Recommended direct solver for sparse positive definite linear system in scipy? - scipy

I'm sorry if this is explained clearly in the scipy.sparse documentation.
When using scipy, what function would you recommend using to solve a sparse positive definite linear system of equations? I want to use a direct method, and I want the columns to be reordered so as to preserve sparsity as much as possible in the Cholesky factorization of the coefficient matrix. Ideally I'd be able to experiment with various options for reordering.
Does a direct solver for sparse positive definite systems exist in scipy.sparse? Is scikit.sparse the way to go?

scipy.sparse.linalg.spsolve
is clear enough,
but it seems that for speed you must
pip install scikit-umfpack
or else
build UMFPACK and AMD from
SuiteSparse
then rebuild scipy from source, with
[umfpack] umfpack_libs = ... in site.cfg .
otherwise scipy.sparse.linalg defaults to the slower SuperLU.
Is scikit.sparse the way to go?
Compared to what, with what criteria ?
If C / C++ is enough for you, use SuiteSparse directly.
Any tool depends on what you're comfortable with,
and on users: one, two, many.
Maybe better visualization would help your project more than faster spsolve.
Some pretty obvious pluses and minuses of scipy.sparse:
+ python for fast development, data input -- matrices -- visualize
+ several packages have built on scipy.sparse;
ask around in your application area (which is ?)
- rough edges (matrices are a pain), with afaik no wiki to collect hints and code snippets
- layers upon layers, scipy.sparse -- SuiteSparse -- ... BLAS ...
make timing and debugging tough.
Fwiw, solver times vary a lot on my iMac.
These are all with default args, without umfpack.
(This is NOT a realistic testcase; but
satisficing is often good enough.)
X = sparse.rand( m, n, dens, format="csr" )
A = 1e-6 * sparse.eye(m) + X * X.T
linalg solvers( A, b )
(5000, 5000) density 0.42 % --
51 msec spsolve
5 msec bicg
3 msec bicgstab
2 msec cg
4 msec cgs
3 msec gmres
4 msec lgmres
1 msec minres
6 msec qmr
5 msec lsmr
(5000, 5000) density 0.84 % --
428 msec spsolve
12 msec bicg
7 msec bicgstab
5 msec cg
10 msec cgs
6 msec gmres
8 msec lgmres
2 msec minres
13 msec qmr
12 msec lsmr
(5000, 5000) density 1.3 % --
1462 msec spsolve
16 msec bicg
9 msec bicgstab
7 msec cg
11 msec cgs
7 msec gmres
10 msec lgmres
1 msec minres
18 msec qmr
14 msec lsmr

Related

Mixed-effects linear regression model using multiple independent measurements

I am trying to implement a linear mixed effect (LME) regression model for an x-ray imaging quality metric "CNR" (contrast-to-noise ratio) for which I measured for various tube potentials (kV) and filtration materials (Filter). CNR was measured for 3 consecutive slices so I have a standard deviation of the CNR from these independent measurements as well. I am wondering how I can incorporate these multiple independent measurements in my analysis. A representation of the data for a single measurement and my first attempt using fitlme is shown below. I tried looking at online resources but could not find an answer to my specific questions.
kV=[80 90 100 80 90 100 80 90 100]';
Filter={'Al','Al','Al','Cu','Cu','Cu','Ti','Ti','Ti'}';
CNR=[10 9 8 10.1 8.9 7.9 7 6 5]';
T=table(kV,Filter,CNR);
kV Filter CNR
___ ______ ___
80 'Al' 10
90 'Al' 9
100 'Al' 8
80 'Cu' 10.1
90 'Cu' 8.9
100 'Cu' 7.9
80 'Ti' 7
90 'Ti' 6
100 'Ti' 5
OUTPUT
Linear mixed-effects model fit by ML
Model information:
Number of observations 9
Fixed effects coefficients 4
Random effects coefficients 0
Covariance parameters 1
Formula:
CNR ~ 1 + kV + Filter
Model fit statistics:
AIC BIC LogLikelihood Deviance
-19.442 -18.456 14.721 -29.442
Fixed effects coefficients (95% CIs):
Name Estimate SE pValue
'(Intercept)' 18.3 0.17533 1.5308e-09
'kV' -0.10333 0.0019245 4.2372e-08
'Filter_Cu' -0.033333 0.03849 -0.86603
'Filter_Ti' -3 0.03849 -77.942
Random effects covariance parameters (95% CIs):
Group: Error
Name Estimate Lower Upper
'Res Std' 0.04714 0.0297 0.074821
Questions/Issues with current implementation:
How is the fixed effects coefficients for '(Intercept)' with P=1.53E-9 interpreted?
I only included fixed effects. Should the standard deviation of the ROI measurements somehow be incorporated into the random effects as well?
How do I incorporate the three independent measurements of CNR for three consecutive slices for a give kV/filter combination? Should I just add more rows to the table "T"? This would result in a total of 27 observations.

