Develop Reference

How to add noise to gradient descent?

I have the following code where I have implemented gradient descent for a function using pyTorch. How do I add noise to the code so that it identifies both local minima?
import torch
startVal = -5.0
alpha = 0.001
space = " "
progressionCheck = True
x = torch.tensor(startVal, requires_grad=True)
def function(a):
f = a**4 - a**3 - a**2 + a - 1
return f
for i in range(1000):
function(x).backward()
newVal = x - alpha * (x.grad)
progressionCheck = function(newVal) < function(startVal)
x = newVal.detach().clone().requires_grad_()
print(x)
print("The minimum value occurs at" + space + str(float(x)))
print("The minimum value is" + space + str(function(float(x))))

I assume you intend to disturb the gradients by some noise. To do so, you could specify a distribution e.g. as follows
low, high = -0.1, 0.1
dist = torch.distributions.uniform.Uniform(low, high)
and then sample from it to update the gradients, i.e. adjust
newVal = x - alpha * (x.grad)
to
newVal = x - alpha * (x.grad) * dist.sample([1]).item()
Altneratively, sample the noise in advance
noise = dist.sample([1000])
and then index it
newVal = x - alpha * (x.grad) * noise[i]
However, I doubt this fulfils the purpose and don't see how you could avoid multiple runs coupled with using varying start values (or, less beautiful, very large noise or step size) to find multiple local minima.

Is there a fast running rolling standard deviation algorithm?

I have a Python script in which for every new sample i have to update the standard deviation of this samples array using a rolling window of length N. Using the simple formula of the standard deviation the code is really slow. I found many different solutions for online calculation but all of them are NOT considering a rolling window for the update. Some of alternative way for computing the variance are explained here (Welford algorithm, parallel, ...)
https://en.m.wikipedia.org/wiki/Algorithms_for_calculating_variance
but none of them are actually using a rolling window through data set.
What i'm looking for is a fast algorithm which won't be prone to catastrophic cancellation phenomenon.
Formulas will be appreciated.
Thanks for your help guys.

Here's an adaptation of the code at the link I put in a comment. It takes O(1) (constant) time for each element added, regardless of window size. Note that if you configure for a window of N elements, the first N-1 results are more-than-less gibberish: it initializes the data to N zeroes. The class also saves the most recent N entries in a collections.deque of maximum size N. Note that this computes the "sample" standard deviation, not "population". Season to taste ;-)
from collections import deque
from math import sqrt
class Running:
def __init__(self, n):
self.n = n
self.data = deque([0.0] * n, maxlen=n)
self.mean = self.variance = self.sdev = 0.0
def add(self, x):
n = self.n
oldmean = self.mean
goingaway = self.data[0]
self.mean = newmean = oldmean + (x - goingaway) / n
self.data.append(x)
self.variance += (x - goingaway) * (
(x - newmean) + (goingaway - oldmean)) / (n - 1)
self.sdev = sqrt(self.variance)
Here's a by-eyeball sanity check. Seems fine. But note that the statistics module makes heroic (and slow!) efforts to maximize floating-point accuracy. The code above just accepts accumulating half a dozen fresh rounding errors per element added.
import statistics
from random import random
from math import ulp
r = Running(50)
for i in range(1000000):
r.add(random() * 100)
assert len(r.data) == 50
if i % 1000 == 0.0:
a, b = r.mean, statistics.mean(r.data)
print(i, "mean", a, b, (a - b) / ulp(b))
a, b = r.sdev, statistics.stdev(r.data)
print(i, "sdev", a, b, (a - b) / ulp(b))
Sample output (will vary across runs):
0 mean 1.4656985567210468 1.4656985567210468 0.0
0 sdev 10.364053886327875 10.364053886327877 -1.0
1000 mean 50.73313401192864 50.73313401192878 -20.0
1000 sdev 31.06576415649153 31.06576415649151 5.0
2000 mean 50.4175663202043 50.41756632020437 -10.0
2000 sdev 27.692406266774878 27.69240626677488 -1.0
3000 mean 53.054435599235525 53.0544355992356 -11.0
3000 sdev 32.439246859431556 32.439246859431606 -7.0
4000 mean 51.66216784517698 51.662167845177 -3.0
4000 sdev 31.026902004950404 31.02690200495047 -18.0
5000 mean 54.08949367166644 54.089493671666425 2.0
5000 sdev 29.405357061221196 29.40535706122128 -24.0
...

