def calcColNormalizer(inMatrix):
#Theano function for calculating logSum, i.e., calculate ln(X + Y) based on ln(X) and ln(Y).
maxExp = -4950.0
x, y = T.fscalars(2)
yMinusx = y - x ## this part is for the condition which x > y
xMinusy = x - y # if x < y
bigger = T.switch(T.gt(x, y), x, y)
YSubtractX = T.switch(T.gt(x,y), yMinusx, xMinusy)
x_prime = T.log(1 + T.exp(YSubtractX)) + bigger
calcSum = T.switch(T.lt(YSubtractX, maxExp), bigger, x_prime)
logSum = function([x, y], calcSum, allow_input_downcast=True)
####### end of logSum ###############
# now we caclculate sum of log joint as normalizer
if len(inMatrix.shape) < 2:
raise Exception ("calcColNormalizer expect a 2D matrix")
nRows, nCols = inMatrix.shape
columnAccumLogSum = np.zeros(nCols)
for col in range(nCols):
currLogSum = np.NINF
for j in range(nRows):
if inMatrix[j,col] == np.NINF:
continue
currLogSum = logSum(currLogSum, inMatrix[j,col])
columnAccumLogSum[col] = currLogSum
return columnAccumLogSum
def advance(u, u_1, u_2, f_a, Cx2, Cy2, dt2, V=None, step1=False):
u_in, u_1_in, u_2_in = T.fmatrices('u_in','u_1_in','u_2_in')
f_a_in, V_in = T.fmatrices('f_in','V_in')
step1_in = T.lscalar('step1_in')
Cx2_in, Cy2_in, dt2_in = T.fscalars('Cx2_in','Cy2_in','dt2_in')
u_out = T.fmatrix('u_out')
if V is None:
V = np.zeros_like(f_a)
step_f = theano.function([u_in, u_1_in, u_2_in, f_a_in, Cx2_in, Cy2_in, dt2_in, V_in, step1_in], u_out, step, on_unused_input='ignore')
u_out = step_f(u, u_1, u_2, f_a, Cx2, Cy2, dt2, V, step1)
return u_out
import sys, os, math
import NamedMatrix
import scipy.special as ss
from utilTCI import *
###############################################################################################
"""
The following block define the Theano functions that going to be used in the regular python funcitons
"""
"""
1. Calculate ln(X + Y) based on ln(X) and ln(Y) using theano library
"""
########### Theano function for calculating logSum
maxExp = -4950.0
x, y = T.fscalars(2)
yMinusx = y - x ## this part is for the condition which x > y
xMinusy = x - y # if x < y
bigger = T.switch(T.gt(x, y), x, y)
YSubtractX = T.switch(T.gt(x,y), yMinusx, xMinusy)
x_prime = T.log(1 + T.exp(YSubtractX)) + bigger
calcSum = T.switch(T.lt(YSubtractX, maxExp), bigger, x_prime)
logSum = function([x, y], calcSum, allow_input_downcast=True)
####### end of logSum ###############
def calcTDI (mutcnaMatrixFN, degMatrixFN, alphaNull = [1, 1], alphaIJKList = [2, 1, 1, 2],
v0=0.3, outputPath = ".", opFlag = None, rowBegin=0, rowEnd = None, GeGlobalDriverDict = None):
"""
# Multiple outputs
a, b = T.fmatrices('a', 'b')
diff = a - b
absDiff = abs(diff)
sqrDiff = diff**2
f = function([a, b], [diff, absDiff, sqrDiff])
mat2 = [[10, 5], [5, 10]]
mat3 = [[5, 10], [10, 5]]
print(f(mat2, mat3))
# Default values
x, y = T.fscalars('x', 'y')
z = x + y
f = function([x, In(y, value=0)], z)
print(f(20))
print(f(20, 10))
# Shared variables
state = shared(0)
inc = T.iscalar('inc')
accumulator = function([inc], state, updates=[(state, state + inc)])
print("state : ", state.get_value())
# Multiple outputs
a, b = T.fmatrices('a', 'b')
diff = a - b
absDiff = abs(diff)
sqrDiff = diff ** 2
f = function([a, b], [diff, absDiff, sqrDiff])
mat2 = [[10, 5], [5, 10]]
mat3 = [[5, 10], [10, 5]]
print(f(mat2, mat3))
# Default values
x, y = T.fscalars('x', 'y')
z = x + y
f = function([x, In(y, value=0)], z)
print(f(20))
print(f(20, 10))
# Shared variables
state = shared(0)
inc = T.iscalar('inc')
accumulator = function([inc], state, updates=[(state, state + inc)])
print("state : ", state.get_value())
x = T.fvector('x')
x1 = T.fscalar('x1')
y = 1/(1 + T.exp(-x))
y1 = 1/(1 + T.exp(-x1))
logistic = function([x], y)
logistic1 = function([x1], y1)
grady = T.grad(y1, x1)
