How to create a non learnable layer for upsampling in Matlab?

Hello Matlab-Community,

I am currently trying to deploy a UNET model created and trained in Keras in Matlab using the function importKerasLayers.

However, in the Matlab version I am using (R2018b) there is no UpSampling2D (see https://keras.io/layers/convolutional/) layer available. Thus, these layers are replaced by a placeholderLayer.

The function replaceLayer can replace such placeholderLayer's in the network graph by any priorly defined layers. As far as I can tell, Matlab does not offer such a simple upsampling layer which does repeat the existing values (without interpolation and without learning anything).

Does anybody know, how to define an easy upsampling layer (as done in Keras by UpSampling2D(size = (2,2))) and include it into the network graph?

Thanks in advance for your help.

Best regards

Adidue

Answer 1

[Hazar Benan Unal]

Since I myself had a hard time figuring out the implementation of upsampling layer in MATLAB, below is the code to do 3D and 2D upsampling in case someone needs in future. Hopefully 2020 version will have this commonly used layer.

%3D Upsampling layer.
%%Example: Input: 6x8x10x7 ---> Output: 12x16x20x7

%Number of filters needs to be the same as number of channels
%in the previous layer.
num_filters = 7

%Set learning rate factors to zero since upsampling layer does not learn any parameters.
upsample = transposedConv3dLayer([2,2,2],num_filters,'Stride',[2,2,2],...
    'WeightLearnRateFactor',0,'BiasLearnRateFactor',0, 'Name', 'upsampling_3D');
upsample.Weights = zeros(2,2,2,num_filters,num_filters);
upsample.Bias = zeros(1,1,1,num_filters);

%First filter should act only on the first channel in the previous layer.
%Second filter should act only on the second channel in the previous layer and so on.
%So we need to 'turn on' only the corresponding filter for the channel of interest.
for j = 1:num_filters
    upsample.Weights(:,:,:,j,j) = 1;
end

%%If you don't do this for loop and simply set all weights to 1 with,

%upsample.Weights = ones(2,2,2,nf1,nf1);
%upsample.Bias = zeros(1,1,1,nf1);

%Then each filter acts on ALL of the channels in the previous layer and gives you wrong result. 

Similarly, for 2D Upsampling:

%2D Upsampling layer.
%%Example: Input: 6x8x7 ---> Output: 12x16x7

%Number of filters needs to be the same as number of channels
%in the previous layer.
num_filters = 7

%Set learning rate factors to zero since upsampling layer does not learn any parameters.
upsample = transposedConv2dLayer([2,2],num_filters,'Stride',[2,2],...
    'WeightLearnRateFactor',0,'BiasLearnRateFactor',0, 'Name', 'upsampling_2D');
upsample.Weights = zeros(2,2,num_filters,num_filters);
upsample.Bias = zeros(1,1,nf1);

%First filter should act only on the first channel in the previous layer.
%Second filter should act only on the second channel in the previous layer and so on.
%So we need to 'turn on' only the corresponding filter for the channel of interest.
for j = 1:num_filters
    upsample.Weights(:,:,j,j) = 1;
end

And a toy example for visual verification:

%3D toy example
input_size = [6,6,4,3];

num_filters = input_size(4); 
upsample = transposedConv3dLayer([2,2,2],num_filters,'Stride',[2,2,2],...
    'WeightLearnRateFactor',0,'BiasLearnRateFactor',0, 'Name', 'ups_3d');
upsample.Weights = zeros(2,2,2,num_filters,num_filters);
upsample.Bias = zeros(1,1,1,num_filters);

for j = 1:num_filters
    upsample.Weights(:,:,:,j,j) = 1;
end

layers = [image3dInputLayer(input_size, 'Name', 'input'), 
          upsample,
          regressionLayer('Name', 'reg')];
layers(1).Mean = 0; %zerocenter input

lgraph = layerGraph(layers);
net = assembleNetwork(lgraph);

input_image = rand(input_size);
size(input_image)   %Size before upsampling
predict(net,input_image);

act = activations(net, input_image, 'ups_3d');
size(act)  %Size after upsampling

figure
subplot(3,2,1)
imshow(input_image(:,:,1,1),[])
title('Input(6x6)')
subplot(3,2,3)
imshow(input_image(:,:,2,1),[])
subplot(3,2,5)
imshow(input_image(:,:,3,1),[])

subplot(3,2,2)
imshow(act(:,:,2,1),[])
title('Upsampled(12x12)')
subplot(3,2,4)
imshow(act(:,:,4,1),[])
subplot(3,2,6)
imshow(act(:,:,6,1),[])

%Images on the same row should be the same, while the ones on the right are higher resolution.

2D case:

%2D toy example
input_size = [6,6,3];

num_filters = input_size(3); 
upsample = transposedConv2dLayer([2,2],num_filters,'Stride',[2,2],...
    'WeightLearnRateFactor',0,'BiasLearnRateFactor',0, 'Name', 'ups_2d');
upsample.Weights = zeros(2,2,num_filters,num_filters);
upsample.Bias = zeros(1,1,num_filters);

for j = 1:num_filters
    upsample.Weights(:,:,j,j) = 1;
end

layers = [imageInputLayer(input_size, 'Name', 'input'), 
          upsample,
          regressionLayer('Name', 'reg')];
layers(1).Mean = 0; %zerocenter input

lgraph = layerGraph(layers);
net = assembleNetwork(lgraph);

input_image = rand(input_size);
size(input_image)   %Size before upsampling
predict(net,input_image);

act = activations(net, input_image, 'ups_2d');
size(act)  %Size after upsampling

figure
subplot(3,2,1)
imshow(input_image(:,:,1),[])
title('Input(6x6)')
subplot(3,2,3)
imshow(input_image(:,:,2),[])
subplot(3,2,5)
imshow(input_image(:,:,3),[])

subplot(3,2,2)
imshow(act(:,:,1),[])
title('Upsampled(12x12)')
subplot(3,2,4)
imshow(act(:,:,2),[])
subplot(3,2,6)
imshow(act(:,:,3),[])

%Images on the same row should be the same, while the ones on the right are higher resolution.

I hope this helps.