I have a function that I use for standard deviation. I modified it from a php function that calculated standard deviation. You can input an array (or a slice of an array) into it, and it will calculate the standard deviation for that array or slice.
def calc_std_dev(lst, precision=0, sample=True):
"""
:param: lst A Python list containing values
:param: precision The number of decimal places desired.
:param: sample Is the data a sample or a population? Set to True by
default.
"""
sum = 0;
length = len(lst)
if length == 0:
print("The array has zero elements.")
return false
elif (length == 1) and (sample == True):
print("The array has only 1 element.")
return false
else:
sum = math.fsum(lst)
# Calculate the arithmetic mean
mean = sum / length
carry = 0.0
for i in lst:
dev = i - mean
carry += dev * dev
if sample == True:
length = length - 1
variance = carry / length
std_dev = math.sqrt(variance)
std_dev = round(std_dev, precision)
return std_dev
When I need a rolling standard deviation, I pass in a slice of the total list to calculate the value.

Calculating inverse trigonometric functions with formulas

I have been trying to create custom calculator for calculating trigonometric functions. Aside from Chebyshev pylonomials and/or Cordic algorithm I have used Taylor series which have been accurate by few places of decimal.
This is what i have created to calculate simple trigonometric functions without any modules:
from __future__ import division
def sqrt(n):
ans = n ** 0.5
return ans
def factorial(n):
k = 1
for i in range(1, n+1):
k = i * k
return k
def sin(d):
pi = 3.14159265359
n = 180 / int(d) # 180 degrees = pi radians
x = pi / n # Converting degrees to radians
ans = x - ( x ** 3 / factorial(3) ) + ( x ** 5 / factorial(5) ) - ( x ** 7 / factorial(7) ) + ( x ** 9 / factorial(9) )
return ans
def cos(d):
pi = 3.14159265359
n = 180 / int(d)
x = pi / n
ans = 1 - ( x ** 2 / factorial(2) ) + ( x ** 4 / factorial(4) ) - ( x ** 6 / factorial(6) ) + ( x ** 8 / factorial(8) )
return ans
def tan(d):
ans = sin(d) / sqrt(1 - sin(d) ** 2)
return ans
Unfortunately i could not find any sources that would help me interpret inverse trigonometric function formulas for Python. I have also tried putting sin(x) to the power of -1 (sin(x) ** -1) which didn't work as expected.
What could be the best solution to do this in Python (In the best, I mean simplest with similar accuracy as Taylor series)? Is this possible with power series or do i need to use cordic algorithm?

The question is broad in scope, but here are some simple ideas (and code!) that might serve as a starting point for computing arctan. First, the good old Taylor series. For simplicity, we use a fixed number of terms; in practice, you might want to decide the number of terms to use dynamically based on the size of x, or introduce some kind of convergence criterion. With a fixed number of terms, we can evaluate efficiently using something akin to Horner's scheme.
def arctan_taylor(x, terms=9):
"""
Compute arctan for small x via Taylor polynomials.
Uses a fixed number of terms. The default of 9 should give good results for
abs(x) < 0.1. Results will become poorer as abs(x) increases, becoming
unusable as abs(x) approaches 1.0 (the radius of convergence of the
series).
"""
# Uses Horner's method for evaluation.
t = 0.0
for n in range(2*terms-1, 0, -2):
t = 1.0/n - x*x*t
return x * t
The above code gives good results for small x (say smaller than 0.1 in absolute value), but the accuracy drops off as x becomes larger, and for abs(x) > 1.0, the series never converges, no matter how many terms (or how much extra precision) we throw at it. So we need a better way to compute for larger x. One solution is to use argument reduction, via the identity arctan(x) = 2 * arctan(x / (1 + sqrt(1 + x^2))). This gives the following code, which builds on arctan_taylor to give reasonable results for a wide range of x (but beware possible overflow and underflow when computing x*x).
import math
def arctan_taylor_with_reduction(x, terms=9, threshold=0.1):
"""
Compute arctan via argument reduction and Taylor series.
Applies reduction steps until x is below `threshold`,
then uses Taylor series.
"""
reductions = 0
while abs(x) > threshold:
x = x / (1 + math.sqrt(1 + x*x))
reductions += 1
return arctan_taylor(x, terms=terms) * 2**reductions
Alternatively, given an existing implementation for tan, you could simply find a solution y to the equation tan(y) = x using traditional root-finding methods. Since arctan is already naturally bounded to lie in the interval (-pi/2, pi/2), bisection search works well:
def arctan_from_tan(x, tolerance=1e-15):
"""
Compute arctan as the inverse of tan, via bisection search. This assumes
that you already have a high quality tan function.
"""
low, high = -0.5 * math.pi, 0.5 * math.pi
while high - low > tolerance:
mid = 0.5 * (low + high)
if math.tan(mid) < x:
low = mid
else:
high = mid
return 0.5 * (low + high)
Finally, just for fun, here's a CORDIC-like implementation, which is really more appropriate for a low-level implementation than for Python. The idea here is that you precompute, once and for all, a table of arctan values for 1, 1/2, 1/4, etc., and then use those to compute general arctan values, essentially by computing successive approximations to the true angle. The remarkable part is that, after the precomputation step, the arctan computation involves only additions, subtractions, and multiplications by by powers of 2. (Of course, those multiplications aren't any more efficient than any other multiplication at the level of Python, but closer to the hardware, this could potentially make a big difference.)
cordic_table_size = 60
cordic_table = [(2**-i, math.atan(2**-i))
for i in range(cordic_table_size)]
def arctan_cordic(y, x=1.0):
"""
Compute arctan(y/x), assuming x positive, via CORDIC-like method.
"""
r = 0.0
for t, a in cordic_table:
if y < 0:
r, x, y = r - a, x - t*y, y + t*x
else:
r, x, y = r + a, x + t*y, y - t*x
return r
Each of the above methods has its strengths and weaknesses, and all of the above code can be improved in a myriad of ways. I encourage you to experiment and explore.
To wrap it all up, here are the results of calling the above functions on a small number of not-very-carefully-chosen test values, comparing with the output of the standard library math.atan function:
test_values = [2.314, 0.0123, -0.56, 168.9]
for value in test_values:
print("{:20.15g} {:20.15g} {:20.15g} {:20.15g}".format(
math.atan(value),
arctan_taylor_with_reduction(value),
arctan_from_tan(value),
arctan_cordic(value),
))
Output on my machine:
1.16288340166519 1.16288340166519 1.16288340166519 1.16288340166519
0.0122993797673 0.0122993797673 0.0122993797673002 0.0122993797672999
-0.510488321916776 -0.510488321916776 -0.510488321916776 -0.510488321916776
1.56487573286064 1.56487573286064 1.56487573286064 1.56487573286064

