def get_path_rnd(x, initpv=valley, maxtrials=300):
pv = [initpv]
count = 0
while count < maxtrials:
pv += [cf.get_next_pv(pv[-1][0], pv[-1][1], get_a_rnd(x, *pv[-1]))]
count += 1
return np.array(pv)
def get_path_rnd(x, initpv=valley, maxtrials=300):
pv = [initpv]
count = 0
while count < maxtrials:
pv += [cf.get_next_pv(pv[-1][0], pv[-1][1], get_a_rnd(x, *pv[-1]))]
count += 1
return np.array(pv)
def parallel_sampling_keepU(self, step, eta, run, rate, T=500, r0=.5, tm=20, ts=2,
reset=0, gamma=1, trials=1000, mode='fix', maxsteps=300,
initpv=None, initW=None, samples=100):
np.random.seed(run)
if initW is None:
self.unlearn(r0, mode)
else:
self.W = initW
if isinstance(T, (int, long, float, complex)):
Tmax = T
else:
Tmax = T[1]
seq = [[None]] * trials # sequences might have differnt length -> list instead array
scount = np.zeros((trials, self.K))
uinit = np.zeros(self.K)
uinit[self.r] = rate / 1000.
for t in xrange(trials):
print 'run', run, ' trial', t, '/', trials
stdout.flush()
pvls = [[8, 16]] if initpv is None else [initpv]
a = []
r = []
res = cfn.runpopU_js(self.W / self.pop_size, uinit, step, self.pop_size, rate,
Tmax, tm, ts, 1. * reset / self.pop_size, run)
uinit = res[1]
scount[t] = np.sum(res[0], axis=0)
for counter in xrange(maxsteps):
a += [self.get_a(scount[t], *pvls[-1])]
pvls += [cf.get_next_pv(pvls[-1][0], pvls[-1][1], a[-1])]
r += [cf.get_R(*pvls[-1])]
seq[t] = [pvls[:-1], a, r]
for i in range(samples):
p = 16 * np.random.rand()
v = 32 * np.random.rand()
for a in range(3):
pvs = cf.get_next_pvs(p, v, a, self.steps)
rr = cf.get_R(*pvs)
self.update_weights_continuous([p, v], a, rr, pvs,
eta, gamma**self.steps)
return np.array([scount, np.array(seq), np.copy(self.W)])
