Perform index-wise matrix operation in julia - matrix

I want to perform index-wise operation on a matrix. I know that you can write a regular function and perform it on each entry of the matrix e.g.
function foo(x::Int64)
return x * 2
end
myArray = [1 2 3; 4 5 6]
foo.(myArray)
how would I go about doing something like x * x.elementCol + x.elementrow? essentially the following code in parallel:
function goo(x::Array{Int64,2})
for j = 1:size(x,2)
for i = 1:size(x,1)
x[i,j] = (x[i,j] * j) + i
end
end
return x
end


You can write:
x .= x .* indices(x, 2)' .+ indices(x, 1)

Related

Tricks to improve the performance of a cunstom function in Julia

I am replicating using Julia a sequence of steps originally made in Matlab. In Octave, this procedure takes 1.4582 seconds and in Julia (using Jupyter) it takes approximately 10 seconds. I'll try to be brief in the scripts. My goal is to achieve or improve Octave's performance. First of all, I will describe my variables and some function:
zgrid (double 1x7 size)
kgrid (double 500x1 size)
V0 (double 500x7 size)
P (double 7x7 size) a transition matrix
delta and beta are fixed parameters.
F(z,k) and u(c) are particular functions and are specified in the Julia script.
% Octave script
% V0 is given
[K, Z, K2] = meshgrid(kgrid, zgrid, kgrid);
K = permute(K, [2, 1, 3]);
Z = permute(Z, [2, 1, 3]);
K2 = permute(K2, [2, 1, 3]);
C = max(f(Z,K) + (1-delta)*K - K2,0);
U = u(C);
EV = V0*P';% EV is a 500x7 matrix size
EV = permute(repmat(EV, 1, 1, 500), [3, 2, 1]);
H = U + beta*EV;
[TV, index] = max(H, [], 3);
In Julia, I created a function that replicates this procedure. I used loops, but it has a performance 9 times longer.
% Julia script
% V0 is the input of my T operator function
V0 = repeat(sqrt.(kgrid), outer = [1,7]);
F = (z,k) -> exp(z)*(k^α);
u = (c) -> (c^(1-μ) - 1)/(1-μ)
% parameters
α = 1/3
β = 0.987
δ = 0.012;
μ = 2
Kss = 48.1905148382166
kgrid = range(0.75*Kss, stop=1.25*Kss, length=500);
zgrid = [-0.06725382459813659, -0.044835883065424395, -0.0224179415327122, 0 , 0.022417941532712187, 0.04483588306542438, 0.06725382459813657]
function T(V)
E=V*P'
T1 = zeros(Float64, 500, 7 )
aux = zeros(Float64, 500)
for i = 1:7
for j = 1:500
for l = 1:500
c= maximum( (F(zrid[i],kgrid[j]) +(1-δ)*kgrid[j] - kgrid[l],0))
aux[l] = u(c) + β*E[l,i]
end
T1[j,i] = maximum(aux)
end
end
return T1
end
I would very much like to improve my performance in Julia. I believe there is a way to improve, but I am new in Julia programming.

This code runs for me in 5ms. Note that I have made F and u into proper (not anonymous) functions, F_ and u_, but you could get a similar effect by making the anonymous functions const.
Your main problem is that you have a lot of non-const global variables, and also that your main function is doing unnecessary work multiple times, and creating an unnecessary array, aux.
The performance tips section in the manual is essential reading: https://docs.julialang.org/en/v1/manual/performance-tips/
F_(z,k) = exp(z) * (k^(1/3)); # you can still use α, but it must be const
u_(c) = (c^(1-2) - 1)/(1-2)
function T_(V, P, kgrid, zgrid, β, δ)
E = V * P'
T1 = similar(V)
for i in axes(T1, 2)
for j in axes(T1, 1)
temp = F_(zgrid[i], kgrid[j]) + (1-δ)*kgrid[j]
aux = -Inf
for l in eachindex(kgrid)
c = max(0.0, temp - kgrid[l])
aux = max(aux, u_(c) + β * E[l, i])
end
T1[j,i] = aux
end
end
return T1
end
Benchmark:
V0 = repeat(sqrt.(kgrid), outer = [1,7]);
zgrid = sort!(rand(1, 7); dims=2)
kgrid = sort!(rand(500, 1); dims=1)
P = rand(length(zgrid), length(zgrid))
#btime T_($V0, $P, $kgrid, $zgrid, $β, $δ);
# output: 5.126 ms (4 allocations: 54.91 KiB)

The following should perform much better. The most noticeable differences are that it calculates F 500x less, and doesn't rely on global variables.
function T(V,kgrid,zgrid,β,δ)
E=V*P'
T1 = zeros(Float64, 500, 7)
for j = 1:500
for i = 1:7
x = F(zrid[i],kgrid[j]) +(1-δ)*kgrid[j]
T1[j,i] = maximum(u(max(x - kgrid[l], 0)) + β*E[l,i] for l in 1:500)
end
end
return T1
end

Which Lua function is better to use?

