Related
I'm running functions to create cyclical datetime features, so I have converted timestamps to sine and cosine representations for ML model training.
In one sample, x = 305.2116709309027, giving np.sin(x) = -0.459279 and np.cos(x) = -0.888292, my question is how to retrieve x from these sin and cos features later?
I assumed np.arcsin(-0.459279) == 305.2116709309027 and I could then decode the timestamp used from there but I'm not having any luck.
You should be aware that mathematically, sin(x) and cos(x) are periodic functions, meaning multiple different values as input can yield the same output.
For example, x=0, x=2pi, and x=4pi can all yield the same value. So you can't decode the x from y, except you know that the input is restricted within a period, such as between [0, 2pi].
HOWEVER, for arcsin(x), since the domain of x is limited, and each y corresponds to a unique x, you can get the x from y.
Suppose there are three variables that take on discrete integer values, say w1 = {1,2,3,4,5,6,7,8,9,10,11,12}, w2 = {1,2,3,4,5,6,7,8,9,10,11,12}, and w3 = {1,2,3,4,5,6,7,8,9,10,11,12}. The task is to pick one value from each set such that the resulting triplet minimizes some (black box, computationally expensive) cost function.
I've tried the surrogate optimization in Matlab but I'm not sure it is appropriate. I've also heard about simulated annealing but found no implementation applied to this instance.
Which algorithm, apart from exhaustive search, can solve this combinatorial optimization problem?
Any help would be much appreciated.
The requirement/benefit of Simulated Annealing (SA), is that the objective surface is somewhat smooth, that is, we can be close to a solution.
For a completely random spiky surface- you might as well do a random search
If it is anything smooth, or even sometimes, it makes sense to try SA.
The idea is that (sometimes) changing only 1 of the 3 values, we have little effect on out blackbox function.
Here is a basic example to do this with Simulated Annealing, using frigidum in Python
import numpy as np
w1 = np.array( [1,2,3,4,5,6,7,8,9,10,11,12] )
w2 = np.array( [1,2,3,4,5,6,7,8,9,10,11,12] )
w3 = np.array( [1,2,3,4,5,6,7,8,9,10,11,12] )
W = np.array([w1,w2,w3])
LENGTH = 12
I define a black-box using the Rastrigin function.
def rastrigin_function_n( x ):
"""
N-dimensional Rastrigin
https://en.wikipedia.org/wiki/Rastrigin_function
x_i is in [-5.12, 5.12]
"""
A = 10
n = x.shape[0]
return A*n + np.sum( x**2- A*np.cos(2*np.pi * x) )
def black_box( x ):
"""
Transform from domain [1,12] to [-5,5]
to be able to push to rastrigin
"""
x = (x - 6.5) * (5/5.5)
return rastrigin_function_n(x)
Simulated Annealing needs to modify state X. Instead of taking/modifying values directly, we keep track of indices. This simplifies creating new proposals as an index is always an integer we can simply add/subtract 1 modulo LENGTH.
def random_start():
"""
returns 3 random indices
"""
return np.random.randint(0, LENGTH, size=3)
def random_small_step(x):
"""
change only 1 index
"""
d = np.array( [1,0,0] )
if np.random.random() < .5:
d = np.array( [-1,0,0] )
np.random.shuffle(d)
return (x+d) % LENGTH
def random_big_step(x):
"""
change 2 indici
"""
d = np.array( [1,-1,0] )
np.random.shuffle(d)
return (x+d) % LENGTH
def obj(x):
"""
We have a triplet of indici,
1. Calculate corresponding values in W = [w1,w2,w3]
2. Push the values in out black-box function
"""
indices = x
values = W[np.array([0,1,2]), indices]
return black_box(values)
And throw a SA Scheme at it
import frigidum
local_opt = frigidum.sa(random_start=random_start,
neighbours=[random_small_step, random_big_step],
objective_function=obj,
T_start=10**4,
T_stop=0.000001,
repeats=10**3,
copy_state=frigidum.annealing.naked)
I am not sure what the minimum for this function should be, but it found a objective with 47.9095 with indicis np.array([9, 2, 2])
Edit:
For frigidum to change the cooling schedule, use alpha=.9. My experience is that all the work of experiment which cooling scheme works best doesn't out-weight simply let it run a little longer. The multiplication you proposed, (sometimes called geometric) is the standard one, also implemented in frigidum. So to implement Tn+1 = 0.9*Tn you need a alpha=.9. Be aware this cooling step is done after N repeats, so if repeats=100, it will first do 100 proposals before lowering the temperature with factor alpha
Simple variations on current state often works best. Since its best practice to set the initial temperature high enough to make most proposals (>90%) accepted, it doesn't matter the steps are small. But if you fear its soo small, try 2 or 3 variations. Frigidum accepts a list of proposal functions, and combinations can enforce each other.
