process_training_data.m

% [input_data, mapped_label_data] = process_training_data(directory, n_chunks, cut_attack_conf)
% 
% Authors:  Babak Toghiani-Rizi & Marcus Windmark
% Date:    2016
% 
% Function Description: 
%
%   Creates feature vectors with labels of all files in the given
%   directory.
% 
% Input Parameters: 
%
%         directory: (string) Path to a directory to perform on.
%
%         n_chunks: (integer) Number of partitions to separate the
%                    spectrum in.
%
%         cut_attack_cond: (string) Which mode to operate in;
%                           empty string to perform normally, 'only_attack'
%                           to use only the attack as feature vector
%                           and 'without_attack' to exclude the attack.
% 
% Returned Value: 
%
%         input_data: (array) The feature vectors of all samples in given
%                      directory.
%
%         label_data: (array) The labels belonging to the returned feature 
%                      vectors.
%
%



function [input_data, mapped_label_data] = process_training_data(directory, n_chunks, cut_attack_conf)
    file_data = process_file_names(directory);
    training_data = cell(length(file_data),2);
    onset_length_ms = 100;

    for n = 1:length(file_data)
        fprintf('%d of %d\n', n, length(file_data));
        filename = file_data(n,3);
        filename = filename{1};

        [y,Fs] = audioread(filename);

        if not(isempty(cut_attack_conf))
           [only_attack, without_attack] = cut_attack(filename, onset_length_ms);

           if strcmp(cut_attack_conf, 'only_attack')
                y = only_attack;
           else
               if strcmp(cut_attack_conf, 'without_attack')
                    y = without_attack;
               end
           end
        end   

        Nsamps = length(y);

        % Do Fourier Transform
        y_fft = abs(fft(y));           %Retain Magnitude
        y_fft = y_fft(1:round(Nsamps/2));      %Discard Half of Points
        f = Fs*(0:Nsamps-1)/Nsamps;   %Prepare freq data for plot
        f = f(1:round(1000*Nsamps/Fs));

        % Shift frequency
        tone =  file_data(n,2);
        frequency_diff = frequency_difference(tone{1});
        shifted_fft = shift_frequency(y_fft, frequency_diff, Fs, Nsamps);

        % Limit shifted_fft
        shifted_fft = shifted_fft(1:round(1000*Nsamps/Fs));

        % Extract feature vector
        feature_vector = create_feature_vector(shifted_fft,n_chunks);

        feature_vector = normalize_data(feature_vector);

        training_data(n, 1) = file_data(n,1);
        training_data(n, 2) = {feature_vector};
    end

    input_data = transpose(cell2mat(training_data(:, 2)));
    label_data = training_data(:, 1);

    instrument_set = unique(label_data);
    instrument_value = (1:length(instrument_set));
    instrumentLabelMap = containers.Map(instrument_set, instrument_value);

    mapped_label_data = zeros(length(unique(label_data)),length(label_data));

    for N = 1:length(label_data)
        instrument = label_data(N);
        mapped_label_data(instrumentLabelMap(instrument{1}), N) = 1;
    end
end