Calculating Q value in dqn with experience replay

consider the Deep Q-Learning algorithm
1 initialize replay memory D
2 initialize action-value function Q with random weights
3 observe initial state s
4 repeat
5 select an action a
6 with probability ε select a random action
7 otherwise select a = argmaxa’Q(s,a’)
8 carry out action a
9 observe reward r and new state s’
10 store experience <s, a, r, s’> in replay memory D
11
12 sample random transitions <ss, aa, rr, ss’> from replay memory D
13 calculate target for each minibatch transition
14 if ss’ is terminal state then tt = rr
15 otherwise tt = rr + γmaxa’Q(ss’, aa’)
16 train the Q network using (tt - Q(ss, aa))^2 as loss
17
18 s = s'
19 until terminated
In step 16 the value of Q(ss, aa) is used to calculate the loss. When is this Q value calculated? At the time the action was taken or during the training itself?
Since replay memory only stores < s,a,r,s' > and not the q-value, is it safe to assume the q value will be calculated during the time of training?
Yes, in step 16, when training the network, you are using the the loss function (tt - Q(ss, aa))^2 because you want to update network weights in order to approximate the most recent Q-values, computed as rr + γmaxa’Q(ss’, aa’) and used as target. Therefore, Q(ss, aa) is the current estimation, which is typically computed during training time.
Here you can find a Jupyter Notebook with a simply Deep Q-learning implementation that maybe is helpful.

Forcestop GlobalSearch function in matlab

I'm having an issue with an optimization problem. I think the code works fine, but I would like to force-stop it when it has reached a few optimal values(under 1e-11). The only solution I have found so far is to set MaxTime to 120
optionsMS = ptimoptions(#fmincon,'MaxIter',1000000,'MaxFunEvals',1000000,...
'Algorithm','sqp');
problem = createOptimProblem('fmincon','lb',LB,'ub',UB,'x0',vx,...
'options',optionsMS,...
'objective',#(v)opti(v,x0,N,imag(wFs),imag(wPs),epsilon,R,W));
gs=GlobalSearch('Display','iter','StartPointsToRun','bounds','MaxTime',120);
[x3,fval3] = run(gs,problem)
What else could I do?
EDIT:
I have set TolFun to 1e-11
optionsGS = optimoptions(#fmincon,'MaxIter',1000000,'MaxFunEvals',1000000,...
'Algorithm','interior-point','TolFun',1e-11);
And these are the results
Num Pts Best Current Threshold Local Local
Analyzed F-count f(x) Penalty Penalty f(x) exitflag Procedure
0 2961 0.006034 0.006034 2 Initial Point
200 16921 0.006034 0.01213 2 Stage 1 Local
300 17023 0.006034 6.352 1.086 Stage 2 Search
400 17123 0.006034 6.322 4.191 Stage 2 Search
432 26074 4.379e-13 6.228 6.475 4.379e-13 2 Stage 2 Local
433 35845 4.379e-13 6.205 6.228 4.244e-13 2 Stage 2 Local
434 44004 4.379e-13 1.995 6.205 0.01337 2 Stage 2 Local
435 52661 3.944e-13 1.514 1.995 3.944e-13 2 Stage 2 Local
474 54906 3.944e-13 1.939 2.017 0.01789 -1 Stage 2 Local
GlobalSearch stopped because maximum time is exceeded.
GlobalSearch called the local solver 7 times before exceeding
the clock time limit (MaxTime = 120 seconds).
6 local solver runs converged with a positive local solver exit flag.
Three results are below 1e-11 so I think it should stop but I dont know how to set that on the optimoptions.
Thanks in advance.
You can set a TolFun or TolX value for your ptimoptions, depending on what your stopping criteria is.
If you want to set a lower bound on the size step, go for TolX. If your goal is to constrain the lower bound in the change of the objective function, set TolFun instead. The Matlab documentation has a nice graph showing the difference between them:
I believe you are looking for the TolFun option. That being said, besides changing the MaxTime, you could try the following:
optionsMS = ptimoptions(#fmincon,'MaxIter',1000000,'MaxFunEvals',1000000,...
'Algorithm','sqp','TolFun',1e-11);

