MP5.zip_HMM_human

clc datadir = 'AV_DATA'; %% Load Data % Files for digit 2 audio_files_2 = dir(strcat(datadir, filesep, '2*a.fea')); video_files_2 = dir(strcat(datadir, filesep, '2*v.fea')); % Files for digit 5 audio_files_5 = dir(strcat(datadir, filesep, '5*a.fea')); video_files_5 = dir(strcat(datadir, filesep, '5*v.fea')); nsequences = 10; audio_data_2 = cell(1, nsequences); video_data_2 = cell(1, nsequences); audio_data_5 = cell(1, nsequences); video_data_5 = cell(1, nsequences); av_data_2 = cell(1, nsequences); av_data_5 = cell(1, nsequences); for i=1:nsequences % Read in audio files for '2' audio_name_2 = strcat(datadir, filesep, audio_files_2(i).name); audio_data_2(i) = {dlmread(audio_name_2)}; % Read in video files for '2' video_name_2 = strcat(datadir, filesep, video_files_2(i).name); video_data_2(i) = {dlmread(video_name_2)}; % Read in audio files for '5' audio_name_5 = strcat(datadir, filesep, audio_files_5(i).name); audio_data_5(i) = {dlmread(audio_name_5)}; % Read in video files for '5' video_name_5 = strcat(datadir, filesep, video_files_5(i).name); video_data_5(i) = {dlmread(video_name_5)}; % Generate AV data through concatenation for '2' av_data_2{i} = [audio_data_2{i}, video_data_2{i}]; % Generate AV data through concatenation for '5' av_data_5{i} = [audio_data_5{i}, video_data_5{i}]; end %% Declare HMM parameters nlatent = 5; A_init = [0.8 0.2 0.0 0.0 0.0; 0.0 0.8 0.2 0.0 0.0; 0.0 0.0 0.8 0.2 0.0; 0.0 0.0 0.0 0.8 0.2; 0.0 0.0 0.0 0.0 1.0;]; audio_scores = zeros(2,1); % first row -> '2', second row -> '5' video_scores = zeros(2,1); av_scores = zeros(2,1); %% Train the audio HMM for i=1:20 fprintf('\n --- Test %d --- \n', i); % Form training set based on test example if i <= 10 Y = audio_data_2{i}; Yseq_2 = [audio_data_2(1:i-1), audio_data_2(i+1:end)]; Yseq_5 = audio_data_5; else Y = audio_data_5{i-10}; Yseq_2 = audio_data_2; Yseq_5 = [audio_data_5(1:i-11), audio_data_5(i-9:end)]; end disp('Training audio HMM for "2"') [P0_2,A_2,mu_2,sigma_2] = ghmm_learn(Yseq_2,nlatent,A_init); disp('Training audio HMM for "5"') [P0_5,A_5,mu_5,sigma_5] = ghmm_learn(Yseq_5,nlatent,A_init); % Test via ML rule: log P(O|lambda) = sum(scale) [~,scale_2] = ghmm_fwd(Y,A_2,P0_2,mu_2,sigma_2); [~,scale_5] = ghmm_fwd(Y,A_5,P0_5,mu_5,sigma_5); score_2 = sum(scale_2); score_5 = sum(scale_5); fprintf('Likelihood for 2: %f\n', score_2); fprintf('Likelihood for 5: %f\n', score_5); % Correct prediction for '2' if (i <= 10) && (score_2 > score_5) audio_scores(1) = audio_scores(1) + 1; end % Correct prediction for '5' if (i >= 11) && (score_5 > score_2) audio_scores(2) = audio_scores(2) + 1; end end %% Train the video HMM for i=1:20 fprintf('\n --- Test %d --- \n', i); % Form training set based on test example if i <= 10 Y = video_data_2{i}; Yseq_2 = [video_data_2(1:i-1), video_data_2(i+1:end)]; Yseq_5 = video_data_5; else Y = video_data_5{i-10}; Yseq_2 = video_data_2; Yseq_5 = [video_data_5(1:i-11), video_data_5(i-9:end)]; end disp('Training video HMM for "2"') [P0_2,A_2,mu_2,sigma_2] = ghmm_learn(Yseq_2,nlatent,A_init); disp('Training video HMM for "5"') [P0_5,A_5,mu_5,sigma_5] = ghmm_learn(Yseq_5,nlatent,A_init); % Test via ML rule: log P(O|lambda) = sum(scale) [~,scale_2] = ghmm_fwd(Y,A_2,P0_2,mu_2,sigma_2); [~,scale_5] = ghmm_fwd(Y,A_5,P0_5,mu_5,sigma_5); score_2 = sum(scale_2); score_5 = sum(scale_5); fprintf('Likelihood for 2: %f\n', score_2); fprintf('Likelihood for 5: %f\n', score_5); % Correct prediction for '2' if (i <= 10) && (score_2 > score_5) video_scores(1) = video_scores(1) + 1; end % Correct prediction for '5' if (i >= 11) && (score_5 > score_2) video_scores(2) = video_scores(2) + 1; end end %% Train the AV HMM for i=1:20 fprintf('\n --- Test %d --- \n', i); % Form training set based on test example if i <= 10 Y = av_data_2{i}; Yseq_2 = [av_data_2(1:i-1), av_data_2(i+1:end)]; Yseq_5 = av_data_5; else Y = av_data_5{i-10}; Yseq_2 = av_data_2; Yseq_5 = [av_data_5(1:i-11), av_data_5(i-9:end)]; end disp('Training AV HMM for "2"') [P0_2,A_2,mu_2,sigma_2] = ghmm_learn(Yseq_2,nlatent,A_init); disp('Training AV HMM for "5"') [P0_5,A_5,mu_5,sigma_5] = ghmm_learn(Yseq_5,nlatent,A_init); % Test via ML rule: log P(O|lambda) = sum(scale) [~,scale_2] = ghmm_fwd(Y,A_2,P0_2,mu_2,sigma_2); [~,scale_5] = ghmm_fwd(Y,A_5,P0_5,mu_5,sigma_5); score_2 = sum(scale_2); score_5 = sum(scale_5); fprintf('Likelihood for 2: %f\n', score_2); fprintf('Likelihood for 5: %f\n', score_5); % Correct prediction for '2' if (i <= 10) && (score_2 > score_5) av_scores(1) = av_scores(1) + 1; end % Correct prediction for '5' if (i >= 11) && (score_5 > score_2) av_scores(2) = av_scores(2) + 1; end end