I took two ways to round numbers to decimals. First function just rounds the number:
function round(num)
local under = math.floor(num)
local over = math.floor(num) + 1
local underV = -(under - num)
local overV = over - num
if overV > underV then
return under
else
return over
end
end
The next two functions use this function to round a number to decimals:
function roundf(num, dec)
return round(num * (1 * dec)) / (1 * dec)
end
function roundf_alt(num, dec)
local r = math.exp(1 * math.log(dec));
return round(r * num) / r;
end

Why not simply
function round(num)
return num >= 0 and math.floor(num+0.5) or math.ceil(num-0.5)
end
Instead of math.floor(num) + 1 you can simply use math.ceil(num) btw.
Why do you multiply with 1 multiple times?
There are many things to consider when rounding numbers. Please do some research on how to handle special cases.

How to calculate the mean of 3D matrices in an image without NaN?

I need to calculate the mean of a 3D matrices (last step in the code). However, there are many NaNs in the (diff_dataframe./dataframe_vor) calculation. So when I use this code, some results will be NaN. How could I calculate the mean of this matrix by ignoring the NaNs? I attached the code as below.
S.amplitude = 1:20;%:20;
S.blocksize = [1 2 3 4 5 6 8 10 12 15 20];
S.frameWidth = 1920;
S.frameHeight = 1080;
S.quality=0:10:100;
image = 127*ones(S.frameHeight,S.frameWidth,3);
S.yuv2rgb = [1 0 1.28033; 1 -0.21482 -0.38059; 1 2.12798 0];
i_bs = 0;
for BS = S.blocksize
i_bs = i_bs + 1;
hblocks = S.frameWidth / BS;
vblocks = S.frameHeight / BS;
i_a = 0;
dataU = randi([0 1],vblocks,hblocks);
dataV = randi([0 1],vblocks,hblocks);
dataframe_yuv = zeros(S.frameHeight, S.frameWidth, 3);
for x = 1 : hblocks
for y = 1 : vblocks
dataframe_yuv((y-1)*BS+1:y*BS, ...
(x-1)*BS+1:x*BS, 2) = dataU(y,x) * 2 - 1;
dataframe_yuv((y-1)*BS+1:y*BS, ...
(x-1)*BS+1:x*BS, 3) = dataV(y,x) * 2 - 1;
end
end
dataframe_rgb(:,:,1) = S.yuv2rgb(1,1) * dataframe_yuv(:,:,1) + ...
S.yuv2rgb(1,2) * dataframe_yuv(:,:,2) + ...
S.yuv2rgb(1,3) * dataframe_yuv(:,:,3);
dataframe_rgb(:,:,2) = S.yuv2rgb(2,1) * dataframe_yuv(:,:,1) + ...
S.yuv2rgb(2,2) * dataframe_yuv(:,:,2) + ...
S.yuv2rgb(2,3) * dataframe_yuv(:,:,3);
dataframe_rgb(:,:,3) = S.yuv2rgb(3,1) * dataframe_yuv(:,:,1) + ...
S.yuv2rgb(3,2) * dataframe_yuv(:,:,2) + ...
S.yuv2rgb(3,3) * dataframe_yuv(:,:,3);
for A = S.amplitude
i_a = i_a + 1;
i_q = 0;
image1p = round(image + dataframe_rgb * A);
image1n = round(image - dataframe_rgb * A);
dataframe_vor = ((image1p-image1n)/2)/255;
for Q = S.quality
i_q = i_q + 1;
namestrp = ['greyjpegs/Img_BS' num2str(BS) '_A' num2str(A) '_Q' num2str(Q) '_1p.jpg'];
namestrn = ['greyjpegs/Img_BS' num2str(BS) '_A' num2str(A) '_Q' num2str(Q) '_1n.jpg'];
imwrite(image1p/255,namestrp,'jpg', 'Quality', Q);
imwrite(image1n/255,namestrn,'jpg', 'Quality', Q);
error_mean(i_bs, i_a, i_q) = mean2((abs(diff_dataframe./dataframe_vor)));
end
end
end

mean2 is a shortcut function that's part of the image processing toolbox that finds the entire average of a 2D region which doesn't include handling NaN. In that case, simply remove all values that are NaN and find the resulting average. Note that the removal of NaN unrolls the 2D region into a 1D vector, so we can simply use mean in this case. As an additional check, let's make sure there are no divide by 0 errors, so also check for Inf as well.
Therefore, replace this line:
error_mean(i_bs, i_a, i_q) = mean2((abs(diff_dataframe./dataframe_vor)));
... with:
tmp = abs(diff_dataframe ./ dataframe_vor);
mask = ~isnan(tmp) | ~isinf(tmp);
tmp = tmp(mask);
if isempty(tmp)
error_mean(i_bs, i_a, i_q) = 0;
else
error_mean(i_bs, i_a, i_q) = mean(tmp);
We first assign the desired operation to a temporary variable, use isnan and isinf to remove out the offending values, then find the average of the rest. One intricacy is that if your entire region is NaN or Inf, then the removal of all these entries in the region results in the empty vector, and finding the mean of this undefined. A separate check is there to be sure that if it's empty, simply assign the value of 0 instead.