How to compute & plot Equal Error Rate (EER) from FAR/FRR values using matlab

I have the following values against FAR/FRR. i want to compute EER rates and then plot in matlab.
FAR FRR
19.64 20
21.29 18.61
24.92 17.08
19.14 20.28
17.99 21.39
16.83 23.47
15.35 26.39
13.20 29.17
7.92 42.92
3.96 60.56
1.82 84.31
1.65 98.33
26.07 16.39
29.04 13.13
34.49 9.31
40.76 6.81
50.33 5.42
66.83 1.67
82.51 0.28
Is there any matlab function available to do this. can somebody explain this to me. Thanks.
Let me try to answer your question
1) For your data EER can be the mean/max/min of [19.64,20]
1.1) The idea of EER is try to measure the system performance against another system (the lower the better) by finding the equal(if not equal then at least nearly equal or have the min distance) between False Alarm Rate (FAR) and False Reject Rate (FRR, or missing rate) .
Refer to your data, [19.64,20] gives min distance, thus it could used as EER, you can take mean/max/min value of these two value, however since it means to compare between systems, thus make sure other system use the same method(mean/max/min) to pick EER value.
The difference among mean/max/min can be ignored if the there are large amount of data. In some speaker verification task, there will be 100k data sample.
2) To understand EER ,better compute it by yourself, here is how:
two things you need to know:
A) The system score for each test case (trial)
B) The true/false for each trial
After you have A and B, then you can create [trial, score,true/false] pairs then sort it by the score value, after that loop through the score, eg from min-> max. At each loop assume threshold is that score and compute the FAR,FRR. After loop through the score find the FAR,FRR with "equal" value.
For the code you can refer to my pyeer.py , in function processDataTable2
https://github.com/StevenLOL/Research_speech_speaker_verification_nist_sre2010/blob/master/SRE2010/sid/pyeer.py
This function is written for the NIST SRE 2010 evaluation.
4) There are other measures similar to EER, such as minDCF which only play with the weights of FAR and FRR. You can refer to "Performance Measure" of http://www.nist.gov/itl/iad/mig/sre10results.cfm
5) You can also refer to this package https://sites.google.com/site/bosaristoolkit/ and DETware_v2.1.tar.gz at http://www.itl.nist.gov/iad/mig/tools/ for computing and plotting EER in Matlab
Plotting in DETWare_v2.1
Pmiss=1:50;Pfa=50:-1:1;
Plot_DET(Pmiss/100.0,Pfa/100.0,'r')
FAR(t) and FRR(t) are parameterized by threshold, t. They are cumulative distributions, so they should be monotonic in t. Your data is not shown to be monotonic, so if it is indeed FAR and FRR, then the measurements were not made in order. But for the sake of clarity, we can order:
FAR FRR
1 1.65 98.33
2 1.82 84.31
3 3.96 60.56
4 7.92 42.92
5 13.2 29.17
6 15.35 26.39
7 16.83 23.47
8 17.99 21.39
9 19.14 20.28
10 19.64 20
11 21.29 18.61
12 24.92 17.08
13 26.07 16.39
14 29.04 13.13
15 34.49 9.31
16 40.76 6.81
17 50.33 5.42
18 66.83 1.67
19 82.51 0.28
This is for increasing FAR, which assumes a distance score; if you have a similarity score, then FAR would be sorted in decreasing order.
Loop over FAR until it is larger than FRR, which occurs at row 11. Then interpolate the cross over value between rows 10 and 11. This is your equal error rate.