The simplest way to do any inverse function is to use binary search.
definitions
let assume function
x = g(y)
And we want to code its inverse:
y = f(x) = f(g(y))
x = <x0,x1>
y = <y0,y1>
bin search on floats
You can do it on integer math accessing mantissa bits like in here:
Any Faster RMS Value Calculation in C?
but if you do not know the exponent of the result prior to computation then you need to use floats for bin search too.
so the idea behind binary search is to change mantissa of y from y1 to y0 bit by bit from MSB to LSB. Then call direct function g(y) and if the result cross x revert the last bit change.
In case of using floats you can use variable that will hold approximate value of the mantissa bit targeted instead of integer bit access. That will eliminate unknown exponent problem. So at the beginning set y = y0 and actual bit to MSB value so b=(y1-y0)/2. After each iteration halve it and do as many iterations as you got mantissa bits n... This way you obtain result in n iterations within (y1-y0)/2^n accuracy.
If your inverse function is not monotonic break it into monotonic intervals and handle each as separate binary search.
The function increasing/decreasing just determine the crossing condition direction (use of < or >).
C++ acos example
so y = acos(x) is defined on x = <-1,+1> , y = <0,M_PI> and decreasing so:
double f64_acos(double x)
{
const int n=52; // mantisa bits
double y,y0,b;
int i;
// handle domain error
if (x<-1.0) return 0;
if (x>+1.0) return 0;
// x = <-1,+1> , y = <0,M_PI> , decreasing
for (y= 0.0,b=0.5*M_PI,i=0;i<n;i++,b*=0.5) // y is min, b is half of max and halving each iteration
{
y0=y; // remember original y
y+=b; // try set "bit"
if (cos(y)<x) y=y0; // if result cross x return to original y decreasing is < and increasing is >
}
return y;
}
I tested it like this:
double x0,x1,y;
for (x0=0.0;x0<M_PI;x0+=M_PI*0.01) // cycle all angle range <0,M_PI>
{
y=cos(x0); // direct function (from math.h)
x1=f64_acos(y); // my inverse function
if (fabs(x1-x0)>1e-9) // check result and output to log if error
Form1->mm_log->Lines->Add(AnsiString().sprintf("acos(%8.3lf) = %8.3lf != %8.3lf",y,x0,x1));
}
Without any difference found... so the implementation is working correctly. Of coarse binary search on 52 bit mantissa is usually slower then polynomial approximation ... on the other hand the implementation is so simple ...
[Notes]
If you do not want to take care of the monotonic intervals you can try
approximation search
As you are dealing with goniometric functions you need to handle singularities to avoid NaN or division by zero etc ...
If you're interested here more bin search examples (mostly on integers)
Power by squaring for negative exponents it contains

Why does simple gradient descent diverge?