Filling a matrix using parallel processing in Julia

I'm trying to speed up the solution time for a dynamic programming problem in Julia (v. 0.5.0), via parallel processing. The problem involves choosing the optimal values for every element of a 1073 x 19 matrix at every iteration, until successive matrix differences fall within a tolerance. I thought that, within each iteration, filling in the values for each element of the matrix could be parallelized. However, I'm seeing a huge performance degradation using SharedArray, and I'm wondering if there's a better way to approach parallel processing for this problem.
I construct the arguments for the function below:
est_params = [.788,.288,.0034,.1519,.1615,.0041,.0077,.2,0.005,.7196]
r = 0.015
tau = 0.35
rho =est_params[1]
sigma =est_params[2]
delta = 0.15
gamma =est_params[3]
a_capital =est_params[4]
lambda1 =est_params[5]
lambda2 =est_params[6]
s =est_params[7]
theta =est_params[8]
mu =est_params[9]
p_bar_k_ss =est_params[10]
beta = (1+r)^(-1)
sigma_range = 4
gz = 19
gp = 29
gk = 37
lnz=collect(linspace(-sigma_range*sigma,sigma_range*sigma,gz))
z=exp(lnz)
gk_m = fld(gk,2)
# Need to add mu somewhere to k_ss
k_ss = (theta*(1-tau)/(r+delta))^(1/(1-theta))
k=cat(1,map(i->k_ss*((1-delta)^i),collect(1:gk_m)),map(i->k_ss/((1-delta)^i),collect(1:gk_m)))
insert!(k,gk_m+1,k_ss)
sort!(k)
p_bar=p_bar_k_ss*k_ss
p = collect(linspace(-p_bar/2,p_bar,gp))
#Tauchen
N = length(z)
Z = zeros(N,1)
Zprob = zeros(Float32,N,N)
Z[N] = lnz[length(z)]
Z[1] = lnz[1]
zstep = (Z[N] - Z[1]) / (N - 1)
for i=2:(N-1)
Z[i] = Z[1] + zstep * (i - 1)
end
for a = 1 : N
for b = 1 : N
if b == 1
Zprob[a,b] = 0.5*erfc(-((Z[1] - mu - rho * Z[a] + zstep / 2) / sigma)/sqrt(2))
elseif b == N
Zprob[a,b] = 1 - 0.5*erfc(-((Z[N] - mu - rho * Z[a] - zstep / 2) / sigma)/sqrt(2))
else
Zprob[a,b] = 0.5*erfc(-((Z[b] - mu - rho * Z[a] + zstep / 2) / sigma)/sqrt(2)) -
0.5*erfc(-((Z[b] - mu - rho * Z[a] - zstep / 2) / sigma)/sqrt(2))
end
end
end
# Collecting tauchen results in a 2 element array of linspace and array; [2] gets array
# Zprob=collect(tauchen(gz, rho, sigma, mu, sigma_range))[2]
Zcumprob=zeros(Float32,gz,gz)
# 2 in cumsum! denotes the 2nd dimension, i.e. columns
cumsum!(Zcumprob, Zprob,2)
gm = gk * gp
control=zeros(gm,2)
for i=1:gk
control[(1+gp*(i-1)):(gp*i),1]=fill(k[i],(gp,1))
control[(1+gp*(i-1)):(gp*i),2]=p
end
endog=copy(control)
E=Array(Float32,gm,gm,gz)
for h=1:gm
for m=1:gm
for j=1:gz
# set the nonzero net debt indicator
if endog[h,2]<0
p_ind=1
else
p_ind=0
end
# set the investment indicator
if (control[m,1]-(1-delta)*endog[h,1])!=0
i_ind=1
else
i_ind=0
end
E[m,h,j] = (1-tau)*z[j]*(endog[h,1]^theta) + control[m,2]-endog[h,2]*(1+r*(1-tau)) +
delta*endog[h,1]*tau-(control[m,1]-(1-delta)*endog[h,1]) -
(i_ind*gamma*endog[h,1]+endog[h,1]*(a_capital/2)*(((control[m,1]-(1-delta)*endog[h,1])/endog[h,1])^2)) +
s*endog[h,2]*p_ind
elem = E[m,h,j]
if E[m,h,j]<0
E[m,h,j]=elem+lambda1*elem-.5*lambda2*elem^2
else
E[m,h,j]=elem
end
end
end
end
I then constructed the function with serial processing. The two for loops iterate through each element to find the largest value in a 1072-sized (=the gm scalar argument in the function) array:
function dynam_serial(E,gm,gz,beta,Zprob)
v = Array(Float32,gm,gz )
fill!(v,E[cld(gm,2),cld(gm,2),cld(gz,2)])
Tv = Array(Float32,gm,gz)
# Set parameters for the loop
convcrit = 0.0001 # chosen convergence criterion
diff = 1 # arbitrary initial value greater than convcrit
while diff>convcrit
exp_v=v*Zprob'
for h=1:gm
for j=1:gz
Tv[h,j]=findmax(E[:,h,j] + beta*exp_v[:,j])[1]
end
end
diff = maxabs(Tv - v)
v=copy(Tv)
end
end
Timing this, I get:
#time dynam_serial(E,gm,gz,beta,Zprob)
> 106.880008 seconds (91.70 M allocations: 203.233 GB, 15.22% gc time)
Now, I try using Shared Arrays to benefit from parallel processing. Note that I reconfigured the iteration so that I only have one for loop, rather than two. I also use v=deepcopy(Tv); otherwise, v is copied as an Array object, rather than a SharedArray:
function dynam_parallel(E,gm,gz,beta,Zprob)
v = SharedArray(Float32,(gm,gz),init = S -> S[Base.localindexes(S)] = myid() )
fill!(v,E[cld(gm,2),cld(gm,2),cld(gz,2)])
# Set parameters for the loop
convcrit = 0.0001 # chosen convergence criterion
diff = 1 # arbitrary initial value greater than convcrit
while diff>convcrit
exp_v=v*Zprob'
Tv = SharedArray(Float32,gm,gz,init = S -> S[Base.localindexes(S)] = myid() )
#sync #parallel for hj=1:(gm*gz)
j=cld(hj,gm)
h=mod(hj,gm)
if h==0;h=gm;end;
#async Tv[h,j]=findmax(E[:,h,j] + beta*exp_v[:,j])[1]
end
diff = maxabs(Tv - v)
v=deepcopy(Tv)
end
end
Timing the parallel version; and using a 4-core 2.5 GHz I7 processor with 16GB of memory, I get:
addprocs(3)
#time dynam_parallel(E,gm,gz,beta,Zprob)
> 164.237208 seconds (2.64 M allocations: 201.812 MB, 0.04% gc time)
Am I doing something incorrect here? Or is there a better way to approach parallel processing in Julia for this particular problem? I've considered using Distributed Arrays, but it's difficult for me to see how to apply them to the present problem.
UPDATE:
Per #DanGetz and his helpful comments, I turned instead to trying to speed up the serial processing version. I was able to get performance down to 53.469780 seconds (67.36 M allocations: 103.419 GiB, 19.12% gc time) through:
1) Upgrading to 0.6.0 (saved about 25 seconds), which includes the helpful #views macro.
2) Preallocating the main array I'm trying to fill in (Tv), per the section on Preallocating Outputs in the Julia Performance Tips: https://docs.julialang.org/en/latest/manual/performance-tips/. (saved another 25 or so seconds)
The biggest remaining slow-down seems to be coming from the add_vecs function, which sums together subarrays of two larger matrices. I've tried devectorizing and using BLAS functions, but haven't been able to produce better performance.
In any event, the improved code for dynam_serial is below:
function add_vecs(r::Array{Float32},h::Int,j::Int,E::Array{Float32},exp_v::Array{Float32},beta::Float32)
#views r=E[:,h,j] + beta*exp_v[:,j]
return r
end
function dynam_serial(E::Array{Float32},gm::Int,gz::Int,beta::Float32,Zprob::Array{Float32})
v = Array{Float32}(gm,gz)
fill!(v,E[cld(gm,2),cld(gm,2),cld(gz,2)])
Tv = Array{Float32}(gm,gz)
r = Array{Float32}(gm)
# Set parameters for the loop
convcrit = 0.0001 # chosen convergence criterion
diff = 1 # arbitrary initial value greater than convcrit
while diff>convcrit
exp_v=v*Zprob'
for h=1:gm
for j=1:gz
#views Tv[h,j]=findmax(add_vecs(r,h,j,E,exp_v,beta))[1]
end
end
diff = maximum(abs,Tv - v)
v=copy(Tv)
end
return Tv
end