MATLAB repeat numbers based on a vector of lengths

Is there a vectorised way to do the following? (shown by an example):
input_lengths = [ 1 1 1 4 3 2 1 ]
result = [ 1 2 3 4 4 4 4 5 5 5 6 6 7 ]
I have spaced out the input_lengths so it is easy to understand how the result is obtained
The resultant vector is of length: sum(lengths). I currently calculate result using the following loop:
result = ones(1, sum(input_lengths ));
counter = 1;
for i = 1:length(input_lengths)
start_index = counter;
end_index = counter + input_lengths (i) - 1;
result(start_index:end_index) = i;
counter = end_index + 1;
end
EDIT:
I can also do this using arrayfun (although that is not exactly a vectorised function)
cell_result = arrayfun(#(x) repmat(x, 1, input_lengths(x)), 1:length(input_lengths), 'UniformOutput', false);
cell_result : {[1], [2], [3], [4 4 4 4], [5 5 5], [6 6], [7]}
result = [cell_result{:}];
result : [ 1 2 3 4 4 4 4 5 5 5 6 6 7 ]
A fully vectorized version:
selector=bsxfun(#le,[1:max(input_lengths)]',input_lengths);
V=repmat([1:size(selector,2)],size(selector,1),1);
result=V(selector);
Downside is, the memory usage is O(numel(input_lengths)*max(input_lengths))
Benchmark of all solutions
Following the previous benchmark, I group all solutions given here in a script and run it a few hours for a benchmark. I've done this because I think it's good to see what is the performance of each proposed solution with the input lenght as parameter - my intention is not here to put down the quality of the previous one, which gives additional information about the effect of JIT. Moreover, and every participant seems to agree with that, quite a good work was done in all answers, so this great post deserves a conclusion post.
I won't post the code of the script here, this is quite long and very uninteresting. The procedure of the benchmark is to run each solution for a set of different lengths of input vectors: 10, 20, 50, 100, 200, 500, 1000, 2000, 5000, 10000, 20000, 50000, 100000, 200000, 500000, 1000000. For each input length, I've generated a random input vector based on Poisson law with parameter 0.8 (to avoid big values):
input_lengths = round(-log(1-rand(1,ILen(i)))/poisson_alpha)+1;
Finally, I average the computation times over 100 runs per input length.
I've run the script on my laptop computer (core I7) with Matlab R2013b; JIT is activated.
And here are the plotted results (sorry, color lines), in a log-log scale (x-axis: input length; y-axis: computation time in seconds):
So Luis Mendo is the clear winner, congrats!
For anyone who wants the numerical results and/or wants to replot them, here they are (cut the table into 2 parts and approximated to 3 digits, for a better display):
N 10 20 50 100 200 500 1e+03 2e+03
-------------------------------------------------------------------------------------------------------------
OP's for-loop 8.02e-05 0.000133 0.00029 0.00036 0.000581 0.00137 0.00248 0.00542
OP's arrayfun 0.00072 0.00117 0.00255 0.00326 0.00514 0.0124 0.0222 0.047
Daniel 0.000132 0.000132 0.000148 0.000118 0.000126 0.000325 0.000397 0.000651
Divakar 0.00012 0.000114 0.000132 0.000106 0.000115 0.000292 0.000367 0.000641
David's for-loop 9.15e-05 0.000149 0.000322 0.00041 0.000654 0.00157 0.00275 0.00622
David's arrayfun 0.00052 0.000761 0.00152 0.00188 0.0029 0.00689 0.0122 0.0272
Luis Mendo 4.15e-05 4.37e-05 4.66e-05 3.49e-05 3.36e-05 4.37e-05 5.87e-05 0.000108
Bentoy13's cumsum 0.000104 0.000107 0.000111 7.9e-05 7.19e-05 8.69e-05 0.000102 0.000165
Bentoy13's sparse 8.9e-05 8.82e-05 9.23e-05 6.78e-05 6.44e-05 8.61e-05 0.000114 0.0002
