Develop Reference

How can I stop my Runge-Kutta2 (Heun) method from exploding?

I am currently trying to write some python code to solve an arbitrary system of first order ODEs, using a general explicit Runge-Kutta method defined by the values alpha, gamma (both vectors of dimension m) and beta (lower triangular matrix of dimension m x m) of the Butcher table which are passed in by the user. My code appears to work for single ODEs, having tested it on a few different examples, but I'm struggling to generalise my code to vector valued ODEs (i.e. systems).
In particular, I try to solve a Van der Pol oscillator ODE (reduced to a first order system) using Heun's method defined by the Butcher Tableau values given in my code, but I receive the errors
"RuntimeWarning: overflow encountered in double_scalars f = lambda t,u: np.array(... etc)" and
"RuntimeWarning: invalid value encountered in add kvec[i] = f(t+alpha[i]*h,y+h*sum)"
followed by my solution vector that is clearly blowing up. Note that the commented out code below is one of the examples of single ODEs that I tried and is solved correctly. Could anyone please help? Here is my code:
import numpy as np
def rk(t,y,h,f,alpha,beta,gamma):
'''Runga Kutta iteration'''
return y + h*phi(t,y,h,f,alpha,beta,gamma)
def phi(t,y,h,f,alpha,beta,gamma):
'''Phi function for the Runga Kutta iteration'''
m = len(alpha)
count = np.zeros(len(f(t,y)))
kvec = k(t,y,h,f,alpha,beta,gamma)
for i in range(1,m+1):
count = count + gamma[i-1]*kvec[i-1]
return count
def k(t,y,h,f,alpha,beta,gamma):
'''returning a vector containing each step k_{i} in the m step Runga Kutta method'''
m = len(alpha)
kvec = np.zeros((m,len(f(t,y))))
kvec[0] = f(t,y)
for i in range(1,m):
sum = np.zeros(len(f(t,y)))
for l in range(1,i+1):
sum = sum + beta[i][l-1]*kvec[l-1]
kvec[i] = f(t+alpha[i]*h,y+h*sum)
return kvec
def timeLoop(y0,N,f,alpha,beta,gamma,h,rk):
'''function that loops through time using the RK method'''
t = np.zeros([N+1])
y = np.zeros([N+1,len(y0)])
y[0] = y0
t[0] = 0
for i in range(1,N+1):
y[i] = rk(t[i-1],y[i-1], h, f,alpha,beta,gamma)
t[i] = t[i-1]+h
return t,y
#################################################################
'''f = lambda t,y: (c-y)**2
Y = lambda t: np.array([(1+t*c*(c-1))/(1+t*(c-1))])
h0 = 1
c = 1.5
T = 10
alpha = np.array([0,1])
gamma = np.array([0.5,0.5])
beta = np.array([[0,0],[1,0]])
eff_rk = compute(h0,Y(0),T,f,alpha,beta,gamma,rk, Y,11)'''
#constants
mu = 100
T = 1000
h = 0.01
N = int(T/h)
#initial conditions
y0 = 0.02
d0 = 0
init = np.array([y0,d0])
#Butcher Tableau for Heun's method
alpha = np.array([0,1])
gamma = np.array([0.5,0.5])
beta = np.array([[0,0],[1,0]])
#rhs of the ode system
f = lambda t,u: np.array([u[1],mu*(1-u[0]**2)*u[1]-u[0]])
#solving the system
time, sol = timeLoop(init,N,f,alpha,beta,gamma,h,rk)
print(sol)

Your step size is not small enough. The Van der Pol oscillator with mu=100 is a fast-slow system with very sharp turns at the switching of the modes, so rather stiff. With explicit methods this requires small step sizes, the smallest sensible step size is 1e-5 to 1e-6. You get a solution on the limit cycle already for h=0.001, with resulting velocities up to 150.
You can reduce some of that stiffness by using a different velocity/impulse variable. In the equation
x'' - mu*(1-x^2)*x' + x = 0
you can combine the first two terms into a derivative,
mu*v = x' - mu*(1-x^2/3)*x
so that
x' = mu*(v+(1-x^2/3)*x)
v' = -x/mu
The second equation is now uniformly slow close to the limit cycle, while the first has long relatively straight jumps when v leaves the cubic v=x^3/3-x.
This integrates nicely with the original h=0.01, keeping the solution inside the box [-3,3]x[-2,2], even if it shows some strange oscillations that are not present for smaller step sizes and the exact solution.

