def _generate_sample_from_state(self, state, random_state=None):
"""Generate random samples from the output distribution of a state.
Returns
-------
sample : mp.array, shape (`n_features`,)
"""
x = vonmises.rvs(self._kappas_[state], self._means_[state])
return circwrap(x)
def create_timeseries(means, kappas, transmat):
"""Construct a random timeseries based on a specified Markov model."""
numStates = len(means)
state = random.randint(0, numStates - 1)
cdf = np.cumsum(transmat, 1)
numFrames = 1000
X = np.empty((numFrames, 1))
for i in range(numFrames):
rand = random.random()
state = (cdf[state] > rand).argmax()
X[i, 0] = circwrap(vonmises.rvs(kappas[state], means[state]))
return X
def create_timeseries(means, kappas, transmat):
"""Construct a random timeseries based on a specified Markov model."""
numStates = len(means)
state = random.randint(0, numStates - 1)
cdf = np.cumsum(transmat, 1)
numFrames = 1000
X = np.empty((numFrames, 1))
for i in range(numFrames):
rand = random.random()
state = (cdf[state] > rand).argmax()
X[i, 0] = circwrap(vonmises.rvs(kappas[state], means[state]))
return X
def adjust(self, node):
""" sample from posterior """
D = np.vstack([d.pos for d in node.data ])
old=np.array([node.get_likelihood_angle(d) for d in node.data ])
old=np.log(old).mean()
#import ipdb as pdb; pdb.set_trace()
D -= node.parent.pos
D = D.T
old_angle = node.angle
# determine posterior of angle given prior and data
x_ang = np.arctan2(D[1,:],D[0,:])
R1 = self.angle_kappa * np.cos(node.parent.angle) + np.sum(np.cos(x_ang))
R2 = self.angle_kappa * np.sin(node.parent.angle) + np.sum(np.sin(x_ang))
#R1 = np.sum(np.cos(x_ang))
#R2 = np.sum(np.sin(x_ang))
mu = np.arctan2(R2,R1)
Rn = R1 / np.cos(mu)
node.angle = vonmises.rvs(Rn,loc=mu)
#node.angle = mu
X = D.copy()
# rotate around newly drawn angle
D = np.dot(rotmat(-node.angle), D)
assert D.shape[1] == len(node.data)
# determine posterior of length given prior and data
aN = self.length_a0 + D.shape[1]/2.0
bN = self.length_b0 + D.shape[1]/2.0 * np.var(D[0,:])
#aN = D.shape[1]/2.0
#bN = D.shape[1]/2.0 * np.var(D[0,:])
#import ipdb; ipdb.set_trace()
node.length = gamma.rvs(aN,loc=1.0/bN)
#node.length = np.var(D[0,:])
old_pos = node.pos.copy()
node.update_position()
def f():
plt.plot(X[0,:],X[1,:],".b")
#plt.plot(D[0,:],D[1,:],".r")
plt.plot(node.pos[0]-node.parent.pos[0],node.pos[1]-node.parent.pos[1],"*r")
plt.plot(old_pos[0]-node.parent.pos[0],old_pos[1]-node.parent.pos[1],"*b")
#plt.plot(node.parent.pos[0],node.parent.pos[1],"*r")
plt.show()
new=np.array([node.get_likelihood_angle(d) for d in node.data ])
new=np.log(new).mean()
print('old likelihood: %f new likelihood: %f'%(old,new))
