problem with tbl,xvar,yvar form of plot()

Greetings, as per title i have a problem with tbl,xvar,yvar form of plot().

I have an array with history of the evolution of the parameters of a function during parameter optimization.

num_iterations = size(param_history, 1);
num_chains = size(param_history, 2);
num_dimensions = size(param_history, 3);
param_at_step = zeros(num_iterations, num_chains * num_dimensions);

The array is of the form 10000x8 - 10k iterations by 2 parameters for each of the 4 independent optimization chains. I expected:

for iteration = 1 : num_iterations
    for chain_id = 1 : num_chains
        for param = 1 : num_dimensions
            column = ((chain_id-1) * 2) + param;
            param_at_step(iteration, column) = param_history(iteration, chain_id, param);
        end
    end
end
plot( param_at_step, [1 3 5 7], [2 4 6 8] );

to plot the evolution of the parameters in the parameter space, but i get an error:

Error using plot

Specify the coordinates as vectors or matrices of the same size, or as a vector and a matrix that share the same length in at least one dimension.

I am confused, what am I doing wrong?

[Voss]

The code involving three nested for loops for constructing param_at_step from param_history

param_history = randn(10000,4,2); % random data

num_iterations = size(param_history, 1);
num_chains = size(param_history, 2);
num_dimensions = size(param_history, 3);
param_at_step = zeros(num_iterations, num_chains * num_dimensions);

for iteration = 1 : num_iterations
    for chain_id = 1 : num_chains
        for param = 1 : num_dimensions
            column = ((chain_id-1) * 2) + param;
            param_at_step(iteration, column) = param_history(iteration, chain_id, param);
        end
    end
end

can be more succinctly written using reshape and permute:

num_iterations = size(param_history, 1);
param_at_step_test = reshape(permute(param_history,[1 3 2]),num_iterations,[]);

And the result is the same:

isequal(param_at_step_test,param_at_step)

ans = logical
   1

Answer 1

[Voss]

what am I doing wrong?

param_at_step is a numeric matrix, not a table, so you are not using the tbl,xvar,yvar form of plot().

[Voss]

param_history = randn(10000,4,2); % random data

num_iterations = size(param_history, 1);
num_chains = size(param_history, 2);
num_dimensions = size(param_history, 3);
param_at_step = zeros(num_iterations, num_chains * num_dimensions);

for iteration = 1 : num_iterations
    for chain_id = 1 : num_chains
        for param = 1 : num_dimensions
            column = ((chain_id-1) * 2) + param;
            param_at_step(iteration, column) = param_history(iteration, chain_id, param);
        end
    end
end

To get the plot you want, you can either convert param_at_step into a table and use the tble,xvar,yvar syntax in plot:

figure
plot( array2table(param_at_step), [1 3 5 7], [2 4 6 8] );

Or, more simply, just plot the columns of param_at_step as it is:

figure
plot( param_at_step(:,[1 3 5 7]), param_at_step(:,[2 4 6 8]) );

[Maciej]

I'll accept this, it answers the question just as it is written.

I have a follow up-however. What if the number of chains and the number of dimensions (aka parameters) are variables that can be configured at the start of the run (lets limit ourselves to 3 dimensions max)?

For e.g. what if i have 6 chains operating in 3 dimensional space?

I could use simple if-elseif-end to switch between plot and plot3 but not sure how to generate those (:,[numbers]) expressions programmaticaly : O

[Voss]

In my opinion, it's easier to plot from the 3D array param_history rather than rearranging it into the matrix param_at_step first.

param_history = randn(10000,6,3); % random data

[num_iterations,num_chains,num_dimensions] = size(param_history);

Here's one simple way:

figure
if num_dimensions > 2
    plot3(param_history(:,:,1),param_history(:,:,2),param_history(:,:,3))
else
    plot(param_history(:,:,1),param_history(:,:,2))
end

Here's another way:

if num_dimensions > 2
    plot_func = @plot3;
else
    plot_func = @plot;
end
plot_args = permute(num2cell(param_history,1),[3 2 1]);
figure
plot_func(plot_args{:})

If you need to use param_at_step for whatever reason, then here's how you can do it, generating the vectors of column numbers programmatically:

param_at_step = reshape(permute(param_history,[1 3 2]),num_iterations,[]);

figure
if num_dimensions > 2
    plot3(param_at_step(:,1:num_dimensions:end),param_at_step(:,2:num_dimensions:end),param_at_step(:,3:num_dimensions:end))
else
    plot(param_at_step(:,1:num_dimensions:end),param_at_step(:,2:num_dimensions:end))
end