If add_vecs seems to be the critical function, writing an explicit for loop could offer more optimization. How does the following benchmark:
function add_vecs!(r::Array{Float32},h::Int,j::Int,E::Array{Float32},
exp_v::Array{Float32},beta::Float32)
#inbounds for i=1:size(E,1)
r[i]=E[i,h,j] + beta*exp_v[i,j]
end
return r
end
UPDATE
To continue optimizing dynam_serial I have tried to remove more allocations. The result is:
function add_vecs_and_max!(gm::Int,r::Array{Float64},h::Int,j::Int,E::Array{Float64},
exp_v::Array{Float64},beta::Float64)
#inbounds for i=1:gm
r[i] = E[i,h,j]+beta*exp_v[i,j]
end
return findmax(r)[1]
end
function dynam_serial(E::Array{Float64},gm::Int,gz::Int,
beta::Float64,Zprob::Array{Float64})
v = Array{Float64}(gm,gz)
fill!(v,E[cld(gm,2),cld(gm,2),cld(gz,2)])
r = Array{Float64}(gm)
exp_v = Array{Float64}(gm,gz)
# Set parameters for the loop
convcrit = 0.0001 # chosen convergence criterion
diff = 1.0 # arbitrary initial value greater than convcrit
while diff>convcrit
A_mul_Bt!(exp_v,v,Zprob)
diff = -Inf
for h=1:gm
for j=1:gz
oldv = v[h,j]
newv = add_vecs_and_max!(gm,r,h,j,E,exp_v,beta)
v[h,j]= newv
diff = max(diff, oldv-newv, newv-oldv)
end
end
end
return v
end
Switching the functions to use Float64 should increase speed (as CPUs are inherently optimized for 64-bit word lengths). Also, using the mutating A_mul_Bt! directly saves another allocation. Avoiding the copy(...) by switching the arrays v and Tv.
How do these optimizations improve your running time?
2nd UPDATE
Updated the code in the UPDATE section to use findmax. Also, changed dynam_serial to use v without Tv, as there was no need to save the old version except for the diff calculation, which is now done inside the loop.

