Skip to content
New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Jump to bottom

mnistAutoencoder #189

Open
wants to merge 1 commit into
base: master
Choose a base branch
from
Open

mnistAutoencoder #189

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
1 commit
Select commit
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
279 changes: 279 additions & 0 deletions examples/mnistAutoencoder/cnn_mnist_autoencoder.m
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,279 @@
function [net, opts, imdb, info] = cnn_mnist_autoencoder
%CNN_MNIST_AUTOENCODER Summary of this function goes here
% Detailed explanation goes here

net = getMnistAutoencoderNet;
opts = getMnistAutoencoderOpts;

if exist(opts.imdbPath, 'file')

load(opts.imdbPath);

else

imdb = getMnistAutoencoderImdb(opts);

if ~exist(opts.expDir, 'dir')

mkdir(opts.expDir);

end

save(opts.imdbPath, 'imdb');

end

% [net, info] = cnn_train(net, imdb, @(imdb, batch) getMnistAutoencoderBatch(imdb, batch), opts);
[net, info] = cnn_train_adagrad(net, imdb, @(imdb, batch) getMnistAutoencoderBatch(imdb, batch), opts);

net.layers{end} = struct('name', 'data_hat_sigmoid', ...
'type', 'sigmoid' );

net.layers{end + 1} = struct('type', 'euclideanloss');

end

% -------------------------------------------------------------------------
function net = getMnistAutoencoderNet
% -------------------------------------------------------------------------

% Layer 1

net.layers{1} = struct('biases' , zeros(1, 1000, 'single') , ...
'biasesLearningRate' , 1 , ...
'biasesWeightDecay' , 0 , ...
'filters' , sparse_initialization([1 1 784 1000]), ...
'filtersLearningRate', 1 , ...
'filtersWeightDecay' , 1 , ...
'name' , 'encoder_1' , ...
'pad' , [0 0 0 0] , ...
'stride' , [1 1] , ...
'type' , 'conv' );

net.layers{2} = struct('name', 'encoder_1_sigmoid', ...
'type', 'sigmoid' );

% Layer 2

net.layers{3} = struct('biases' , zeros(1, 500, 'single') , ...
'biasesLearningRate' , 1 , ...
'biasesWeightDecay' , 0 , ...
'filters' , sparse_initialization([1 1 1000 500]), ...
'filtersLearningRate', 1 , ...
'filtersWeightDecay' , 1 , ...
'name' , 'encoder_2' , ...
'pad' , [0 0 0 0] , ...
'stride' , [1 1] , ...
'type' , 'conv' );

net.layers{4} = struct('name', 'encoder_2_sigmoid', ...
'type', 'sigmoid' );

% Layer 3

net.layers{5} = struct('biases' , zeros(1, 250, 'single') , ...
'biasesLearningRate' , 1 , ...
'biasesWeightDecay' , 0 , ...
'filters' , sparse_initialization([1 1 500 250]), ...
'filtersLearningRate', 1 , ...
'filtersWeightDecay' , 1 , ...
'name' , 'encoder_3' , ...
'pad' , [0 0 0 0] , ...
'stride' , [1 1] , ...
'type' , 'conv' );

net.layers{6} = struct('name', 'encoder_3_sigmoid', ...
'type', 'sigmoid' );

% Layer 4

net.layers{5} = struct('biases' , zeros(1, 30, 'single') , ...
'biasesLearningRate' , 1 , ...
'biasesWeightDecay' , 0 , ...
'filters' , sparse_initialization([1 1 250 30]), ...
'filtersLearningRate', 1 , ...
'filtersWeightDecay' , 1 , ...
'name' , 'code' , ...
'pad' , [0 0 0 0] , ...
'stride' , [1 1] , ...
'type' , 'conv' );

% Layer 5

net.layers{6} = struct('biases' , zeros(1, 250, 'single') , ...
'biasesLearningRate' , 1 , ...
'biasesWeightDecay' , 0 , ...
'filters' , sparse_initialization([1 1 30 250]), ...
'filtersLearningRate', 1 , ...
'filtersWeightDecay' , 1 , ...
'name' , 'decoder_3' , ...
'pad' , [0 0 0 0] , ...
'stride' , [1 1] , ...
'type' , 'conv' );

net.layers{7} = struct('name', 'decoder_3_sigmoid', ...
'type', 'sigmoid' );

% Layer 6

net.layers{8} = struct('biases' , zeros(1, 500, 'single') , ...
'biasesLearningRate' , 1 , ...
'biasesWeightDecay' , 0 , ...
'filters' , sparse_initialization([1 1 250 500]), ...
'filtersLearningRate', 1 , ...
'filtersWeightDecay' , 1 , ...
'name' , 'decoder_2' , ...
'pad' , [0 0 0 0] , ...
'stride' , [1 1] , ...
'type' , 'conv' );

net.layers{9} = struct('name', 'decoder_2_sigmoid', ...
'type', 'sigmoid' );