How do I implement the Probability density function of a Gaussian Distribution

I need to implement a class in Python, that represents a Univariate (for now) Normal Distribution. What I have in mind is as follows
class Norm():
def __init__(self, mu=0, sigma_sq=1):
self.mu = mu
self.sigma_sq = sigma_sq
# some initialization if necessary
def sample(self):
# generate a sample, where the probability of the value
# of the sample being generated is distributed according
# a normal distribution with a particular mean and variance
pass
N = Norm()
N.sample()
The generated samples should be distributed according to the following probability density function
I know that scipy.stats and Numpy provide functions to do this, but I need to understand how these functions are implemented. Any help would be appreciated, thanks :)

I ended up using the advice by #sascha. I looked at both this wikipedia article and the Numpy source and found this randomkit.c file that implemented the functions rk_gauss (which implements the Box Muller Transform), rk_double and rk_random (which implements the Mersenne Twister Random Number Generator that simulates a Uniformly Distributed Random Variable, required by the Box Muller Transform).
I then adapted the Mersenne Twister Generator from here and implemented the Box Muller Transform to simulate a gaussian (more information about Random Twister Generator here).
Here is the code I ended up writing:
import numpy as np
from datetime import datetime
import matplotlib.pyplot as plt
class Distribution():
def __init__(self):
pass
def plot(self, number_of_samples=100000):
# the histogram of the data
n, bins, patches = plt.hist([self.sample() for i in range(number_of_samples)], 100, normed=1, facecolor='g', alpha=0.75)
plt.show()
def sample(self):
# dummy sample function (to be overridden)
return 1
class Uniform_distribution(Distribution):
# Create a length 624 list to store the state of the generator
MT = [0 for i in xrange(624)]
index = 0
# To get last 32 bits
bitmask_1 = (2 ** 32) - 1
# To get 32. bit
bitmask_2 = 2 ** 31
# To get last 31 bits
bitmask_3 = (2 ** 31) - 1
def __init__(self, seed):
self.initialize_generator(seed)
def initialize_generator(self, seed):
"Initialize the generator from a seed"
global MT
global bitmask_1
MT[0] = seed
for i in xrange(1,624):
MT[i] = ((1812433253 * MT[i-1]) ^ ((MT[i-1] >> 30) + i)) & bitmask_1
def generate_numbers(self):
"Generate an array of 624 untempered numbers"
global MT
for i in xrange(624):
y = (MT[i] & bitmask_2) + (MT[(i + 1 ) % 624] & bitmask_3)
MT[i] = MT[(i + 397) % 624] ^ (y >> 1)
if y % 2 != 0:
MT[i] ^= 2567483615
def sample(self):
"""
Extract a tempered pseudorandom number based on the index-th value,
calling generate_numbers() every 624 numbers
"""
global index
global MT
if index == 0:
self.generate_numbers()
y = MT[index]
y ^= y >> 11
y ^= (y << 7) & 2636928640
y ^= (y << 15) & 4022730752
y ^= y >> 18
index = (index + 1) % 624
# divide by 4294967296, which is the largest 32 bit number
# to normalize the output value to the range [0,1]
return y*1.0/4294967296
class Norm(Distribution):
def __init__(self, mu=0, sigma_sq=1):
self.mu = mu
self.sigma_sq = sigma_sq
self.uniform_distribution_1 = Uniform_distribution(datetime.now().microsecond)
self.uniform_distribution_2 = Uniform_distribution(datetime.now().microsecond)
# some initialization if necessary
def sample(self):
# generate a sample, where the value of the sample being generated
# is distributed according a normal distribution with a particular
# mean and variance
u = self.uniform_distribution_1.sample()
v = self.uniform_distribution_2.sample()
return ((self.sigma_sq**0.5)*((-2*np.log(u))**0.5)*np.cos(2*np.pi*v)) + self.mu
This works perfectly, and generates a pretty good Gaussian
Norm().plot(10000)

Using the Box-Müller method:
def sample(self):
x = np.random.uniform(0,1,[2])
z = np.sqrt(-2*np.log(x[0]))*np.cos(2*np.pi*x[1])
return z * self.sigma_sq + self.mu

Have I implemented Milstein's method/Euler-Maruyama correctly?