Here's the code I copied-and-pasted, provided by Dan Getz above. I include the array and scalar definitions exactly as I ran them. Performance was: 39.507005 seconds (11 allocations: 486.891 KiB) when running #time dynam_serial(E,gm,gz,beta,Zprob).
using SpecialFunctions
est_params = [.788,.288,.0034,.1519,.1615,.0041,.0077,.2,0.005,.7196]
r = 0.015
tau = 0.35
rho =est_params[1]
sigma =est_params[2]
delta = 0.15
gamma =est_params[3]
a_capital =est_params[4]
lambda1 =est_params[5]
lambda2 =est_params[6]
s =est_params[7]
theta =est_params[8]
mu =est_params[9]
p_bar_k_ss =est_params[10]
beta = (1+r)^(-1)
sigma_range = 4
gz = 19 #15 #19
gp = 29 #19 #29
gk = 37 #25 #37
lnz=collect(linspace(-sigma_range*sigma,sigma_range*sigma,gz))
z=exp.(lnz)
gk_m = fld(gk,2)
# Need to add mu somewhere to k_ss
k_ss = (theta*(1-tau)/(r+delta))^(1/(1-theta))
k=cat(1,map(i->k_ss*((1-delta)^i),collect(1:gk_m)),map(i->k_ss/((1-delta)^i),collect(1:gk_m)))
insert!(k,gk_m+1,k_ss)
sort!(k)
p_bar=p_bar_k_ss*k_ss
p = collect(linspace(-p_bar/2,p_bar,gp))
#Tauchen
N = length(z)
Z = zeros(N,1)
Zprob = zeros(Float64,N,N)
Z[N] = lnz[length(z)]
Z[1] = lnz[1]
zstep = (Z[N] - Z[1]) / (N - 1)
for i=2:(N-1)
Z[i] = Z[1] + zstep * (i - 1)
end
for a = 1 : N
for b = 1 : N
if b == 1
Zprob[a,b] = 0.5*erfc(-((Z[1] - mu - rho * Z[a] + zstep / 2) / sigma)/sqrt(2))
elseif b == N
Zprob[a,b] = 1 - 0.5*erfc(-((Z[N] - mu - rho * Z[a] - zstep / 2) / sigma)/sqrt(2))
else
Zprob[a,b] = 0.5*erfc(-((Z[b] - mu - rho * Z[a] + zstep / 2) / sigma)/sqrt(2)) -
0.5*erfc(-((Z[b] - mu - rho * Z[a] - zstep / 2) / sigma)/sqrt(2))
end
end
end
# Collecting tauchen results in a 2 element array of linspace and array; [2] gets array
# Zprob=collect(tauchen(gz, rho, sigma, mu, sigma_range))[2]
Zcumprob=zeros(Float64,gz,gz)
# 2 in cumsum! denotes the 2nd dimension, i.e. columns
cumsum!(Zcumprob, Zprob,2)
gm = gk * gp
control=zeros(gm,2)
for i=1:gk
control[(1+gp*(i-1)):(gp*i),1]=fill(k[i],(gp,1))
control[(1+gp*(i-1)):(gp*i),2]=p
end
endog=copy(control)
E=Array(Float64,gm,gm,gz)
for h=1:gm
for m=1:gm
for j=1:gz
# set the nonzero net debt indicator
if endog[h,2]<0
p_ind=1
else
p_ind=0
end
# set the investment indicator
if (control[m,1]-(1-delta)*endog[h,1])!=0
i_ind=1
else
i_ind=0
end
E[m,h,j] = (1-tau)*z[j]*(endog[h,1]^theta) + control[m,2]-endog[h,2]*(1+r*(1-tau)) +
delta*endog[h,1]*tau-(control[m,1]-(1-delta)*endog[h,1]) -
(i_ind*gamma*endog[h,1]+endog[h,1]*(a_capital/2)*(((control[m,1]-(1-delta)*endog[h,1])/endog[h,1])^2)) +
s*endog[h,2]*p_ind
elem = E[m,h,j]
if E[m,h,j]<0
E[m,h,j]=elem+lambda1*elem-.5*lambda2*elem^2
else
E[m,h,j]=elem
end
end
end
end
function add_vecs_and_max!(gm::Int,r::Array{Float64},h::Int,j::Int,E::Array{Float64},
exp_v::Array{Float64},beta::Float64)
maxr = -Inf
#inbounds for i=1:gm r[i] = E[i,h,j]+beta*exp_v[i,j]
maxr = max(r[i],maxr)
end
return maxr
end
function dynam_serial(E::Array{Float64},gm::Int,gz::Int,
beta::Float64,Zprob::Array{Float64})
v = Array{Float64}(gm,gz)
fill!(v,E[cld(gm,2),cld(gm,2),cld(gz,2)])
Tv = Array{Float64}(gm,gz)
r = Array{Float64}(gm)
exp_v = Array{Float64}(gm,gz)
# Set parameters for the loop
convcrit = 0.0001 # chosen convergence criterion
diff = 1.0 # arbitrary initial value greater than convcrit
while diff>convcrit
A_mul_Bt!(exp_v,v,Zprob)
diff = -Inf
for h=1:gm
for j=1:gz
Tv[h,j]=add_vecs_and_max!(gm,r,h,j,E,exp_v,beta)
diff = max(abs(Tv[h,j]-v[h,j]),diff)
end
end
(v,Tv)=(Tv,v)
end
return v
end
