Stacked Denoising Auto-Encoder of DeepLearnToolbox (https://github.com/rasmusbergpalm/DeepLearnToolbox) - a simple example

trySAE.matlab

% Playing with the Stacked Denoising Auto-Encoder in the DeepLearnToolbox 
% (https://github.com/rasmusbergpalm/DeepLearnToolbox), 
% and trying to use it to extract features.

%% Loads data
load mnist_uint8;

% I don't have much memory
clear test_x test_y

% The original MNIST that comes along with DeepLearnToolbox 
% contains 60000 samples, which will cause out-of-memory error
% on my old computer. 

% Fortunately, the data of 10 numbers are distributed uniformly.
train_x = train_x(1:10000, :);
train_y = train_y(1:10000, :);

train_x = double(train_x) / 255;
train_y = double(train_y);


%% SAE setup & training
rand('twister', 5489);  % Matlab's default setting
sae = saesetup([784 200 100]);
nAutoEncoders = numel(sae.ae);

for i = 1:nAutoEncoders
	% Why does the author mix camelCase and under_score? :S
	sae.ae{i}.activation_function = 'sigm';
	sae.ae{i}.learningRate = 1;
	sae.ae{i}.inputZeroMaskedFraction = 0.5;
end

opts.numepochs = 1;
opts.batchsize = 100;

sae = saetrain(sae, train_x, opts);


%% Output

% In the `test_example_SAE` demo, the learned weights are feed into an FFNN.
% However here we just use the SAE directly.
fprintf('Please enter the index of data:\n');
while true
	index = input('#');
	if isempty(index)
		break;
	end
	x = train_x(index, :)';
	x0 = x;  % backup

	% The first half
	for i = 1:nAutoEncoders
		activ = str2func(sae.ae{i}.activation_function);
		x = activ(sae.ae{i}.W{1} * [1; x]);
	end

	% Now `x` is the extracted feature vector.

	% The second half
	for i = 1:nAutoEncoders
		activ = str2func(sae.ae{i}.activation_function);
		x = activ(sae.ae{nAutoEncoders + 1 - i}.W{2} * [1; x]);
	end
	
	% Visualizes input/output.
	% Theoretically, they should look similar.
	fprintf('Number = %d\n', find(train_y(index, :)) - 1);
	subplot(1, 2, 1);
	imshow(reshape(x0, [28 28])');
	subplot(1, 2, 2);
	imshow(reshape(x, [28 28])');
end