def rotate(M):
"""Rotates a RTBM model by a Pi/2 angle in counterclockwise direction.
"""
nh = M._Nh
nv = M._Nv
R = np.array([[0, -1], [1, 0]])
Rt = R.T
Rpinv = sp.linalg.inv(Rt.dot(R)).dot(Rt)
t = sp.linalg.inv(R.dot(sp.linalg.inv(M.t).dot(R.T)))
w = Rpinv.T.dot(M.w).reshape(1, nh * nv)
bv = Rpinv.T.dot(M.bv)
rM = rtbm.RTBM(nv, nh)
params = np.concatenate(
(bv, M.bh, w, t[np.triu_indices(nv)], M.q[np.triu_indices(nh)]),
axis=None)
rM.set_parameters(params)
return rM
def conditional(model, d):
"""Generates the conditional RTBM.
Args:
model (theta.rtbm.RTBM): RTBM modelling the original PDF
d (numpy.array): column vector containing the values for the conditional
Returns:
theta.rtbm.RTBM: RTBM modelling the conditional probability P(y|d)
"""
nh = model._Nh
nv = model._Nv
assert (nv > 1), "cannot do the conditional probability of a 1d distribution"
assert (d.size < nv), "d larger than Nv"
k = int(nv-d.size)
cmodel = rtbm.RTBM(k, nh)
# Matrix A
t = model.t[:k,:k]
t = t[np.triu_indices(k)]
q = model.q[np.triu_indices(nh)]
w = model.w[:k]
# Biases
bh = model.bh + np.dot(model.w[k:].T, d)
bv = model.bv[:k] + np.dot(model.t[:k,k:], d)
cparams = np.concatenate((bv, bh, w, t, q), axis = None)
cmodel.set_parameters(cparams)
return cmodel
sys.path.append('../../main')
import conditional as con
import utils as ut
import numpy as np
import matplotlib.pyplot as plt
from theta import rtbm, minimizer
from theta.costfunctions import logarithmic
# Loading data
n = 5000
data = np.load('gaussian-mix-sample.npy')
# RTBM model
M = rtbm.RTBM(2, 4, random_bound=1, init_max_param_bound=30)
# Training
minim = minimizer.CMA(True)
#solution = minim.train(logarithmic(), M, data.T, tolfun=1e-3) # decomment if you want training to take place,
# otherwise use the pre-trained model below
# Pre-trained model
M.set_parameters(
np.array([
1.02735619e-05, -3.63860337e-04, 2.15571853e+00, -2.96561961e-01,
1.47561952e+01, 9.63622840e-01, -1.03625774e+01, -8.31445704e+00,
-2.55729017e-01, -7.01183003e+00, 4.09970744e+00, -3.38018523e+00,
1.70457837e+00, 6.56773389e-01, 1.49908760e+01, 1.15771402e+00,
1.50000000e+01, 1.50000000e+01, 4.59256632e+00, 3.45065143e-01,
2.12470448e+00, 1.50000000e+01, 4.29228972e-02, -1.66464616e+00,
for i in range(n):
if u[i] < 0.6:
v[i] = np.random.normal(-0.7, 0.3)
elif u[i] < 0.7:
v[i] = np.random.normal(0.2, .1)
else:
v[i] = np.random.normal(.5, .5)
return v
n = 10000
data = gaussian_mixture(n)
data = data.reshape(1, data.size)
# RTBM model
model = rtbm.RTBM(1, 2, random_bound=1)
# Training
minim = CMA_es(KullbackLeibler(model), model, data, empirical_cost=True)
#minim = AMSGrad(KullbackLeibler(model), model, data, empirical_cost=True)
solution = minim.train(popsize=30, m2=20) # this is for CMA_es
#solution = minim.train() # this is for AMSGrad
# Plotting
figure = plt.axes()
figure.hist(data.flatten(), bins=50, density=True, color='grey')
x = np.linspace(figure.get_xlim()[0],
figure.get_xlim()[1], 100).reshape(1, 100)
data = student.sampling()
# Building grid for evaluation
n = 50
xx = np.linspace(-5,5, n)
yy = np.linspace(-5,5, n)
X,Y = np.meshgrid(xx,yy)
grid = np.vstack([X,Y]).reshape(2,-1)
# Analytic pdf
pdf = np.asarray([student.pdf(grid.T[i]) for i in range(n**2)]).flatten()
# RTBM model
M = rtbm.RTBM(2,2, random_bound=1, init_max_param_bound=60)
# Training
minim = minimizer.CMA(True)
#solution = minim.train(logarithmic, M, data.T, tolfun=1e-3) # decomment if you want training to take place,
# otherwise use the pre-trained model below
# Pre-trained model
M.set_parameters(np.array([ 1.74587313e-03, -9.13477122e-04, 8.22063600e+00, 1.74029397e+01,-1.10959475e+00, 1.02118882e+00, -6.59149043e-01, 5.99301112e-01,5.57746556e-01, 1.84310133e-01, 3.04781259e-01, 2.41516805e+01,-4.40890318e-01, 4.15665779e+01]))
# Analytic conditional
x = np.array([-1, 0.1, 1.9]) # fixed x at which we calculate the conditional probability p(y|x)
cstudent1 = student.conditional(x[0].reshape(1,1))