Now, here's another version of the algorithm and inputs. The functions are similar to what Dan Getz suggested, except that I use findmax rather than an iterated max function to find the array maximum. In the input construction, I am using both Float32 and mixing different bit-types together. However, I've consistently achieved better performance this way: 24.905569 seconds (1.81 k allocations: 46.829 MiB, 0.01% gc time). But it's not clear at all why.
using SpecialFunctions
est_params = [.788,.288,.0034,.1519,.1615,.0041,.0077,.2,0.005,.7196]
r = 0.015
tau = 0.35
rho =est_params[1]
sigma =est_params[2]
delta = 0.15
gamma =est_params[3]
a_capital =est_params[4]
lambda1 =est_params[5]
lambda2 =est_params[6]
s =est_params[7]
theta =est_params[8]
mu =est_params[9]
p_bar_k_ss =est_params[10]
beta = Float32((1+r)^(-1))
sigma_range = 4
gz = 19
gp = 29
gk = 37
lnz=collect(linspace(-sigma_range*sigma,sigma_range*sigma,gz))
z=exp(lnz)
gk_m = fld(gk,2)
# Need to add mu somewhere to k_ss
k_ss = (theta*(1-tau)/(r+delta))^(1/(1-theta))
k=cat(1,map(i->k_ss*((1-delta)^i),collect(1:gk_m)),map(i->k_ss/((1-delta)^i),collect(1:gk_m)))
insert!(k,gk_m+1,k_ss)
sort!(k)
p_bar=p_bar_k_ss*k_ss
p = collect(linspace(-p_bar/2,p_bar,gp))
#Tauchen
N = length(z)
Z = zeros(N,1)
Zprob = zeros(Float32,N,N)
Z[N] = lnz[length(z)]
Z[1] = lnz[1]
zstep = (Z[N] - Z[1]) / (N - 1)
for i=2:(N-1)
Z[i] = Z[1] + zstep * (i - 1)
end
for a = 1 : N
for b = 1 : N
if b == 1
Zprob[a,b] = 0.5*erfc(-((Z[1] - mu - rho * Z[a] + zstep / 2) / sigma)/sqrt(2))
elseif b == N
Zprob[a,b] = 1 - 0.5*erfc(-((Z[N] - mu - rho * Z[a] - zstep / 2) / sigma)/sqrt(2))
else
Zprob[a,b] = 0.5*erfc(-((Z[b] - mu - rho * Z[a] + zstep / 2) / sigma)/sqrt(2)) -
0.5*erfc(-((Z[b] - mu - rho * Z[a] - zstep / 2) / sigma)/sqrt(2))
end
end
end
# Collecting tauchen results in a 2 element array of linspace and array; [2] gets array
# Zprob=collect(tauchen(gz, rho, sigma, mu, sigma_range))[2]
Zcumprob=zeros(Float32,gz,gz)
# 2 in cumsum! denotes the 2nd dimension, i.e. columns
cumsum!(Zcumprob, Zprob,2)
gm = gk * gp
control=zeros(gm,2)
for i=1:gk
control[(1+gp*(i-1)):(gp*i),1]=fill(k[i],(gp,1))
control[(1+gp*(i-1)):(gp*i),2]=p
end
endog=copy(control)
E=Array(Float32,gm,gm,gz)
for h=1:gm
for m=1:gm
for j=1:gz
# set the nonzero net debt indicator
if endog[h,2]<0
p_ind=1
else
p_ind=0
end
# set the investment indicator
if (control[m,1]-(1-delta)*endog[h,1])!=0
i_ind=1
else
i_ind=0
end
E[m,h,j] = (1-tau)*z[j]*(endog[h,1]^theta) + control[m,2]-endog[h,2]*(1+r*(1-tau)) +
delta*endog[h,1]*tau-(control[m,1]-(1-delta)*endog[h,1]) -
(i_ind*gamma*endog[h,1]+endog[h,1]*(a_capital/2)*(((control[m,1]-(1-delta)*endog[h,1])/endog[h,1])^2)) +
s*endog[h,2]*p_ind
elem = E[m,h,j]
if E[m,h,j]<0
E[m,h,j]=elem+lambda1*elem-.5*lambda2*elem^2
else
E[m,h,j]=elem
end
end
end
end
function add_vecs!(gm::Int,r::Array{Float32},h::Int,j::Int,E::Array{Float32},
exp_v::Array{Float32},beta::Float32)
#inbounds #views for i=1:gm
r[i]=E[i,h,j] + beta*exp_v[i,j]
end
return r
end
function dynam_serial(E::Array{Float32},gm::Int,gz::Int,beta::Float32,Zprob::Array{Float32})
v = Array{Float32}(gm,gz)
fill!(v,E[cld(gm,2),cld(gm,2),cld(gz,2)])
Tv = Array{Float32}(gm,gz)
# Set parameters for the loop
convcrit = 0.0001 # chosen convergence criterion
diff = 1.00000 # arbitrary initial value greater than convcrit
iter=0
exp_v=Array{Float32}(gm,gz)
r=Array{Float32}(gm)
while diff>convcrit
A_mul_Bt!(exp_v,v,Zprob)
for h=1:gm
for j=1:gz
Tv[h,j]=findmax(add_vecs!(gm,r,h,j,E,exp_v,beta))[1]
end
end
diff = maximum(abs,Tv - v)
(v,Tv)=(Tv,v)
end
return v
end