% Layer 7

net.layers{10} = struct('biases' , zeros(1, 1000, 'single') , ...
'biasesLearningRate' , 1 , ...
'biasesWeightDecay' , 0 , ...
'filters' , sparse_initialization([1 1 500 1000]), ...
'filtersLearningRate', 1 , ...
'filtersWeightDecay' , 1 , ...
'name' , 'decoder_1' , ...
'pad' , [0 0 0 0] , ...
'stride' , [1 1] , ...
'type' , 'conv' );

net.layers{11} = struct('name', 'decoder_1_sigmoid', ...
'type', 'sigmoid' );

% Layer 8

net.layers{12} = struct('biases' , zeros(1, 784, 'single') , ...
'biasesLearningRate' , 1 , ...
'biasesWeightDecay' , 0 , ...
'filters' , sparse_initialization([1 1 1000 784]), ...
'filtersLearningRate', 1 , ...
'filtersWeightDecay' , 1 , ...
'name' , 'data_hat' , ...
'pad' , [0 0 0 0] , ...
'stride' , [1 1] , ...
'type' , 'conv' );

net.layers{13} = struct('type', 'sigmoidcrossentropyloss');

vl_simplenn_display(net);

end

% -------------------------------------------------------------------------
function filters = sparse_initialization(d)
% -------------------------------------------------------------------------

filters = zeros(d, 'single');

for index = 1 : d(4)

p = randperm(d(3), 15);

filters(1, 1, p, index) = randn(1, 1, 15, 1);

end

end

% -------------------------------------------------------------------------
function opts = getMnistAutoencoderOpts
% -------------------------------------------------------------------------

opts.batchSize = 100;
opts.conserveMemory = false;
opts.continue = false;
opts.dataDir = fullfile('data','mnist');
opts.display = 10;
opts.delta = 1e-8;
opts.errorType = 'euclideanloss';
opts.expDir = fullfile('data','mnistAutoencoder');
opts.imdbPath = fullfile(opts.expDir, 'imdb.mat');
% opts.learningRate = 1e-4;
opts.learningRate = 1e-2;
% opts.momentum = 0.9;
% opts.numEpochs = 6667; % 6667 epochs is ~4000000 iterations.
opts.numEpochs = 108; % 108 epochs is ~65000 iterations.
opts.plotDiagnostics = false;
opts.prefetch = false;
opts.snapshot = 10;
opts.sync = true;
opts.test_interval = 10;
opts.train = [];
opts.useGpu = true;
opts.val = [];
opts.weightDecay = 5e-4;

end

% -------------------------------------------------------------------------
function imdb = getMnistAutoencoderImdb(opts)
% -------------------------------------------------------------------------
% Preapre the imdb structure, returns image data with mean image subtracted
files = {'train-images-idx3-ubyte', ...
'train-labels-idx1-ubyte', ...
't10k-images-idx3-ubyte', ...
't10k-labels-idx1-ubyte'} ;

if ~exist(opts.dataDir, 'dir')
mkdir(opts.dataDir) ;
end

for i=1:4
if ~exist(fullfile(opts.dataDir, files{i}), 'file')
url = sprintf('http://yann.lecun.com/exdb/mnist/%s.gz',files{i}) ;
fprintf('downloading %s\n', url) ;
gunzip(url, opts.dataDir) ;
end
end

f=fopen(fullfile(opts.dataDir, 'train-images-idx3-ubyte'),'r') ;
x1=fread(f,inf,'uint8');
fclose(f) ;
x1=permute(reshape(x1(17:end),28,28,60e3),[2 1 3]) ;

f=fopen(fullfile(opts.dataDir, 't10k-images-idx3-ubyte'),'r') ;
x2=fread(f,inf,'uint8');
fclose(f) ;
x2=permute(reshape(x2(17:end),28,28,10e3),[2 1 3]) ;

f=fopen(fullfile(opts.dataDir, 'train-labels-idx1-ubyte'),'r') ;
y1=fread(f,inf,'uint8');
fclose(f) ;
y1=double(y1(9:end)')+1 ;

f=fopen(fullfile(opts.dataDir, 't10k-labels-idx1-ubyte'),'r') ;
y2=fread(f,inf,'uint8');
fclose(f) ;
y2=double(y2(9:end)')+1 ;

set = [ones(1,numel(y1)) 2*ones(1,numel(y2))];
% data = single(reshape(cat(3, x1, x2),28,28,1,[]));
% dataMean = mean(data(:,:,:,set == 1), 4);
% data = bsxfun(@minus, data, dataMean) ;
data = single(reshape(cat(3, x1, x2), 1, 1, 784, []));
data = data - min(data(:)); data = data / max(data(:));