I have an stochastic differential equation (SDE) that I am trying to solve using Milsteins method but am getting results that disagree with experiment.
The SDE is
which I have broken up into 2 first order equations:
eq1:
eq2:
Then I have used the Ito form:
So that for eq1:
and for eq2:
My python code used to attempt to solve this is like so:
# set constants from real data
Gamma0 = 4000 # defines enviromental damping
Omega0 = 75e3*2*np.pi # defines the angular frequency of the motion
eta = 0 # set eta 0 => no effect from non-linear p*q**2 term
T_0 = 300 # temperature of enviroment
k_b = scipy.constants.Boltzmann
m = 3.1e-19 # mass of oscillator
# set a and b functions for these 2 equations
def a_p(t, p, q):
return -(Gamma0 - Omega0*eta*q**2)*p
def b_p(t, p, q):
return np.sqrt(2*Gamma0*k_b*T_0/m)
def a_q(t, p, q):
return p
# generate time data
dt = 10e-11
tArray = np.arange(0, 200e-6, dt)
# initialise q and p arrays and set initial conditions to 0, 0
q0 = 0
p0 = 0
q = np.zeros_like(tArray)
p = np.zeros_like(tArray)
q[0] = q0
p[0] = p0
# generate normally distributed random numbers
dwArray = np.random.normal(0, np.sqrt(dt), len(tArray)) # independent and identically distributed normal random variables with expected value 0 and variance dt
# iterate through implementing Milstein's method (technically Euler-Maruyama since b' = 0
for n, t in enumerate(tArray[:-1]):
dw = dwArray[n]
p[n+1] = p[n] + a_p(t, p[n], q[n])*dt + b_p(t, p[n], q[n])*dw + 0
q[n+1] = q[n] + a_q(t, p[n], q[n])*dt + 0
Where in this case p is velocity and q is position.
I then get the following plots of q and p:
I expected the resulting plot of position to look something like the following, which I get from experimental data (from which the constants used in the model are determined):
Have I implemented Milstein's method correctly?
If I have, what else might be wrong my process of solving the SDE that'd causing this disagreement with the experiment?

You missed a term in the drift coefficient, note that to the right of dp there are two dt terms. Thus
def a_p(t, p, q):
return -(Gamma0 - Omega0*eta*q**2)*p - Omega0**2*q
which is actually the part that makes the oscillator into an oscillator. With that corrected the solution looks like
And no, you did not implement the Milstein method as there are no derivatives of b_p which are what distinguishes Milstein from Euler-Maruyama, the missing term is +0.5*b'(X)*b(X)*(dW**2-dt).
There is also a derivative-free version of Milsteins method as a two-stage kind-of Runge-Kutta method, documented in wikipedia or the original in arxiv.org (PDF).
The step there is (vector based, duplicate into X=[p,q], K1=[k1_p,k1_q] etc. to be close to your conventions)
S = random_choice_of ([-1,1])
K1 = a(X )*dt + b(X )*(dW - S*sqrt(dt))
Xh = X + K1
K2 = a(Xh)*dt + b(Xh)*(dW + S*sqrt(dt))
X = X + 0.5 * (K1+K2)