Luis Mendo's optim. 3.99e-05 3.96e-05 4.08e-05 4.3e-05 4.61e-05 5.86e-05 7.66e-05 0.000111
N 5e+03 1e+04 2e+04 5e+04 1e+05 2e+05 5e+05 1e+06
-------------------------------------------------------------------------------------------------------------
OP's for-loop 0.0138 0.0278 0.0588 0.16 0.264 0.525 1.35 2.73
OP's arrayfun 0.118 0.239 0.533 1.46 2.42 4.83 12.2 24.8
Daniel 0.00105 0.0021 0.00461 0.0138 0.0242 0.0504 0.126 0.264
Divakar 0.00127 0.00284 0.00655 0.0203 0.0335 0.0684 0.185 0.396
David's for-loop 0.015 0.0286 0.065 0.175 0.3 0.605 1.56 3.16
David's arrayfun 0.0668 0.129 0.299 0.803 1.33 2.64 6.76 13.6
Luis Mendo 0.000236 0.000446 0.000863 0.00221 0.0049 0.0118 0.0299 0.0637
Bentoy13's cumsum 0.000318 0.000638 0.00107 0.00261 0.00498 0.0114 0.0283 0.0526
Bentoy13's sparse 0.000414 0.000774 0.00148 0.00451 0.00814 0.0191 0.0441 0.0877
Luis Mendo's optim. 0.000224 0.000413 0.000754 0.00207 0.00353 0.00832 0.0216 0.0441
Ok, I've added another solution to the list ... I could not prevent myself to optimize the best-so-far solution of Luis Mendo. No credit for that, it's just a variant from Luis Mendo's, I'll explain it later.
Clearly, the solutions using arrayfun are very time-consuming. The solutions using an explicit for loop are faster, yet still slow compared with others solutions. So yes, vectorizing is still a major option for optimizing a Matlab script.
Since I've seen a big dispersion on the computing times of the fastest solutions, especially with input lengths between 100 and 10000, I decide to benchmark more precisely. So I've put the slowest apart (sorry), and redo the benchmark over the 6 other solutions which run much faster. The second benchmark over this reduced list of solutions is identical except that I've average over 1000 runs.
(No table here, unless you really want to, it's quite the same numbers as before)
As it was remarked, the solution by Daniel is a little faster than the one by Divakar because it seems that the use of bsxfun with #times is slower than using repmat. Still, they are 10 times faster than for-loop solutions: clearly, vectorizing in Matlab is a good thing.
The solutions of Bentoy13 and Luis Mendo are very close; the first one uses more instructions, but the second one uses an extra allocation when concatenating 1 to cumsum(input_lengths(1:end-1)). And that's why we see that Bentoy13's solution tends to be a bit faster with big input lengths (above 5.10^5), because there is no extra allocation. From this consideration, I've made an optimized solution where there is no extra allocation; here is the code (Luis Mendo can put this one in his answer if he wants to :) ):
result = zeros(1,sum(input_lengths));
result(1) = 1;
result(1+cumsum(input_lengths(1:end-1))) = 1;
result = cumsum(result);
Any comment for improvement is welcome.
More of a comment than anything, but I did some tests. I tried a for loop, and an arrayfun, and I tested your for loop and arrayfun version. Your for loop was the fastest. I think this is because it is simple, and allows the JIT compilation to do the most optimisation. I am using Matlab, octave might be different.
And the timing:
Solution: With JIT Without JIT
Sam for 0.74 1.22
Sam arrayfun 2.85 2.85
My for 0.62 2.57
My arrayfun 1.27 3.81
Divakar 0.26 0.28
Bentoy 0.07 0.06
Daniel 0.15 0.16
Luis Mendo 0.07 0.06
So Bentoy's code is really fast, and Luis Mendo's is almost exactly the same speed. And I rely on JIT way too much!
And the code for my attempts
clc,clear
input_lengths = randi(20,[1 10000]);
% My for loop
tic()
C=cumsum(input_lengths);
D=diff(C);
results=zeros(1,C(end));
results(1,1:C(1))=1;
for i=2:length(input_lengths)