Fastest solution for all possible combinations, taking k elements out of n possible with k>2 and n large

I am using MATLAB to find all of the possible combinations of k elements out of n possible elements. I stumbled across this question, but unfortunately it does not solve my problem. Of course, neither does nchoosek as my n is around 100.
Truth is, I don't need all of the possible combinations at the same time. I will explain what I need, as there might be an easier way to achieve the desired result. I have a matrix M of 100 rows and 25 columns.
Think of a submatrix of M as a matrix formed by ALL columns of M and only a subset of the rows. I have a function f that can be applied to any matrix which gives a result of either -1 or 1. For example, you can think of the function as sign(det(A)) where A is any matrix (the exact function is irrelevant for this part of the question).
I want to know what is the biggest number of rows of M for which the submatrix A formed by these rows is such that f(A) = 1. Notice that if f(M) = 1, I am done. However, if this is not the case then I need to start combining rows, starting of all combinations with 99 rows, then taking the ones with 98 rows, and so on.
Up to this point, my implementation had to do with nchoosek which worked when M had only a few rows. However, now that I am working with a relatively bigger dataset, things get stuck. Do any of you guys think of a way to implement this without having to use the above function? Any help would be gladly appreciated.
Here is my minimal working example, it works for small obs_tot but fails when I try to use bigger numbers:
value = -1; obs_tot = 100; n_rows = 25;
mat = randi(obs_tot,n_rows);
while value == -1
posibles = nchoosek(1:obs_tot,i);
[num_tries,num_obs] = size(possibles);
num_try = 1;
while value == 0 && num_try <= num_tries
check = mat(possibles(num_try,:),:);
value = sign(det(check));
num_try = num_try + 1;
end
i = i - 1;
end
obs_used = possibles(num_try-1,:)';

