coding structure of gaussian noise layer

I want to create a special layer to add some special noise to the data.

But my matlab version is 2017b, I don't have the example " gaussianNoiseLayer.m".

That file should be located at (matlabroot, 'examples', 'nnet', 'main', 'gaussianNoiseLayer.m') in the matlab 2018b or 2018a version.

I really want to know the coding structure of adding noise layer.

If any kind-hearted person has installed the latest version of matlab, can you send a copy of this file to me?

email: [e-mail] thank you very much！！

[MAHSA YOUSEFI]

Hi Jian.

Did you solve your problem with adding noise?

I want to add Gaussian noide per each layer of hidden layer and input in my costumized training loop.

Answer 1

[Jack Xiao]

here is the code:

classdef gaussianNoiseLayer < nnet.layer.Layer

% gaussianNoiseLayer Gaussian noise layer

% A Gaussian noise layer adds random Gaussian noise to the input.

%

% To create a Gaussian noise layer, use

% layer = gaussianNoiseLayer(sigma, name)

properties

% Standard deviation.

Sigma

end

methods

function layer = gaussianNoiseLayer(sigma, name)

% layer = gaussianNoiseLayer(sigma,name) creates a Gaussian

% noise layer and specifies the standard deviation and layer

% name.

layer.Name = name;

layer.Description = ...

"Gaussian noise with standard deviation " + sigma;

layer.Type = "Gaussian Noise";

layer.Sigma = sigma;

end

function Z = predict(layer, X)

% Z = predict(layer, X) forwards the input data X through the

% layer for prediction and outputs the result Z.

% At prediction time, the output is equal to the input.

Z = X;

end

function [Z, memory] = forward(layer, X)

% Z = forward(layer, X) forwards the input data X through the

% layer and outputs the result Z.

% At training time, the layer adds Gaussian noise to the input.

sigma = layer.Sigma;

noise = randn(size(X)) * sigma;

Z = X + noise;

memory = [];

end

function dLdX = backward(layer, X, Z, dLdZ, memory)

% [dLdX, dLdAlpha] = backward(layer, X, Z, dLdZ, memory)

% backward propagates the derivative of the loss function

% through the layer.

% Since the layer adds a random constant, the derivative dLdX

% is equal to dLdZ.

dLdX = dLdZ;

end

end

end