I have no experience with MINLP. But even if, so many times experiments can surprise us. So if time/cost is small to bring another competitor to the table, yes!
Try every possible combination of the three values and see which has the lowest cost.
In order to make the case simple and intuitive, I will using binary (0 and 1) classification for illustration.
Loss function
loss = np.multiply(np.log(predY), Y) + np.multiply((1 - Y), np.log(1 - predY)) #cross entropy
cost = -np.sum(loss)/m #num of examples in batch is m
Probability of Y
predY is computed using sigmoid and logits can be thought as the outcome of from a neural network before reaching the classification step
predY = sigmoid(logits) #binary case
def sigmoid(X):
return 1/(1 + np.exp(-X))
Problem
Suppose we are running a feed-forward net.
Inputs: [3, 5]: 3 is number of examples and 5 is feature size (fabricated data)
Num of hidden units: 100 (only 1 hidden layer)
Iterations: 10000
Such arrangement is set to overfit. When it's overfitting, we can perfectly predict the probability for the training examples; in other words, sigmoid outputs either 1 or 0, exact number because the exponential gets exploded. If this is the case, we would have np.log(0) undefined. How do you usually handle this issue?
If you don't mind the dependency on scipy, you can use scipy.special.xlogy. You would replace the expression
np.multiply(np.log(predY), Y) + np.multiply((1 - Y), np.log(1 - predY))
with
xlogy(Y, predY) + xlogy(1 - Y, 1 - predY)
If you expect predY to contain very small values, you might get better numerical results using scipy.special.xlog1py in the second term:
xlogy(Y, predY) + xlog1py(1 - Y, -predY)
Alternatively, knowing that the values in Y are either 0 or 1, you can compute the cost in an entirely different way:
Yis1 = Y == 1
cost = -(np.log(predY[Yis1]).sum() + np.log(1 - predY[~Yis1]).sum())/m
How do you usually handle this issue?
Add small number (something like 1e-15) to predY - this number doesn't make predictions much off, and it solves log(0) issue.
BTW if your algorithm outputs zeros and ones it might be useful to check the histogram of returned probabilities - when algorithm is so sure that something's happening it can be a sign of overfitting.
One common way to deal with log(x) and y / x where x is always non-negative but can become 0 is to add a small constant (as written by Jakub).
You can also clip the value (e.g. tf.clip_by_value or np.clip).
I've created a codebook using k-means of size 4000x300 (4000 centroids, each with 300 features). Using the codebook, I then want to label an input vector (for purposes of binning later on). The input vector is of size Nx300, where N is the total number of input instances I receive.
To compute the labels, I calculate the closest centroid for each of the input vectors. To do so, I compare each input vector against all centroids and pick the centroid with the minimum distance. The label is then just the index of that centroid.
My current Matlab code looks like:
function labels = assign_labels(centroids, X)
labels = zeros(size(X, 1), 1);
% for each X, calculate the distance from each centroid
for i = 1:size(X, 1)
% distance of X_i from all j centroids is: sum((X_i - centroid_j)^2)
% note: we leave off the sqrt as an optimization
distances = sum(bsxfun(#minus, centroids, X(i, :)) .^ 2, 2);
[value, label] = min(distances);
labels(i) = label;
end
However, this code is still fairly slow (for my purposes), and I was hoping there might be a way to optimize the code further.
One obvious issue is that there is a for-loop, which is the bane of good performance on Matlab. I've been trying to come up with a way to get rid of it, but with no luck (I looked into using arrayfun in conjunction with bsxfun, but haven't gotten that to work). Alternatively, if someone know of any other way to speed this up, I would be greatly appreciate it.
