def init(self, boot_spec):
ExpSwitcher.init(self, boot_spec)
streamels = boot_spec.get_observations().get_streamels()
if ((streamels.size != 1) or
(streamels[0]['kind'] != ValueFormats.Discrete)):
msg = 'I expect the observations to be symbols (one discrete var).'
raise UnsupportedSpec(msg)
self.lower = streamels[0]['lower']
nsymbols = int(1 + (streamels[0]['upper'] - self.lower))
if nsymbols > 10000:
msg = ('Are you sure you want to deal with %d symbols?'
% nsymbols)
raise Exception(msg)
self.y_stats = SymbolsStatistics(nsymbols, window=self.window)
class SymbolsStats(ExpSwitcher):
'''
A simple agent that estimates various statistics
of the observations.
'''
def __init__(self, beta, window):
ExpSwitcher.__init__(self, beta)
self.window = window
def init(self, boot_spec):
ExpSwitcher.init(self, boot_spec)
streamels = boot_spec.get_observations().get_streamels()
if ((streamels.size != 1) or
(streamels[0]['kind'] != ValueFormats.Discrete)):
msg = 'I expect the observations to be symbols (one discrete var).'
raise UnsupportedSpec(msg)
self.lower = streamels[0]['lower']
nsymbols = int(1 + (streamels[0]['upper'] - self.lower))
if nsymbols > 10000:
msg = ('Are you sure you want to deal with %d symbols?'
% nsymbols)
raise Exception(msg)
self.y_stats = SymbolsStatistics(nsymbols, window=self.window)
def process_observations(self, obs):
y = obs['observations']
dt = obs['dt'].item()
symbol = y[0].item() - self.lower
self.y_stats.update(symbol, dt)
def publish(self, pub):
self.y_stats.publish(pub.section('y_stats'))
def display_solution(name, caption, R, S):
D = euclidean_distances(S)
sec = pub.section(name)
sec.text('info', caption)
with sec.plot('S') as pl:
pl.plot(S[0, :], S[1, :], 'k.')
if S.shape[0] == 3:
import mpl_toolkits.mplot3d.axes3d as p3
for angle in [0, 45, 90]:
with sec.plot('S3d%s' % angle) as pl:
fig = pl.gcf()
ax = p3.Axes3D(fig)
ax.view_init(30, angle)
col = S[2, :]
ax.scatter(S[0, :], S[1, :], S[2, :], c=col)
ax.set_xlabel('X')
ax.set_ylabel('Y')
ax.set_zlabel('Z')
fig.add_axes(ax)
if False: # TODO: use cheaper solution for point cloud as in
# calib
with sec.plot('D_vs_R') as pl:
pl.plot(D.flat, R.flat, 'k.')
pl.xlabel('D')
pl.ylabel('R')
deltas = range(1, self.window)
for d in deltas:
R = self.y_stats.get_transition_matrix(delta=d)
S = self.get_embedding(R, ndim=2)
display_solution('delta%d' % d, 'Using delta=%d' % d, R, S)
S3 = self.get_embedding(R, ndim=3)
display_solution('delta%dS3' % d, 'Using delta=%d' % d, R, S3)
Ts = [self.y_stats.get_transition_matrix(delta=d) for d in deltas]
T = np.array(Ts)
R = np.sum(T, axis=0)
S = self.get_embedding(R, ndim=2)
display_solution('deltasum', 'Summing all of them', R, S)
S3 = self.get_embedding(R, ndim=3)
display_solution('deltasum3', 'Summing all of them', R, S3)
@contract(R='array[NxN]', ndim='int,>=2')
def get_embedding(self, R, ndim):
D = 1 - R
D = D * np.pi / D.max()
np.fill_diagonal(D, 0)
# if pub is not None:
# P = D * D
# B = double_center(P)
# plot_spectrum(pub, 'B', B)
S = mds(D, ndim=ndim)
return S