imdb.images.data = data ;
% imdb.images.data_mean = dataMean;
imdb.images.labels = cat(2, y1, y2) ;
imdb.images.set = set ;
imdb.meta.sets = {'train', 'val', 'test'} ;
imdb.meta.classes = arrayfun(@(x)sprintf('%d',x),0:9,'uniformoutput',false) ;

end

% -------------------------------------------------------------------------
function [im, labels] = getMnistAutoencoderBatch(imdb, batch)
% -------------------------------------------------------------------------

im = imdb.images.data(:, :, :, batch);
labels = im;

end

76 changes: 76 additions & 0 deletions examples/mnistAutoencoder/cnn_mnist_autoencoder_test_demo.m
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,76 @@
%%

close all;
clear all;
clc;

%%

run('~/GitHub/umuguc/matconvnet/matlab/vl_setupnn');

%%

load('net.mat');
load(opts.imdbPath);

%%

N = [5 2];

Y = zeros(N(1) * N(2), 1);

h = figure;

for index = 1 : N(1) * N(2)

im = imdb.images.data(:, :, :, end - index + 1);

if opts.useGpu

im = gpuArray(im);

end

subplot(N(1), 2 * N(2), 2 * index - 1);

imagesc(reshape(im, 28, 28));

axis off;
axis square;

drawnow;

net.layers{end}.class = im;

res = vl_simplenn(net, im, [], [], 'disableDropout', true);

subplot(N(1), 2 * N(2), 2 * index);

imagesc(reshape(res(end - 1).x, 28, 28));

axis off;
axis square;

drawnow;

Y(index) = gather(res(end).x);

end

disp(['Euclidean loss: ' num2str(mean(Y))]);

%%

% Test net:

% layer| 1| 2| 3| 4| 5| 6| 7| 8| 9| 10| 11| 12| 13| 14|
% type| cnv|sigmoid| cnv|sigmoid| cnv| cnv|sigmoid| cnv|sigmoid| cnv|sigmoid| cnv|sigmoid|euclideanloss|
% support| 1x1| 1x1| 1x1| 1x1| 1x1| 1x1| 1x1| 1x1| 1x1| 1x1| 1x1| 1x1| 1x1| 1x1|
% stride| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1|
% pad| 0| 0| 0| 0| 0| 0| 0| 0| 0| 0| 0| 0| 0| 0|
% out dim| 1000| 1000| 500| 500| 30| 250| 250| 500| 500| 1000| 1000| 784| 784| 784|
% filt dim| 784| n/a| 1000| n/a| 250| 30| n/a| 250| n/a| 500| n/a| 1000| n/a| n/a|
% rec. field| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1|
% c/g net KB| 3066/0| 0/0| 1955/0| 0/0| 29/0| 30/0| 0/0| 490/0| 0/0| 1957/0| 0/0| 3066/0| 0/0| 0/0|
% total network CPU/GPU memory: 10.3/0 MB

33 changes: 33 additions & 0 deletions examples/mnistAutoencoder/cnn_mnist_autoencoder_training_demo.m
View file
Open in desktop
Original file line number Diff line number Diff line change
@@ -0,0 +1,33 @@
%%

close all;
clear all;
clc;

%%

run('~/GitHub/umuguc/matconvnet/matlab/vl_setupnn');

%%

rng(0);

[net, opts, imdb, info] = cnn_mnist_autoencoder;

save('net.mat', 'net', 'opts', 'info');

%%

% Training net:

% layer| 1| 2| 3| 4| 5| 6| 7| 8| 9| 10| 11| 12| 13|
% type| cnv|sigmoid| cnv|sigmoid| cnv| cnv|sigmoid| cnv|sigmoid| cnv|sigmoid| cnv|sigmoidcrossentropyloss|
% support| 1x1| 1x1| 1x1| 1x1| 1x1| 1x1| 1x1| 1x1| 1x1| 1x1| 1x1| 1x1| 1x1|
% stride| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1|
% pad| 0| 0| 0| 0| 0| 0| 0| 0| 0| 0| 0| 0| 0|
% out dim| 1000| 1000| 500| 500| 30| 250| 250| 500| 500| 1000| 1000| 784| 784|
% filt dim| 784| n/a| 1000| n/a| 250| 30| n/a| 250| n/a| 500| n/a| 1000| n/a|
% rec. field| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1| 1|
% c/g net KB| 3066/0| 0/0| 1955/0| 0/0| 29/0| 30/0| 0/0| 490/0| 0/0| 1957/0| 0/0| 3066/0| 0/0|
% total network CPU/GPU memory: 10.3/0 MB

Loading