Update
After doing some searching, I couldn't find a great solution using Matlab, so I decided to look at what is used in Python's scikits.learn package for 'euclidean_distance' (shortened):
XX = sum(X * X, axis=1)[:, newaxis]
YY = Y.copy()
YY **= 2
YY = sum(YY, axis=1)[newaxis, :]
distances = XX + YY
distances -= 2 * dot(X, Y.T)
distances = maximum(distances, 0)
which uses the binomial form of the euclidean distance ((x-y)^2 -> x^2 + y^2 - 2xy), which from what I've read usually runs faster. My completely untested Matlab translation is:
XX = sum(data .* data, 2);
YY = sum(center .^ 2, 2);
[val, ~] = max(XX + YY - 2*data*center');
Use the following function to calculate your distances. You should see an order of magnitude speed up
The two matrices A and B have the columns as the dimenions and the rows as each point.
A is your matrix of centroids. B is your matrix of datapoints.
function D=getSim(A,B)
Qa=repmat(dot(A,A,2),1,size(B,1));
Qb=repmat(dot(B,B,2),1,size(A,1));
D=Qa+Qb'-2*A*B';
You can vectorize it by converting to cells and using cellfun:
[nRows,nCols]=size(X);
XCell=num2cell(X,2);
dist=reshape(cell2mat(cellfun(#(x)(sum(bsxfun(#minus,centroids,x).^2,2)),XCell,'UniformOutput',false)),nRows,nRows);
[~,labels]=min(dist);
Explanation:
We assign each row of X to its own cell in the second line
This piece #(x)(sum(bsxfun(#minus,centroids,x).^2,2)) is an anonymous function which is the same as your distances=... line, and using cell2mat, we apply it to each row of X.
The labels are then the indices of the minimum row along each column.
For a true matrix implementation, you may consider trying something along the lines of:
P2 = kron(centroids, ones(size(X,1),1));
Q2 = kron(ones(size(centroids,1),1), X);
distances = reshape(sum((Q2-P2).^2,2), size(X,1), size(centroids,1));
Note
This assumes the data is organized as [x1 y1 ...; x2 y2 ...;...]
You can use a more efficient algorithm for nearest neighbor search than brute force.
The most popular approach are Kd-Tree. O(log(n)) average query time instead of the O(n) brute force complexity.
Regarding a Maltab implementation of Kd-Trees, you can have a look here
I am processing a series of points which all have the same Y value, but different X values. I go through the points by incrementing X by one. For example, I might have Y = 50 and X is the integers from -30 to 30. Part of my algorithm involves finding the distance to the origin from each point and then doing further processing.
After profiling, I've found that the sqrt call in the distance calculation is taking a significant amount of my time. Is there an iterative way to calculate the distance?
In other words:
I want to efficiently calculate: r[n] = sqrt(x[n]*x[n] + y*y)). I can save information from the previous iteration. Each iteration changes by incrementing x, so x[n] = x[n-1] + 1. I can not use sqrt or trig functions because they are too slow except at the beginning of each scanline.
I can use approximations as long as they are good enough (less than 0.l% error) and the errors introduced are smooth (I can't bin to a pre-calculated table of approximations).
Additional information:
x and y are always integers between -150 and 150
I'm going to try a couple ideas out tomorrow and mark the best answer based on which is fastest.
Results
I did some timings
Distance formula: 16 ms / iteration
Pete's interperlating solution: 8 ms / iteration
wrang-wrang pre-calculation solution: 8ms / iteration
I was hoping the test would decide between the two, because I like both answers. I'm going to go with Pete's because it uses less memory.
Just to get a feel for it, for your range y = 50, x = 0 gives r = 50 and y = 50, x = +/- 30 gives r ~= 58.3. You want an approximation good for +/- 0.1%, or +/- 0.05 absolute. That's a lot lower accuracy than most library sqrts do.
Two approximate approaches - you calculate r based on interpolating from the previous value, or use a few terms of a suitable series.
Interpolating from previous r
r = ( x2 + y2 ) 1/2
dr/dx = 1/2 . 2x . ( x2 + y2 ) -1/2 = x/r
double r = 50;
for ( int x = 0; x <= 30; ++x ) {
double r_true = Math.sqrt ( 50*50 + x*x );
System.out.printf ( "x: %d r_true: %f r_approx: %f error: %f%%\n", x, r, r_true, 100 * Math.abs ( r_true - r ) / r );
r = r + ( x + 0.5 ) / r;
}
Gives:
x: 0 r_true: 50.000000 r_approx: 50.000000 error: 0.000000%
x: 1 r_true: 50.010000 r_approx: 50.009999 error: 0.000002%
....
x: 29 r_true: 57.825065 r_approx: 57.801384 error: 0.040953%
x: 30 r_true: 58.335225 r_approx: 58.309519 error: 0.044065%
which seems to meet the requirement of 0.1% error, so I didn't bother coding the next one, as it would require quite a bit more calculation steps.
