IMDBAddImageNet.m

function [ imdb, select ] = IMDBAddImageNet( imdb, path, task_num, varargin )
% IMDBADDIMAGENET   Add to imdb the ImageNet (ILSVRC 2012) dataset.
%   Input:
%     PATH struct generated by GETPATH()
%     TASK_NUM the path number (see CNN_CUSTOMTRAIN) that this dataset corresponds to
%   Options:
%     See code comments
% 
% Authors: Zhizhong Li
% 
% See the COPYING file.

opts.partial = 0; % for >0 partial, e.g. 0.3, only include that much portion of # samples.
opts.trainval = [1 2]; % 1 for train, 2 for val. By default include train+val.
opts.randstream = []; % use randstream if provided
opts = vl_argparse(opts, varargin);

n_synset = numel(path.files_imgtrain_synsets);

if ~isfield(imdb, 'images')
    imdb.images.name = [];
    imdb.images.label = [];
    imdb.images.set = [];
    imdb.images.task = [];
end

% for training
if ismember(1, opts.trainval)
    if exist(path.path_imgIMDB.train, 'file')
        imgimdb = load(path.path_imgIMDB.train);
        name = imgimdb.name;
        label = imgimdb.label;
        set = ones(numel(label), 1);
    else
        name = cell(n_synset,1);
        label = cell(n_synset,1);
        set = cell(n_synset,1);

        % dir through all directories
        for i = 1:n_synset
            synset_name = path.files_imgtrain_synsets{i};
            images_file = dir(fullfile(path.path_imgtrain, synset_name, '*.JPEG'));
            name{i} = strcat(synset_name, '/', {images_file.name});
            n_imgs = numel(name{i});
            label{i} = ones(n_imgs,1) * i;
            set{i} = ones(n_imgs,1);
        end

        % concat
        name = reshape([name{:}], [], 1);
        label = num2cell(cell2mat(label));
        set = cell2mat(set);
        save(path.path_imgIMDB.train, 'name', 'label');
    end

    % only using part of it
    if opts.partial
        if isempty(opts.randstream)
            select.train = randperm(numel(name));
        else
            select.train = randperm(opts.randstream, numel(name));
        end
        select.train = select.train(1:ceil(numel(name) * opts.partial));
        name = name(select.train);
        label = label(select.train);
        set = set(select.train);
    else
        select.train = 1:numel(name);
    end
else
    name = {}; label = {}; set = [];
end
    
% for validation...
if ismember(2, opts.trainval)
    if exist(path.path_imgIMDB.val, 'file')
        imgimdb = load(path.path_imgIMDB.val);
        val_name = imgimdb.val_name;
        val_label = imgimdb.val_label;
        val_set = 2 * ones(numel(val_label), 1);
    else
        % translate the official metadata order (i.e. validation label) to the
        % dir result order (i.e. Alexnet's order)
        meta = load(path.file_imgmeta);
        wnids = cell2mat(cellfun(@(x) str2double(x(2:end)), {meta.synsets.WNID}, 'UniformOutput', false));
        val_officiallabel = load(path.file_imgvalgt);
        val_wnid = mat2cell(wnids(val_officiallabel)',ones(size(val_officiallabel)));

        dirwnid = cellfun(@(x) str2double(x(2:end)), path.files_imgtrain_synsets, 'UniformOutput', false);
        wniddict = containers.Map(dirwnid, 1:numel(dirwnid));
        val_label = values(wniddict, val_wnid); % no cell2mat

        % then get the file names
        val_name = dir(fullfile(path.path_imgval, '*.JPEG'));
        val_name = strcat(path.path_imgval_relative_to_imgtrain, {val_name.name}');
        assert(numel(val_name) == numel(val_label));
        val_set = 2 * ones(numel(val_label), 1);

        save(path.path_imgIMDB.val, 'val_name', 'val_label');
    end

    % only using part of it
    if opts.partial
        if isempty(opts.randstream)
            select.val = randperm(numel(name));
        else
            select.val = randperm(opts.randstream, numel(name));
        end
        select.val = select.val(1:ceil(numel(name) * opts.partial));
        val_name = val_name(select.val);
        val_label = val_label(select.val);
        val_set = val_set(select.val);
    else
        select.val = 1:numel(name);
    end

    name = [ name; val_name ];
    label = [ label; val_label ];
    set = [ set; val_set ];
    
end

imdb.images.name = [ imdb.images.name; name ];
imdb.images.label = [ imdb.images.label; label ];
imdb.images.set = [ imdb.images.set; set ];
imdb.images.task = [ imdb.images.task;
    task_num * ones(size(set,1), 1) ];