Preamble
As yourself noticed in your question, it would be nice not to have nchoosek to return all possible combinations at the same time but rather to enumerate them one by one in order not to explode memory when n becomes large. So something like:
enumerator = CombinationEnumerator(k, n);
while(enumerator.MoveNext())
currentCombination = enumerator.Current;
...
end
Here is an implementation of such enumerator as a Matlab class. It is based on classic IEnumerator<T> interface in C# / .NET and mimics the subfunction combs in nchoosek (the unrolled way):
%
% PURPOSE:
%
% Enumerates all combinations of length 'k' in a set of length 'n'.
%
% USAGE:
%
% enumerator = CombinaisonEnumerator(k, n);
% while(enumerator.MoveNext())
% currentCombination = enumerator.Current;
% ...
% end
%
%% ---
classdef CombinaisonEnumerator < handle
properties (Dependent) % NB: Matlab R2013b bug => Dependent must be declared before their get/set !
Current; % Gets the current element.
end
methods
function [enumerator] = CombinaisonEnumerator(k, n)
% Creates a new combinations enumerator.
if (~isscalar(n) || (n < 1) || (~isreal(n)) || (n ~= round(n))), error('`n` must be a scalar positive integer.'); end
if (~isscalar(k) || (k < 0) || (~isreal(k)) || (k ~= round(k))), error('`k` must be a scalar positive or null integer.'); end
if (k > n), error('`k` must be less or equal than `n`'); end
enumerator.k = k;
enumerator.n = n;
enumerator.v = 1:n;
enumerator.Reset();
end
function [b] = MoveNext(enumerator)
% Advances the enumerator to the next element of the collection.
if (~enumerator.isOkNext),
b = false; return;
end
if (enumerator.isInVoid)
if (enumerator.k == enumerator.n),
enumerator.isInVoid = false;
enumerator.current = enumerator.v;
elseif (enumerator.k == 1)
enumerator.isInVoid = false;
enumerator.index = 1;
enumerator.current = enumerator.v(enumerator.index);
else
enumerator.isInVoid = false;
enumerator.index = 1;
enumerator.recursion = CombinaisonEnumerator(enumerator.k - 1, enumerator.n - enumerator.index);
enumerator.recursion.v = enumerator.v((enumerator.index + 1):end); % adapt v (todo: should use private constructor)
enumerator.recursion.MoveNext();
enumerator.current = [enumerator.v(enumerator.index) enumerator.recursion.Current];
end
else
if (enumerator.k == enumerator.n),
enumerator.isInVoid = true;
enumerator.isOkNext = false;
elseif (enumerator.k == 1)
enumerator.index = enumerator.index + 1;
if (enumerator.index <= enumerator.n)
enumerator.current = enumerator.v(enumerator.index);
else
enumerator.isInVoid = true;
enumerator.isOkNext = false;
end
else
if (enumerator.recursion.MoveNext())
enumerator.current = [enumerator.v(enumerator.index) enumerator.recursion.Current];
else
enumerator.index = enumerator.index + 1;
if (enumerator.index <= (enumerator.n - enumerator.k + 1))
enumerator.recursion = CombinaisonEnumerator(enumerator.k - 1, enumerator.n - enumerator.index);
enumerator.recursion.v = enumerator.v((enumerator.index + 1):end); % adapt v (todo: should use private constructor)
enumerator.recursion.MoveNext();
enumerator.current = [enumerator.v(enumerator.index) enumerator.recursion.Current];
else
enumerator.isInVoid = true;
enumerator.isOkNext = false;
end
end
end
end
b = enumerator.isOkNext;
end
function [] = Reset(enumerator)
% Sets the enumerator to its initial position, which is before the first element.
enumerator.isInVoid = true;
enumerator.isOkNext = (enumerator.k > 0);
end
function [c] = get.Current(enumerator)
if (enumerator.isInVoid), error('Enumerator is positioned (before/after) the (first/last) element.'); end
c = enumerator.current;
end
end
properties (GetAccess=private, SetAccess=private)
k = [];
n = [];
v = [];
index = [];
recursion = [];
current = [];
isOkNext = false;
isInVoid = true;
end
end
We can test implementation is ok from command window like this:
>> e = CombinaisonEnumerator(3, 6);
>> while(e.MoveNext()), fprintf(1, '%s\n', num2str(e.Current)); end
Which returns as expected the following n!/(k!*(n-k)!) combinations:
1 2 3
1 2 4
1 2 5
1 2 6
1 3 4
1 3 5
1 3 6
1 4 5
1 4 6
1 5 6
2 3 4
2 3 5
2 3 6
2 4 5
2 4 6
2 5 6
3 4 5
3 4 6
3 5 6
4 5 6
Implementation of this enumerator may be further optimized for speed, or by enumerating combinations in an order more appropriate for your case (e.g., test some combinations first rather than others) ... Well, at least it works! :)
Problem solving
Now solving your problem is really easy:
n = 100;
m = 25;
matrix = rand(n, m);
k = n;
cont = true;
while(cont && (k >= 1))
e = CombinationEnumerator(k, n);
while(cont && e.MoveNext());
cont = f(matrix(e.Current(:), :)) ~= 1;
end
if (cont), k = k - 1; end
end

Resources