what this function do?

function inputs = getBatchHdd(imdb, batch)

args_data = [{imdb.images.data(batch)} imdb.args];

noise = imdb.images.noise(:,:,:, batch);

label0 = imdb.images.label0(:,:,:, batch);

label1 = imdb.images.label1(:,:,:, batch);

data = vl_imreadjpeg(args_data{:});

data = data{1}/255*2-1; % set in the range of [-1:1]

if (imdb.gpus == true)

      noise = gpuArray(noise);

      data = gpuArray(data);

      label0 = gpuArray(label0);

      label1 = gpuArray(label1);    
  end

inputs = ({'noise', noise, 'data', data, 'label0',label0 'label1',label1} );

return;

[Walter Roberson]

Is this from https://github.com/sunghbae/dcgan-matconvnet/blob/master/get_train_DCGAN.m ?

[mohammed mahmoud]

yes

Answer 1

[Walter Roberson]

It pre-processes the data for the given batch number to scale it to [-1 +1], and creates a cell array with the corresponding noise array, data, and two label entries. If the gpus option is set it creates those arrays on the GPU. Oh, and it looks like it might read the image from disk along the way.

[mohammed mahmoud]

Thanks, but i don`t understand the step of add noise.

[Walter Roberson]

The function does not add noise: it retrieves a noise array that has already been calculated during the database creation. The noise will not be added to the data until later in the processing.

The noise array is created in src/cnn_train_dag_dcgan as

   imdb.images.noise = single(randn(1, 1, 100, imdb.nSample, 'single'));

The surrounding comments indicate that the noise is pre-created for the purposes of reproducibility.