def fix(d, T, p, q, iters, epsilon, steps, nu):
start = time.time()
info = 'd=%d_T=%d_p=%0.3f_q=%0.3f_steps=%0.1E' % (d, T, p, q, steps)
name = 'plots/data_' + info
parent = "./"
directory = os.path.join(parent, name)
opd = rt_util.pkl_get(os.path.join(directory,'local_one_particle.pkl'))
full = rt_util.pkl_get(os.path.join(directory,'full_one_particle.pkl'))
plt.plot([full[p] for p in rt_util.bin_tuples(T)], label='full')
plt.plot([opd[p] for p in rt_util.bin_tuples(T)], label='local')
plt.legend(loc=2)
l1 = sum([abs(full[p] - opd[p]) for p in rt_util.bin_tuples(T)])
info = ('L1 distance = %0.4f ' % l1) + info.replace("_", " ")
plt.title(info)
plt.xlabel('One Particle Path')
plt.ylabel('Probability')
plt.savefig(os.path.join(directory,"fig.png"))
F = open(os.path.join(directory,"L1.txt"), "w")
F.write('%f' % l1)
F.close()
F = open(os.path.join(directory,"time.txt"), "w")
F.write('%f' % l1)
F.close()
end = time.time()
F = open(os.path.join(directory,"time.txt"), "w")
F.write('%f' % (end-start))
F.close()
def init_observations(T):
observations = [{} for t in range(T)]
for t in range(1, T):
for p1 in rt_util.bin_tuples(t):
for p2 in rt_util.bin_tuples(t):
observations[t][(p1, p2)] = 0.
return observations
def init_cond(T):
cond = [{} for t in range(T)]
for t in range(1, T):
for p1 in rt_util.bin_tuples(t):
for p2 in rt_util.bin_tuples(t):
cond[t][(p1, p2)] = {}
return cond
def rt_local_approx_iteration(d, T, p, q, steps, nu, cond):
f_new = {}
cond_new = init_cond(T)
one_particle_distr_new = init_one_particle_distr(T)
observed = init_observations(T)
bad = 0.
increment = 1.0 / steps
for step in range(steps):
# X[0] is root. X[1:d+1] is the d children.
X, b = rt_realization(d, T, p, q, nu, cond)
bad += b
# record the realization in f_new
if X.tostring() not in f_new:
f_new[X.tostring()] = increment
else:
f_new[X.tostring()] += increment
# record the realization in cond_new
for t in range(1, T):
for k in range(1, d + 1):
# TODO: See email clarification and change to t instead of t-1
# maybe.
other_children = tuple(X[t - 1, 1:k]) + \
tuple(X[t - 1, (k + 1):])
root_and_child = (tuple(X[:t, 0]), tuple(X[:t, k]))
observed[t][root_and_child] += 1
if other_children not in cond_new[t][root_and_child]:
cond_new[t][root_and_child][other_children] = 1
else:
cond_new[t][root_and_child][other_children] += 1
# Record realization in one_particle_distr
one_particle_distr_new[tuple(X[:, 0])] += increment
# normalize cond_new
for t in range(1, T):
for p1 in rt_util.bin_tuples(t):
for p2 in rt_util.bin_tuples(t):
p = (p1, p2)
for c in cond_new[t][p]:
if observed[t][p] != 0:
cond_new[t][p][c] = cond_new[t][p][c] / observed[t][p]
return f_new, cond_new, one_particle_distr_new, bad / steps
def compare(d, T, p, q, iters, epsilon, steps, nu):
start = time.time()
info = 'd=%d_T=%d_p=%0.3f_q=%0.3f_steps=%0.1E' % (d, T, p, q, steps)
name = 'data/data_' + info
parent = "./"
rt_util.make_directory(parent, name)
directory = os.path.join(parent, name)
info_file = os.path.join(directory, "info")
f, c, opd, res, op_res, bad = rt_local_approx.rt_local_approx(
d, T, p, q, iters, epsilon, steps, nu, info_file)
rt_util.pkl_save(f, os.path.join(directory, 'local_joint.pkl'))
rt_util.pkl_save(c, os.path.join(directory, 'local_cond.pkl'))
rt_util.pkl_save(opd, os.path.join(directory, 'local_one_particle.pkl'))
rt_util.pkl_save(res, os.path.join(directory, 'local_joint_residual.pkl'))
rt_util.pkl_save(op_res, os.path.join(
directory, 'local_one_particle_residual.pkl'))
full = rt_full_simulation.generate_distr(d, T, p, q, nu, steps)
rt_util.pkl_save(full, os.path.join(directory, 'full_one_particle.pkl'))
plt.plot([full[p] for p in rt_util.bin_tuples(T)], label='full')
plt.plot([opd[p] for p in rt_util.bin_tuples(T)], label='local')
plt.legend(loc=2)
l1 = sum([abs(full[p] - opd[p]) for p in rt_util.bin_tuples(T)])
info = ('L1 distance = %0.4f ' % l1) + info.replace("_", " ")
plt.title(info)
plt.xlabel('One Particle Path')
plt.ylabel('Probability')
plt.savefig(os.path.join(directory,"fig.png"))
F = open(os.path.join(directory,"L1.txt"), "w")
F.write('%f' % l1)
F.close()
F = open(os.path.join(directory,"time.txt"), "w")
F.write('%f' % l1)
F.close()
end = time.time()
F = open(os.path.join(directory,"time.txt"), "w")
F.write('%f' % (end-start))
F.close()
def generate_distr(d, T, p, q, nu, steps):
f = {}
for st in rt_util.bin_tuples(T):
f[st] = 0
inc = 1. / steps
for step in range(steps):
X = rt_full_simulation(d, T, p, q, steps, nu)
f[X] += inc
return f
def rt_local_approx(d, T, p, q, iters, epsilon, steps, nu, file=None):
# f is the joint distribution over the single neighborhood of the root
f = {}
# cond is the conditional law of d-1 children given the root and the dth
cond = init_cond(T)
one_particle_distr = init_one_particle_distr(T)
distances = []
one_particle_distances = []
if file:
F = open(file, "w")
for iteration in range(iters):
f_new, cond_new, one_particle_distr_new, bad = rt_local_approx_iteration(
d, T, p, q, steps, nu, cond)
# distance
l = 1
dist = 0
dist += sum([f_new[k]**l for k in f_new if k not in f])
dist += sum([f[k]**l for k in f if k not in f_new])
dist += sum([abs(f[k] - f_new[k])**l for k in f if k in f_new])
distances.append(dist)
op_dist = 0
op_dist = sum([
abs(one_particle_distr[k] - one_particle_distr_new[k])**l
for k in rt_util.bin_tuples(T)
])
one_particle_distances.append(op_dist)
info_str = 'iteration %d, Full distance %f, OP Distance %f, baddness %f' % (
iteration, dist, op_dist, bad)
if file:
F.write(info_str)
F.write('\n')
else:
print info_str
f = f_new
cond = cond_new
one_particle_distr = one_particle_distr_new
if dist < epsilon:
break
return f, cond, one_particle_distr, distances, one_particle_distances, bad
def init_one_particle_distr(T):
one_particle_distr = {}
for p in rt_util.bin_tuples(T):
one_particle_distr[p] = 0.
return one_particle_distr