results(1,C(i-1)+1:C(i))=i*ones(1,D(i-1));
end
toc()
tic()
A=arrayfun(#(i) i*ones(1,input_lengths(i)),1:length(input_lengths),'UniformOutput',false);
R=[A{:}];
toc()
result = zeros(1,sum(input_lengths));
result(cumsum([1 input_lengths(1:end-1)])) = 1;
result = cumsum(result);
This should be pretty fast. And memory usage is the minimum possible.
An optimized version of the above code, due to Bentoy13 (see his very detailed benchmarking):
result = zeros(1,sum(input_lengths));
result(1) = 1;
result(1+cumsum(input_lengths(1:end-1))) = 1;
result = cumsum(result);
This is a slight variant of #Daniel's answer. The crux of this solution is based on that solution. Now this one avoids repmat, so in that way it's little-more "vectorized" maybe. Here's the code -
selector=bsxfun(#le,[1:max(input_lengths)]',input_lengths); %//'
V = bsxfun(#times,selector,1:numel(input_lengths));
result = V(V~=0)
For all the desperate one-liner searching people -
result = nonzeros(bsxfun(#times,bsxfun(#le,[1:max(input_lengths)]',input_lengths),1:numel(input_lengths)))
I search an elegant solution, and I think David's solution is a good start. What I have in mind is that one can generate the indexes where to add one from previous element.
For that, if we compute the cumsum of the input vector, we get:
cumsum(input_lengths)
ans = 1 2 3 7 10 12 13
This is the indexes of the ends of sequences of identical numbers. That is not what we want, so we flip the vector twice to get the beginnings:
fliplr(sum(input_lengths)+1-cumsum(fliplr(input_lengths)))
ans = 1 2 3 4 8 11 13
Here is the trick. You flip the vector, cumsum it to get the ends of the flipped vector, and then flip back; but you must substract the vector from the total length of the output vector (+1 because index starts at 1) because cumsum applies on the flipped vector.
Once you have done this, it's very straightforward, you just have to put 1 at computed indexes and 0 elsewhere, and cumsum it:
idx_begs = fliplr(sum(input_lengths)+1-cumsum(fliplr(input_lengths)));
result = zeros(1,sum(input_lengths));
result(idx_begs) = 1;
result = cumsum(result);
EDIT
First, please have a look at Luis Mendo's solution, it is very close to mine but is more simpler and a bit faster (I won't edit mine even it is very close). I think at this date this is the fastest solution from all.
Second, while looking at others solutions, I've made up another one-liner, a little different from my initial solution and from the other one-liner. Ok, this won't be very readable, so take a breath:
result = cumsum( full(sparse(cumsum([1,input_lengths(1:end-1)]), ...
ones(1,length(input_lengths)), 1, sum(input_lengths),1)) );
I cut it on two lines. Ok now let's explain it.
The similar part is to build the array of the indexes where to increment the value of the current element. I use the solution of Luis Mendo's for that. To build in one line the solution vector, I use here the fact that it is in fact a sparse representation of the binary vector, the one we will cumsum at the very end. This sparse vector is build using our computed index vector as x positions, a vector of 1 as y positions, and 1 as the value to put at these locations. A fourth argument is given to precise the total size of the vector (important if the last element of input_lengths is not 1). Then we get the full representation of this sparse vector (else the result is a sparse vector with no empty element) and we can cumsum.
There is no use of this solution other than to give another solution to this problem. A benchmark can show that it is slower than my original solution, because of a heavier memory load.