This is my second attempt at implementing gradient descent in one variable and it always diverges. Any ideas?
This is simple linear regression for minimizing the residual sum of squares in one variable.
def gradient_descent_wtf(xvalues, yvalues):
tolerance = 0.1
#y=mx+b
#some line to predict y values from x values
m=1.
b=1.
#a predicted y-value has value mx + b
for i in range(0,10):
#calculate y-value predictions for all x-values
predicted_yvalues = list()
for x in xvalues:
predicted_yvalues.append(m*x + b)
# predicted_yvalues holds the predicted y-values
#now calculate the residuals = y-value - predicted y-value for each point
residuals = list()
number_of_points = len(yvalues)
for n in range(0,number_of_points):
residuals.append(yvalues[n] - predicted_yvalues[n])
## calculate the residual sum of squares from the residuals, that is,
## square each residual and add them all up. we will try to minimize
## the residual sum of squares later.
residual_sum_of_squares = 0.
for r in residuals:
residual_sum_of_squares += r**2
print("RSS = %s" % residual_sum_of_squares)
##
##
##
#now make a version of the residuals which is multiplied by the x-values
residuals_times_xvalues = list()
for n in range(0,number_of_points):
residuals_times_xvalues.append(residuals[n] * xvalues[n])
#now create the sums for the residuals and for the residuals times the x-values
residuals_sum = sum(residuals)
residuals_times_xvalues_sum = sum(residuals_times_xvalues)
# now multiply the sums by a positive scalar and add each to m and b.
residuals_sum *= 0.1
residuals_times_xvalues_sum *= 0.1
b += residuals_sum
m += residuals_times_xvalues_sum
#and repeat until convergence.
#convergence occurs when ||sum vector|| < some tolerance.
# ||sum vector|| = sqrt( residuals_sum**2 + residuals_times_xvalues_sum**2 )
#check for convergence
magnitude_of_sum_vector = (residuals_sum**2 + residuals_times_xvalues_sum**2)**0.5
if magnitude_of_sum_vector < tolerance:
break
return (b, m)
Result:
gradient_descent_wtf([1,2,3,4,5,6,7,8,9,10],[6,23,8,56,3,24,234,76,59,567])
RSS = 370433.0
RSS = 300170125.7
RSS = 4.86943013045e+11
RSS = 7.90447409339e+14
RSS = 1.28312217794e+18
RSS = 2.08287421094e+21
RSS = 3.38110045417e+24
RSS = 5.48849288217e+27
RSS = 8.90939341376e+30
RSS = 1.44624932026e+34
Out[108]:
(-3.475524066284303e+16, -2.4195981188763203e+17)

The gradients are huge -- hence you are following large vectors for long distances (0.1 times a large number is large). Find unit vectors in the appropriate direction. Something like this (with comprehensions replacing your loops):
def gradient_descent_wtf(xvalues, yvalues):
tolerance = 0.1
m=1.
b=1.
for i in range(0,10):
predicted_yvalues = [m*x+b for x in xvalues]
residuals = [y-y_hat for y,y_hat in zip(yvalues,predicted_yvalues)]
residual_sum_of_squares = sum(r**2 for r in residuals) #only needed for debugging purposes
print("RSS = %s" % residual_sum_of_squares)
residuals_times_xvalues = [r*x for r,x in zip(residuals,xvalues)]
residuals_sum = sum(residuals)
residuals_times_xvalues_sum = sum(residuals_times_xvalues)
# (residuals_sum,residual_times_xvalues_sum) is a vector which points in the negative
# gradient direction. *Find a unit vector which points in same direction*
magnitude = (residuals_sum**2 + residuals_times_xvalues_sum**2)**0.5
residuals_sum /= magnitude
residuals_times_xvalues_sum /= magnitude
b += residuals_sum * (0.1)
m += residuals_times_xvalues_sum * (0.1)
#check for convergence -- this needs work!
magnitude_of_sum_vector = (residuals_sum**2 + residuals_times_xvalues_sum**2)**0.5
if magnitude_of_sum_vector < tolerance:
break
return (b, m)
For example:
>>> gradient_descent_wtf([1,2,3,4,5,6,7,8,9,10],[6,23,8,56,3,24,234,76,59,567])
RSS = 370433.0
RSS = 368732.1655050716
RSS = 367039.18363896786
RSS = 365354.0543519137
RSS = 363676.7775934381
RSS = 362007.3533123621
RSS = 360345.7814567845
RSS = 358692.061974069
RSS = 357046.1948108295
RSS = 355408.17991291644
(1.1157111313023558, 1.9932828425473605)
which is certainly much more plausible.
It isn't a trivial matter to make a numerically stable gradient-descent algorithm. You might want to consult a decent textbook in numerical analysis.

First, Your code is right.
But you should consider something about math when you do linear regression.
For example, the residual is -205.8 and your learning rate is 0.1 so you will get a huge descent step -25.8.
It's a so large step that you can't go back to the correct m and b. You have to make your step small enough.
There are two ways to make gradient descent step reasonable:
initialize a small learning rate, such as 0.001 and 0.0003.
Divide your step by the total amount of your input values.

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: