR[-i-1,:]
fsimFinal[-i-1] = fsimR[-i-1]
state[i]= -1 # we stopped after one reflection
actives.remove(i)
else: # contract
if fsimR[-i-1] < fsim[-i-1]:
simC[-i-1, :] = (1 + rho * psi) * xbar - rho * psi* sim[-i-1, :]
applicator(i, simC[-i-1,:])
state[i] = 3
else:
simC[-i- 1, :] = (1 - psi) * xbar + psi * sim[-i-1, :]
applicator(i, simC[-i-1,:])
state[i] = 4
elif state[i] == 2: # the expansion step has been done
fsimE[-i-1] = curfsim
#print 'TTT', fsimE[-i-1] , fsim[0]
if fsimE[-i-1] < fsimR[-i-1]: # still better than the best
simFinal[-i-1, :] = simE[-i-1,:]
fsimFinal[-i-1] = fsimE[-i-1]
state[i] = -2
actives.remove(i)
else:
simFinal[-i-1, :] = simR[-i-1,:]
fsimFinal[-i-1] = fsimR[-i-1]
state[i] = -3
actives.remove(i)
elif state[i] == 3: # contraction1 just done
fsimC[-i-1] = curfsim
if fsimC[-i-1] < fsimR[-i-1]: # still better than the reflected
simFinal[-i-1, :] = simC[-i-1,:]
fsimFinal[-i-1] = fsimC[-i-1]
state[i] = -4
actives.remove(i)
else:
simFinal[-i-1, :] = sim[-i-1,:]
fsimFinal[-i-1] = fsim[-i-1]
shrinks.append(i)
state[i] = -10
actives.remove(i)
elif state[i] == 4: # contraction2 just done
fsimC[-i-1] = curfsim
if fsimC[-i-1] < fsim[-i-1]: # still better than the reflected
simFinal[-i-1, :] = simC[-i-1,:]
fsimFinal[-i-1] = fsimC[-i-1]
state[i] = -5
actives.remove(i)
else:
simFinal[-i-1, :] = sim[-i-1,:]
fsimFinal[-i-1] = fsim[-i-1]
shrinks.append(i)
state[i] = -10
actives.remove(i)
else:
print 'Weird'
for i in range(nactive):
sim[-i-1,:] = simFinal[-i-1,:]
fsim[-i-1] = fsimFinal[-i-1]
#print 'ST1', state
if len(shrinks) == nactive:
print 'Shrink....'
# no change was positive
for j in one2np1:
sim[j] = sim[0] + sigma * (sim[j] - sim[0])
applicator(j, sim[j, :])
for j in one2np1:
fsim[j] = pool.get(j)
ind = np.argsort(fsim)
sim = np.take(sim, ind, 0)
fsim = np.take(fsim, ind, 0)
if callback is not None:
callback(sim[0])
iterations += 1
if retall:
allvecs.append(sim)
#allvecs = allcalls
x = sim[0]
fval = np.min(fsim)
warnflag = 0
if fcalls[0] >= maxfun:
warnflag = 1
msg = _status_message['maxfev']
if disp:
print('Warning: ' + msg)
elif iterations >= maxiter:
warnflag = 2
msg = _status_message['maxiter']
if disp:
print('Warning: ' + msg)
else:
msg = _status_message['success']
if disp:
print(msg)
print(" Current function value: %f" % fval)
print(" Iterations: %d" % iterations)
print(" Function evaluations: %d" % fcalls[0])
result = OptimizeResult(fun=fval, nit=iterations, nfev=fcalls[0],
status=warnflag, success=(warnflag == 0), message=msg,
x=x)
if retall:
result['allvecs'] = allvecs
return result