How to Manually Change dlnet's Learnable Parameters?

Hi,

I want to add some minor noise to a dlnet.Learnables.

Firstly, I tried to convert each learnables from table to cell to a single dlarray by

yy = cell2mat(table2array(dlnet.Learnables(1, 3)));
zz = yy + normrnd(0, 0.005, size(yy));

But I found I could not manage to convert zz back and have it stored in dlnet.Learnables(1, 3).

Then, I tried to use dlupdate() function like

Noise_Adder = @(para, Noise_Sigma) para + normrnd(0, Noise_Sigma, size(para));
G_Add_Noise_Sigma = 0.005;
for l = 1:size(dlnet.Learnables, 1)
    dlnet.Learnables(i, 3) = dlupdate(Noise_Adder, dlnet.Learnables(i, 3), G_Add_Noise_Sigma);
end

Yet I got an error saying

Dot indexing is not supported for variables of this
type.

Error in
deep.internal.recording.containerfeval>iProcessTable_Nout_Nin
(line 262)
    varargvals{input} = varargin{input}.Value;

Error in
deep.internal.recording.containerfeval>iDispatch_Nout_Nin
(line 194)
    [varargout{1:nargout}]  =
    iProcessTable_Nout_Nin(allowNetInput,
    makeNetOutput, fun, paramFun, c, varargin{:});

Error in deep.internal.recording.containerfeval
(line 41)
        [c, varargout{1:nargout-1}] =
        iDispatch_Nout_Nin(allowNetInput,
        makeNetOutput, fun, paramFun, c,
        varargin{:});

Error in deep.internal.networkContainerFun (line
15)
    [c, varargout{1:nargout-1}] =
    deep.internal.recording.containerfeval(...

Error in dlupdate (line 102)
    [P, varargout{1:nargout-1}] =
    deep.internal.networkContainerFun(...

So how can I achieve my goal? Many thanks.

[Theron FARRELL]

Actually, I managed to implement what I want manually by adding noise layer by layer

G_Para_Noise_Sigma = 0.05;
[dlnet_G.Learnables] = Add_Para_Noise(dlnet_G.Learnables, G_Para_Noise_Sigma);

function [para] = Add_Para_Noise(para, sigma)
for layer = 1:size(para, 1)
    para.Value{layer} = para.Value{layer} + normrnd(0, sigma, size(para.Value{layer}));
end
end

But I know it is quite naive and inefficient. Is there any better way to do it, especially by using dlupdate?

Thanks.

Answer 1

[Gautam Pendse]

Hi Theron,

The first syntax for dlupdate may be what you are looking for:

https://www.mathworks.com/help/deeplearning/ref/dlupdate.html

% Update learnable parameters in dlnet by adding 1
myfun = @(x) x + 1;
dlnet = dlupdate(myfun,dlnet);

Hope this helps,

Gautam