Re: Optimisation / parallelisation Problems

[Riccardo Corradini]

Hi,

I would suggest you to use openmpi_ext on octave-forge.

It works on multicore computers.

Bests

---

Da: somealias <address@hidden>
A: address@hidden
Inviato: Venerdì 20 Aprile 2012 7:43
Oggetto: Optimisation / parallelisation Problems

Hi all,

I'm running some machine learning experiments with Octave. I have optimised
(e.g. used sparse matrices, pre-allocated etc.) and vectorised most of my
code but want to improve the performance of some loops. I have tried
vectorising and parallelising (pararrayfun) these loops but without success.
I need some help!

Here's an example of what I'm trying to do.

---
X = rand(10,20); % examples
y = rand(10,1); % labels
[m, n] = size(X); % # examples and dimensions
increment = 1; % # of examples incremented per iteration
iterations = 5; % # of times we randomly pick the examples

% Train the model incrementally
% e.g. increment = 1, train on 1 example, then 2, then 3 etc...
for i = increment:increment:m
    % Initialise running totals for the current # of examples 
    errorTrainTotal = 0;
    errorValTotal = 0;
   
    % randomly select the examples used for training
    % e.g. iterations = 5, then we do this 5 times
    for j = 1:iterations
        % Create matrices of randomly selected examples from the training set
        shuffle = randperm(m);

        % Initialise the new random training matrices
        X_random = X(shuffle(1:i),:);
        y_random = y(shuffle(1:i));

        % Learn the model parameters
        theta = trainLogisticReg(X_random, y_random, lambda);

        % Predict the class of the training examples based on the theta
        predictions = predict(theta, X_random);
        % Predict the class of the validation examples based on the theta
        predictionsval = predict(theta, Xval);

        % Compute training and validation f-score errors
        errorTrainTotal = errorTrainTotal + (1 - fscore(predictions,
y_random));
        errorValTotal = errorValTotal + (1 - fscore(predictionsval, yval));
    end

    % Calculate the training and validation errors for the current # of
examples
    errorTrain(i/increment) = errorTrainTotal / iterations;
    errorVal(i/increment) = errorValTotal / iterations;
end
---

The above code becomes slow for the large number of examples (m) in my real
dataset. I thought the best way to improve performance would be to
parallelise these loops because my other CPU cores are idle while each loop
being processed in a serial manner.

Any suggestions and comments would be greatly appreciated!

Kev

--
View this message in context: http://octave.1599824.n4.nabble.com/Optimisation-parallelisation-Problems-tp4572922p4572922.html
Sent from the Octave - General mailing list archive at Nabble.com.
_______________________________________________
Help-octave mailing list
address@hidden
https://mailman.cae.wisc.edu/listinfo/help-octave