Population Monte Carlo implementation

I am trying to implement the Population Monte Carlo algorithm as described in this paper (see page 78 Fig.3) for a simple model (see function model()) with one parameter using Python. Unfortunately, the algorithm doesn't work and I can't figure out what's wrong. See my implementation below. The actual function is called abc(). All other functions can be seen as helper-functions and seem to work fine.
To check whether the algorithm workds, I first generate observed data with the only parameter of the model set to param = 8. Therefore, the posterior resulting from the ABC algorithm should be centered around 8. This is not the case and I'm wondering why.
I would appreciate any help or comments.
# imports
from math import exp
from math import log
from math import sqrt
import numpy as np
import random
from scipy.stats import norm
# globals
N = 300 # sample size
N_PARTICLE = 300 # number of particles
ITERS = 5 # number of decreasing thresholds
M = 10 # number of words to remember
MEAN = 7 # prior mean of parameter
SD = 2 # prior sd of parameter
def model(param):
recall_prob_all = 1/(1 + np.exp(M - param))
recall_prob_one_item = np.exp(np.log(recall_prob_all) / float(M))
return sum([1 if random.random() < recall_prob_one_item else 0 for item in range(M)])
## example
print "Output of model function: \n" + str(model(10)) + "\n"
# generate data from model
def generate(param):
out = np.empty(N)
for i in range(N):
out[i] = model(param)
return out
## example
print "Output of generate function: \n" + str(generate(10)) + "\n"
# distance function (sum of squared error)
def distance(obsData,simData):
out = 0.0
for i in range(len(obsData)):
out += (obsData[i] - simData[i]) * (obsData[i] - simData[i])
return out
## example
print "Output of distance function: \n" + str(distance([1,2,3],[4,5,6])) + "\n"
# sample new particles based on weights
def sample(particles, weights):
return np.random.choice(particles, 1, p=weights)
## example
print "Output of sample function: \n" + str(sample([1,2,3],[0.1,0.1,0.8])) + "\n"
# perturbance function
def perturb(variance):
return np.random.normal(0,sqrt(variance),1)[0]
## example
print "Output of perturb function: \n" + str(perturb(1)) + "\n"
# compute new weight
def computeWeight(prevWeights,prevParticles,prevVariance,currentParticle):
denom = 0.0
proposal = norm(currentParticle, sqrt(prevVariance))
prior = norm(MEAN,SD)
for i in range(len(prevParticles)):
denom += prevWeights[i] * proposal.pdf(prevParticles[i])
return prior.pdf(currentParticle)/denom
## example
prevWeights = [0.2,0.3,0.5]
prevParticles = [1,2,3]
prevVariance = 1
currentParticle = 2.5
print "Output of computeWeight function: \n" + str(computeWeight(prevWeights,prevParticles,prevVariance,currentParticle)) + "\n"
# normalize weights
def normalize(weights):
return weights/np.sum(weights)
## example
print "Output of normalize function: \n" + str(normalize([3.,5.,9.])) + "\n"
# sampling from prior distribution
def rprior():
return np.random.normal(MEAN,SD,1)[0]
## example
print "Output of rprior function: \n" + str(rprior()) + "\n"
# ABC using Population Monte Carlo sampling
def abc(obsData,eps):
draw = 0
Distance = 1e9
variance = np.empty(ITERS)
simData = np.empty(N)
particles = np.empty([ITERS,N_PARTICLE])
weights = np.empty([ITERS,N_PARTICLE])
for t in range(ITERS):
if t == 0:
for i in range(N_PARTICLE):
while(Distance > eps[t]):
draw = rprior()
simData = generate(draw)
Distance = distance(obsData,simData)
Distance = 1e9
particles[t][i] = draw
weights[t][i] = 1./N_PARTICLE
variance[t] = 2 * np.var(particles[t])
continue
for i in range(N_PARTICLE):
while(Distance > eps[t]):
draw = sample(particles[t-1],weights[t-1])
draw += perturb(variance[t-1])
simData = generate(draw)
Distance = distance(obsData,simData)
Distance = 1e9
particles[t][i] = draw
weights[t][i] = computeWeight(weights[t-1],particles[t-1],variance[t-1],particles[t][i])
weights[t] = normalize(weights[t])
variance[t] = 2 * np.var(particles[t])
return particles[ITERS-1]
true_param = 9
obsData = generate(true_param)
eps = [15000,10000,8000,6000,3000]
posterior = abc(obsData,eps)
#print posterior

I stumbled upon this question as I was looking for pythonic implementations of PMC algorithms, since, quite coincidentally, I'm currently in the process of applying the techniques in this exact paper to my own research.
Can you post the results you're getting? My guess is that 1) you're using a poor choice of distance function (and/or similarity thresholds), or 2) you're not using enough particles. I may be wrong here (I'm not very well-versed in sample statistics), but your distance function implicitly suggests to me that the ordering of your random draws matters. I'd have to think about this more to determine whether it actually has any effect on the convergence properties (it may not), but why don't you simply use the mean or median as your sample statistic?
I ran your code with 1000 particles and a true parameter value of 8, while using the absolute difference between sample means as my distance function, for three iterations with epsilons of [0.5, 0.3, 0.1]; the peak of my estimated posterior distribution seems to be approaching 8 just like it should on each iteration, alongside a reduction in the population variance. Note that there is still a noticeable rightward bias, but this is because of the asymmetry of your model (parameter values of 8 or less can never result in more than 8 observed successes, while all parameters values greater than 8 can, leading to a rightward skewedness in the distribution).
Here's the plot of my results:

Python Implementation of Viterbi Algorithm

I'm doing a Python project in which I'd like to use the Viterbi Algorithm. Does anyone know of a complete Python implementation of the Viterbi algorithm? The correctness of the one on Wikipedia seems to be in question on the talk page. Does anyone have a pointer?