derivada = function([x1], grady)
a = float(input('Introduce el extremo izqdo. \n'))
b = float(input('Introduce el extremo drcho. \n'))
particion = float(input('Introduce la longitud de particion del intervalo. \n'))
pderiv = float(input('Introduce el punto donde hallar su recta tangente. \n'))
xval = arange(a,b,particion, dtype='float32')
z,w,w1=T.fscalars('z', 'w', 'w1')
rectatg2 = (x-z)*w+w1
rectatg3 = function([x, Param(z, default=pderiv), Param(w, default=derivada(pderiv)), Param(w1, default=logistic1(pderiv))], rectatg2)
figure(1)
plot(xval, logistic(xval), linewidth=1.5, color='r')
plot(xval, rectatg3(xval), linewidth=1.0, color='g')
ylim([0,1])
xlabel(r'\textbf{Abcisa}', fontsize=12)
ylabel(r'\textit{Ordenada}',fontsize=12)
title(r"Funcion logistica f(x) = $\displaystyle\frac{1}{1+e^{-x}}$", fontsize=12, color='r')
legend(('Funcion Logistica', 'Recta Tangente'),'upper left', shadow=True, fancybox=True)
leg = gca().get_legend()
def __init__(self, inf=1e37):
pos, vel = T.fmatrices(['pos', 'vel'])
nc, N, n_steps = T.iscalars(['nc', 'N', 'n_steps'])
ra, rb, re, r0 = T.fscalars(['ra', 'rb', 're', 'r0'])
v0, j, b = T.fscalars(['v0', 'J', 'b'])
nu = trng.uniform(size=(N, 2), low=0.0, high=3.14159, dtype='floatX')
def distance_tensor(X):
E = X.reshape((X.shape[0], 1, -1)) - X.reshape((1, X.shape[0], -1))
D = T.sqrt(T.sum(T.square(E), axis=2))
return D
def direction_tensor(X):
E = X.reshape((X.shape[0], 1, -1)) - X.reshape((1, X.shape[0], -1))
L = T.sqrt(T.sum(T.square(E), axis=2))
L = T.pow(L + T.identity_like(L), -1)
L = T.stack([L, L, L], axis=2)
return L * E
def neighbourhood(X):
D = distance_tensor(X)
N = T.argsort(D, axis=0)
mask = T.cast(T.lt(N, nc), 'float32')
return N[1:nc + 1], mask
def alignment(X, Y):
n, d = neighbourhood(X)
return T.sum(Y[n], axis=0)
def cohesion(X, inf=100.0):
D = distance_tensor(X)
E = direction_tensor(X)
n, d = neighbourhood(X)
F = T.zeros_like(E)
D = T.stack([D, D, D], axis=2)
d = T.stack([d, d, d], axis=2)
c1 = T.lt(D, rb)
c2 = T.and_(T.gt(D, rb), T.lt(D, ra))
c3 = T.and_(T.gt(D, ra), T.lt(D, r0))
F = T.set_subtensor(F[c1], -E[c1])
F = T.set_subtensor(F[c2], 0.25 * (D[c2] - re) / (ra - re) * E[c2])
F = T.set_subtensor(F[c3], E[c3])
return T.sum(d * F, axis=0)
def perturbation(nu=nu):
phi = nu[:, 0]
theta = 2.0 * nu[:, 1]
return T.stack([
T.sin(theta) * T.sin(phi),
T.cos(theta) * T.sin(phi),
T.cos(phi)
],
axis=1)
def step(X, dX):
X_ = X + dX
V_ = j * nc / v0 * (alignment(
X, dX)) + b * (cohesion(X)) + nc * (perturbation())
dV = T.sqrt(T.sum(T.square(V_), axis=1)).reshape(V_.shape[0], 1)
dV = T.stack([dV, dV, dV], axis=1)
V = v0 * V_ / dV
return T.cast(X_, 'float32'), T.cast(V, 'float32')
def probability(X, Y):
n, d = neighbourhood(X)
vDv = T.batched_dot(Y[n].swapaxes(0, 1), Y)
p = T.exp((j / 2.0) * T.sum(vDv, axis=1))
return p / T.sum(p)
sim, update = theano.scan(step,
outputs_info=[pos, vel],
n_steps=n_steps)
pos_, vel_ = sim
mean_final_velocity = 1 / (N * v0) * T.sqrt(
T.sum(T.square(T.sum(vel_[-1], axis=0))))
particle_probability = probability(pos_[-1], vel_[-1])
self.f = theano.function(
[pos, vel, nc, ra, rb, r0, re, j, v0, b, N, n_steps], [pos_, vel_],
allow_input_downcast=True)
self.g = theano.function(
[pos, vel, nc, ra, rb, r0, re, j, v0, b, N, n_steps],
mean_final_velocity,
allow_input_downcast=True)
self.h = theano.function(
[pos, vel, nc, ra, rb, r0, re, j, v0, b, N, n_steps],
particle_probability,
allow_input_downcast=True)
import theano
import theano.tensor as T
import numpy as np
from scipy.stats import norm
from theano import function
import _pickle as cPickle
# dictionary with the distributions
data = cPickle.load(
open('C:\\Users\\user\\Desktop\\Master-Project-master\\save.pkl', 'rb'))
dict = {}
words = ['ich', 'das', 'ist', 'du', 'die']