Truncated Series
The taylor series for sqrt ( 1 + x ) for x near zero is
sqrt ( 1 + x ) = 1 + 1/2 x - 1/8 x2 ... + ( - 1 / 2 )n+1 xn
Using r = y sqrt ( 1 + (x/y)2 ) then you're looking for a term t = ( - 1 / 2 )n+1 0.36n with magnitude less that a 0.001, log ( 0.002 ) > n log ( 0.18 ) or n > 3.6, so taking terms to x^4 should be Ok.
Y=10000
Y2=Y*Y
for x=0..Y2 do
D[x]=sqrt(Y2+x*x)
norm(x,y)=
if (y==0) x
else if (x>y) norm(y,x)
else {
s=Y/y
D[round(x*s)]/s
}
If your coordinates are smooth, then the idea can be extended with linear interpolation. For more precision, increase Y.
The idea is that s*(x,y) is on the line y=Y, which you've precomputed distances for. Get the distance, then divide it by s.
I assume you really do need the distance and not its square.
You may also be able to find a general sqrt implementation that sacrifices some accuracy for speed, but I have a hard time imagining that beating what the FPU can do.
By linear interpolation, I mean to change D[round(x)] to:
f=floor(x)
a=x-f
D[f]*(1-a)+D[f+1]*a
This doesn't really answer your question, but may help...
The first questions I would ask would be:
"do I need the sqrt at all?".
"If not, how can I reduce the number of sqrts?"
then yours: "Can I replace the remaining sqrts with a clever calculation?"
So I'd start with:
Do you need the exact radius, or would radius-squared be acceptable? There are fast approximatiosn to sqrt, but probably not accurate enough for your spec.
Can you process the image using mirrored quadrants or eighths? By processing all pixels at the same radius value in a batch, you can reduce the number of calculations by 8x.
Can you precalculate the radius values? You only need a table that is a quarter (or possibly an eighth) of the size of the image you are processing, and the table would only need to be precalculated once and then re-used for many runs of the algorithm.
So clever maths may not be the fastest solution.
Well there's always trying optimize your sqrt, the fastest one I've seen is the old carmack quake 3 sqrt:
http://betterexplained.com/articles/understanding-quakes-fast-inverse-square-root/
That said, since sqrt is non-linear, you're not going to be able to do simple linear interpolation along your line to get your result. The best idea is to use a table lookup since that will give you blazing fast access to the data. And, since you appear to be iterating by whole integers, a table lookup should be exceedingly accurate.
Well, you can mirror around x=0 to start with (you need only compute n>=0, and the dupe those results to corresponding n<0). After that, I'd take a look at using the derivative on sqrt(a^2+b^2) (or the corresponding sin) to take advantage of the constant dx.
If that's not accurate enough, may I point out that this is a pretty good job for SIMD, which will provide you with a reciprocal square root op on both SSE and VMX (and shader model 2).
This is sort of related to a HAKMEM item:
ITEM 149 (Minsky): CIRCLE ALGORITHM
Here is an elegant way to draw almost
circles on a point-plotting display:
NEW X = OLD X - epsilon * OLD Y
NEW Y = OLD Y + epsilon * NEW(!) X
This makes a very round ellipse
centered at the origin with its size
determined by the initial point.
epsilon determines the angular
velocity of the circulating point, and
slightly affects the eccentricity. If
epsilon is a power of 2, then we don't
even need multiplication, let alone
square roots, sines, and cosines! The
"circle" will be perfectly stable
because the points soon become
periodic.
The circle algorithm was invented by
mistake when I tried to save one
register in a display hack! Ben Gurley
had an amazing display hack using only
about six or seven instructions, and
it was a great wonder. But it was
basically line-oriented. It occurred
to me that it would be exciting to
have curves, and I was trying to get a
curve display hack with minimal
instructions.