Here's mine. Its paraphrased directly from the psuedocode implemenation from wikipedia. It uses numpy for conveince of their ndarray but is otherwise a pure python3 implementation.
import numpy as np
def viterbi(y, A, B, Pi=None):
"""
Return the MAP estimate of state trajectory of Hidden Markov Model.
Parameters
----------
y : array (T,)
Observation state sequence. int dtype.
A : array (K, K)
State transition matrix. See HiddenMarkovModel.state_transition for
details.
B : array (K, M)
Emission matrix. See HiddenMarkovModel.emission for details.
Pi: optional, (K,)
Initial state probabilities: Pi[i] is the probability x[0] == i. If
None, uniform initial distribution is assumed (Pi[:] == 1/K).
Returns
-------
x : array (T,)
Maximum a posteriori probability estimate of hidden state trajectory,
conditioned on observation sequence y under the model parameters A, B,
Pi.
T1: array (K, T)
the probability of the most likely path so far
T2: array (K, T)
the x_j-1 of the most likely path so far
"""
# Cardinality of the state space
K = A.shape[0]
# Initialize the priors with default (uniform dist) if not given by caller
Pi = Pi if Pi is not None else np.full(K, 1 / K)
T = len(y)
T1 = np.empty((K, T), 'd')
T2 = np.empty((K, T), 'B')
# Initilaize the tracking tables from first observation
T1[:, 0] = Pi * B[:, y[0]]
T2[:, 0] = 0
# Iterate throught the observations updating the tracking tables
for i in range(1, T):
T1[:, i] = np.max(T1[:, i - 1] * A.T * B[np.newaxis, :, y[i]].T, 1)
T2[:, i] = np.argmax(T1[:, i - 1] * A.T, 1)
# Build the output, optimal model trajectory
x = np.empty(T, 'B')
x[-1] = np.argmax(T1[:, T - 1])
for i in reversed(range(1, T)):
x[i - 1] = T2[x[i], i]
return x, T1, T2

I found the following code in the example repository of Artificial Intelligence: A Modern Approach. Is something like this what you're looking for?
def viterbi_segment(text, P):
"""Find the best segmentation of the string of characters, given the
UnigramTextModel P."""
# best[i] = best probability for text[0:i]
# words[i] = best word ending at position i
n = len(text)
words = [''] + list(text)
best = [1.0] + [0.0] * n
## Fill in the vectors best, words via dynamic programming
for i in range(n+1):
for j in range(0, i):
w = text[j:i]
if P[w] * best[i - len(w)] >= best[i]:
best[i] = P[w] * best[i - len(w)]
words[i] = w
## Now recover the sequence of best words
sequence = []; i = len(words)-1
while i > 0:
sequence[0:0] = [words[i]]
i = i - len(words[i])
## Return sequence of best words and overall probability
return sequence, best[-1]

Hmm I can post mine. Its not pretty though, please let me know if you need clarification. I wrote this relatively recently for specifically part of speech tagging.
class Trellis:
trell = []
def __init__(self, hmm, words):
self.trell = []
temp = {}
for label in hmm.labels:
temp[label] = [0,None]
for word in words:
self.trell.append([word,copy.deepcopy(temp)])
self.fill_in(hmm)
def fill_in(self,hmm):
for i in range(len(self.trell)):
for token in self.trell[i][1]:
word = self.trell[i][0]
if i == 0:
self.trell[i][1][token][0] = hmm.e(token,word)
else:
max = None
guess = None
c = None
for k in self.trell[i-1][1]:
c = self.trell[i-1][1][k][0] + hmm.t(k,token)
if max == None or c > max:
max = c
guess = k
max += hmm.e(token,word)
self.trell[i][1][token][0] = max
self.trell[i][1][token][1] = guess
def return_max(self):
tokens = []
token = None
for i in range(len(self.trell)-1,-1,-1):
if token == None:
max = None
guess = None
for k in self.trell[i][1]:
if max == None or self.trell[i][1][k][0] > max:
max = self.trell[i][1][k][0]
token = self.trell[i][1][k][1]
guess = k
tokens.append(guess)
else:
tokens.append(token)
token = self.trell[i][1][token][1]
tokens.reverse()
return tokens

I have just corrected the pseudo implementation of Viterbi in Wikipedia. From the initial (incorrect) version, it took me a while to figure out where I was going wrong but I finally managed it, thanks partly to Kevin Murphy's implementation of the viterbi_path.m in the MatLab HMM toolbox.
In the context of an HMM object with variables as shown:
hmm = HMM()
hmm.priors = np.array([0.5, 0.5]) # pi = prior probs
hmm.transition = np.array([[0.75, 0.25], # A = transition probs. / 2 states
[0.32, 0.68]])
hmm.emission = np.array([[0.8, 0.1, 0.1], # B = emission (observation) probs. / 3 obs modes
[0.1, 0.2, 0.7]])
The Python function to run Viterbi (best-path) algorithm is below:
def viterbi (self,observations):
"""Return the best path, given an HMM model and a sequence of observations"""
# A - initialise stuff
nSamples = len(observations[0])
nStates = self.transition.shape[0] # number of states
c = np.zeros(nSamples) #scale factors (necessary to prevent underflow)
viterbi = np.zeros((nStates,nSamples)) # initialise viterbi table
psi = np.zeros((nStates,nSamples)) # initialise the best path table
best_path = np.zeros(nSamples); # this will be your output
# B- appoint initial values for viterbi and best path (bp) tables - Eq (32a-32b)
viterbi[:,0] = self.priors.T * self.emission[:,observations(0)]
c[0] = 1.0/np.sum(viterbi[:,0])
viterbi[:,0] = c[0] * viterbi[:,0] # apply the scaling factor
psi[0] = 0;
# C- Do the iterations for viterbi and psi for time>0 until T
for t in range(1,nSamples): # loop through time
for s in range (0,nStates): # loop through the states #(t-1)
trans_p = viterbi[:,t-1] * self.transition[:,s]
psi[s,t], viterbi[s,t] = max(enumerate(trans_p), key=operator.itemgetter(1))
viterbi[s,t] = viterbi[s,t]*self.emission[s,observations(t)]
c[t] = 1.0/np.sum(viterbi[:,t]) # scaling factor
viterbi[:,t] = c[t] * viterbi[:,t]
# D - Back-tracking
best_path[nSamples-1] = viterbi[:,nSamples-1].argmax() # last state
for t in range(nSamples-1,0,-1): # states of (last-1)th to 0th time step
best_path[t-1] = psi[best_path[t],t]
return best_path

This is an old question, but none of the other answers were quite what I needed because my application doesn't have specific observed states.
Taking after #Rhubarb, I've also re-implemented Kevin Murphey's Matlab implementation (see viterbi_path.m), but I've kept it closer to the original. I've included a simple test case as well.
import numpy as np
def viterbi_path(prior, transmat, obslik, scaled=True, ret_loglik=False):
'''Finds the most-probable (Viterbi) path through the HMM state trellis
Notation:
Z[t] := Observation at time t
Q[t] := Hidden state at time t
Inputs:
prior: np.array(num_hid)
prior[i] := Pr(Q[0] == i)
transmat: np.ndarray((num_hid,num_hid))
transmat[i,j] := Pr(Q[t+1] == j | Q[t] == i)
obslik: np.ndarray((num_hid,num_obs))
obslik[i,t] := Pr(Z[t] | Q[t] == i)
scaled: bool
whether or not to normalize the probability trellis along the way
doing so prevents underflow by repeated multiplications of probabilities
ret_loglik: bool
whether or not to return the log-likelihood of the best path
Outputs:
path: np.array(num_obs)
path[t] := Q[t]
'''
num_hid = obslik.shape[0] # number of hidden states
num_obs = obslik.shape[1] # number of observations (not observation *states*)
# trellis_prob[i,t] := Pr((best sequence of length t-1 goes to state i), Z[1:(t+1)])
trellis_prob = np.zeros((num_hid,num_obs))
# trellis_state[i,t] := best predecessor state given that we ended up in state i at t
trellis_state = np.zeros((num_hid,num_obs), dtype=int) # int because its elements will be used as indicies
path = np.zeros(num_obs, dtype=int) # int because its elements will be used as indicies
trellis_prob[:,0] = prior * obslik[:,0] # element-wise mult
if scaled:
scale = np.ones(num_obs) # only instantiated if necessary to save memory
scale[0] = 1.0 / np.sum(trellis_prob[:,0])
trellis_prob[:,0] *= scale[0]
trellis_state[:,0] = 0 # arbitrary value since t == 0 has no predecessor
for t in xrange(1, num_obs):
for j in xrange(num_hid):
trans_probs = trellis_prob[:,t-1] * transmat[:,j] # element-wise mult
trellis_state[j,t] = trans_probs.argmax()
trellis_prob[j,t] = trans_probs[trellis_state[j,t]] # max of trans_probs
trellis_prob[j,t] *= obslik[j,t]
if scaled:
scale[t] = 1.0 / np.sum(trellis_prob[:,t])
trellis_prob[:,t] *= scale[t]
path[-1] = trellis_prob[:,-1].argmax()
for t in range(num_obs-2, -1, -1):
path[t] = trellis_state[(path[t+1]), t+1]
if not ret_loglik:
return path
else:
if scaled:
loglik = -np.sum(np.log(scale))
else:
p = trellis_prob[path[-1],-1]
loglik = np.log(p)
return path, loglik
if __name__=='__main__':
# Assume there are 3 observation states, 2 hidden states, and 5 observations
priors = np.array([0.5, 0.5])
transmat = np.array([
[0.75, 0.25],
[0.32, 0.68]])
emmat = np.array([
[0.8, 0.1, 0.1],
[0.1, 0.2, 0.7]])
observations = np.array([0, 1, 2, 1, 0], dtype=int)
obslik = np.array([emmat[:,z] for z in observations]).T
print viterbi_path(priors, transmat, obslik) #=> [0 1 1 1 0]
print viterbi_path(priors, transmat, obslik, scaled=False) #=> [0 1 1 1 0]
print viterbi_path(priors, transmat, obslik, ret_loglik=True) #=> (array([0, 1, 1, 1, 0]), -7.776472586614755)
print viterbi_path(priors, transmat, obslik, scaled=False, ret_loglik=True) #=> (array([0, 1, 1, 1, 0]), -8.0120386579275227)
Note that this implementation does not use emission probabilities directly but uses a variable obslik. Generally, emissions[i,j] := Pr(observed_state == j | hidden_state == i) for a particular observed state i, making emissions.shape == (num_hidden_states, num_obs_states).
However, given a sequence observations[t] := observation at time t, all the Viterbi Algorithm requires is the likelihood of that observation for each hidden state. Hence, obslik[i,t] := Pr(observations[t] | hidden_state == i). The actual value the of the observed state isn't necessary.

I have modified #Rhubarb's answer for the condition where the marginal probabilities are already known (e.g by computing the Forward Backward algorithm).
def viterbi (transition_probabilities, conditional_probabilities):
# Initialise everything
num_samples = conditional_probabilities.shape[1]
num_states = transition_probabilities.shape[0] # number of states
c = np.zeros(num_samples) #scale factors (necessary to prevent underflow)
viterbi = np.zeros((num_states,num_samples)) # initialise viterbi table
best_path_table = np.zeros((num_states,num_samples)) # initialise the best path table
best_path = np.zeros(num_samples).astype(np.int32) # this will be your output
# B- appoint initial values for viterbi and best path (bp) tables - Eq (32a-32b)
viterbi[:,0] = conditional_probabilities[:,0]
c[0] = 1.0/np.sum(viterbi[:,0])
viterbi[:,0] = c[0] * viterbi[:,0] # apply the scaling factor
# C- Do the iterations for viterbi and psi for time>0 until T
for t in range(1, num_samples): # loop through time
for s in range (0,num_states): # loop through the states #(t-1)
trans_p = viterbi[:, t-1] * transition_probabilities[:,s] # transition probs of each state transitioning
best_path_table[s,t], viterbi[s,t] = max(enumerate(trans_p), key=operator.itemgetter(1))
viterbi[s,t] = viterbi[s,t] * conditional_probabilities[s][t]
c[t] = 1.0/np.sum(viterbi[:,t]) # scaling factor
viterbi[:,t] = c[t] * viterbi[:,t]
## D - Back-tracking
best_path[num_samples-1] = viterbi[:,num_samples-1].argmax() # last state
for t in range(num_samples-1,0,-1): # states of (last-1)th to 0th time step
best_path[t-1] = best_path_table[best_path[t],t]
return best_path