# TensorVariables for mi, mj, si, sj respectivelly.
a, b = T.fscalars('a', 'b') # mu
c, d = T.fscalars('c', 'd') # sigma
# Energy as a TensorVariable
E = 0.5 * (d / c + (a - b)**2 / c - 1 - T.log(d / c))
enrg = function([a, b, c, d], E)
def KLdivergence(data):
dictKLword = []
distList = []
for i in range(len(data)):
mu, std = norm.fit(data[i])
var = np.power(std, 2)
distList.append([mu, var])
x = T.fvector("x")
x1 = T.fscalar("x1")
y = 1/(1 + T.exp(-x))
y1 = 1/(1 + T.exp(-x1))
logistic = function([x], y)
logistic1 = function([x1], y1)
grady = T.grad(y1, x1)
derivada = function([x1], grady)
a = float(input("Introduce el extremo izqdo. \n"))
b = float(input("Introduce el extremo drcho. \n"))
particion = float(input("Introduce la longitud de particion del intervalo. \n"))
pderiv = float(input("Introduce el punto donde hallar su recta tangente. \n"))
xval = arange(a,b,particion, dtype="float32")
z,w,w1=T.fscalars("z", "w’, ‘w1′)
rectatg2 = (x-z)*w+w1
rectatg3 = function([x, Param(z, default=pderiv), Param(w, default=derivada(pderiv)), Param(w1, default=logistic1(pderiv))], rectatg2)
figure(1)
plot(xval, logistic(xval), linewidth=1.5, color=’r")
plot(xval, rectatg3(xval), linewidth=1.0, color=’g")
ylim([0,1])
xlabel(r’\textbf{Abcisa}’, fontsize=12)
ylabel(r’\textit{Ordenada}’,fontsize=12)
title(r”Funcion logistica f(x) = $\displaystyle\frac{1}{1+e^{-x}}$”, fontsize=12, color=’r")
legend((‘Funcion Logistica’, ‘Recta Tangente’),’upper left’, shadow=True, fancybox=True)
leg = gca().get_legend()
def __init__(self,
N=None,
size=1,
mu0=0.1,
sigma_mean0=10,
sigma_std0=1.0,
sigma_min=0.1,
sigma_max=10,
data=None):
self.N = N
self.K = size
# Parameter initialization
#random init
if data is None:
# mu = random normal with std mu0,mean 0
self.mu = mu0 * np.random.randn(self.N, self.K).astype(DTYPE)
# Sigma = random normal with mean sigma_mean0, std sigma_std0, and min/max of sigma_min, sigma_max
self.Sigma = np.random.randn(self.N, 1).astype(DTYPE)
self.Sigma *= sigma_std0
self.Sigma += sigma_mean0
self.Sigma = np.maximum(sigma_min,
np.minimum(self.Sigma, sigma_max))
self.Gaussian = np.concatenate((self.mu, self.Sigma), axis=1)
# TensorVariables for mi, mj, si, sj respectivelly.
a, b = T.fvectors('a', 'b')
c, d = T.fscalars('c', 'd')
# Energy as a TensorVariable
E = -0.5 * (self.K * d / c + T.sum(
(a - b)**2 / c) - self.K - self.K * T.log(d / c))
self.enrg = function([a, b, c, d], E)
g1 = T.grad(E, a) # dE/dmi
self.f1 = function([a, b, c, d], g1)
g2 = T.grad(E, b) # dE/dmj
self.f2 = function([a, b, c, d], g2)
g3 = T.grad(E, c) # dE/dsi
self.f3 = function([a, b, c, d], g3)
g4 = T.grad(E, d) # dE/dsj
self.f4 = function([a, b, c, d], g4)
#non random init
else:
self.mu = []
self.Sigma = []
for i in range(len(data)):
mu, std = norm.fit(data[i])
var = np.power(std, 2)
self.mu.append(mu)
self.Sigma.append(var)
self.Gaussian = np.concatenate(
(np.asarray(self.mu), np.asarray(self.Sigma)), axis=1)
self.Gaussian = np.reshape(self.Gaussian, (2, N)).T
