class CovarianceRemember(Block):
""" Quick hack to remember covariance across executions """
Block.alias("covariance_rem")
Block.config("filename", default="cov_remember.pickle")
Block.input("x")
Block.output("cov_x")
def init(self):
filename = self.config.filename
if os.path.exists(filename):
self.info("Loading state from filename %r" % filename)
self.state = safe_pickle_load(filename)
else:
from boot_agents.utils.mean_covariance import MeanCovariance
self.state = MeanCovariance()
def update(self):
x = self.input.x.astype("float32")
self.state.update(x)
self.output.cov_x = self.state.get_covariance()
safe_pickle_dump(self.state, self.config.filename) # @UndefinedVariable
def init(self, boot_spec):
ExpSwitcher.init(self, boot_spec)
# TODO: check float
if len(boot_spec.get_observations().shape()) != 1:
raise UnsupportedSpec('I assume 2D signals.')
self.yy = Expectation()
self.ylogy = Expectation()
self.einvy = Expectation()
self.ey = Expectation()
self.elogy = Expectation()
self.covy = MeanCovariance()
self.covfy = MeanCovariance()
def init(self):
filename = self.config.filename
if os.path.exists(filename):
self.info("Loading state from filename %r" % filename)
self.state = safe_pickle_load(filename)
else:
from boot_agents.utils.mean_covariance import MeanCovariance
self.state = MeanCovariance()
class EstStatsTh(ExpSwitcher):
'''
'''
def init(self, boot_spec):
ExpSwitcher.init(self, boot_spec)
# TODO: check float
if len(boot_spec.get_observations().shape()) != 1:
raise UnsupportedSpec('I assume 2D signals.')
self.yy = Expectation()
self.ylogy = Expectation()
self.einvy = Expectation()
self.ey = Expectation()
self.elogy = Expectation()
self.covy = MeanCovariance()
self.covfy = MeanCovariance()
def merge(self, other):
self.yy.merge(other.yy)
self.ylogy.merge(other.ylogy)
self.einvy.merge(other.einvy)
self.ey.merge(other.ey)
self.elogy.merge(other.elogy)
self.covy.merge(other.covy)
self.covfy.merge(other.covfy)
def process_observations(self, obs):
y = obs['observations']
n = y.size
# # XXX
# which = np.array(range(y.size)) < 100
# y[which] = (y * y)[which]
dt = obs['dt'].item()
z = y == 0
y[z] = 0.5
yy = outer(y, y)
dt = 1
logy = np.log(y)
# TMP
logy = y * y
# logy[:int(n / 4)] = y[:int(n / 4)]
ylogy = outer(logy, y)
self.yy.update(yy, dt)
self.ylogy.update(ylogy, dt)
invy = 1.0 / y
self.einvy.update(invy, dt)
self.ey.update(y, dt)
self.elogy.update(logy, dt)
self.covy.update(y, dt)
self.covfy.update(logy, dt)
def publish(self, pub):
# pub.text('warn', 'using y^3')
yy = self.yy.get_value()
ylogy = self.ylogy.get_value()
def symmetrize(x):
return 0.5 * (x + x.T)
yy = symmetrize(yy)
ylogy = symmetrize(ylogy)
with pub.subsection('yy') as sub:
sub.array_as_image('val', yy)
pub_svd_decomp(sub, yy)
pub_eig_decomp(sub, yy)
with pub.subsection('ylogy') as sub:
sub.array_as_image('ylogy', ylogy)
pub_svd_decomp(sub, ylogy)
pub_eig_decomp(sub, ylogy)
f = pub.figure()
with f.plot('both') as pylab:
pylab.plot(yy.flat, ylogy.flat, '.')
pylab.xlabel('yy')
pylab.ylabel('ylogy')
pylab.axis('equal')
D = get_deriv_matrix(yy.shape[0])
inv = np.linalg.inv(yy)
x = np.dot(ylogy, inv)
f.array_as_image('inv', inv)
f.array_as_image('ylogy_times_inverse', x)
f.array_as_image('D', D)
f.array_as_image('Dlog', np.dot(D, x))
diag = np.diag(x)
einvy = self.einvy.get_value()
elogy = self.elogy.get_value()
ey = self.ey.get_value()
f = pub.figure()
with f.plot('einvy', caption='E{1/y}') as pylab:
pylab.plot(einvy, 's')
y_axis_set_min(pylab, 0)
with f.plot('elogy', caption='E{logy}') as pylab:
pylab.plot(elogy, 's')
# y_axis_set_min(pylab, 0)
with f.plot('ey', caption='E{y}') as pylab:
pylab.plot(ey, 's')
y_axis_set_min(pylab, 0)
with f.plot('diag', caption='diagonal of x') as pylab:
pylab.plot(diag, 's')