-
Notifications
You must be signed in to change notification settings - Fork 7
/
demo_segm_ibpHMM.m
212 lines (177 loc) · 8.42 KB
/
demo_segm_ibpHMM.m
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
% Main demo scripts for the ICSC-HMM Segmentation Algorithm proposed in:
%
% N. Figueroa and A. Billard, “Transform-Invariant Clustering of SPD Matrices
% and its Application on Joint Segmentation and Action Discovery}”
% Arxiv, 2017.
%
% Author: Nadia Figueroa, PhD Student., Robotics
% Learning Algorithms and Systems Lab, EPFL (Switzerland)
% Email address: nadia.figueroafernandez@epfl.ch
% Website: http://lasa.epfl.ch
% November 2016; Last revision: 25-May-2017
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% --Select a Dataset to Test-- %%
%% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% 1) Toy 2D dataset, 3 Unique Emission models, 3 time-series, same swicthing
clc; clear all; close all;
N_TS = 3; display = 2 ; % 0: no-display, 1: raw data in one plot, 2: ts w/labels
[data, Data, True_states, True_theta] = genToyHMMData_Gaussian( N_TS, display );
super_states = 0;
%% 2a) Toy 2D dataset, 4 Unique Emission models, 5 time-series
clc; clear all; close all;
[data, TruePsi, Data, True_states] = genToySeqData_Gaussian( 4, 2, 2, 500, 0.5 );
dataset_name = '2D';
super_states = 0;
% Feat matrix F (binary 5 x 4 matrix )
if exist('h0','var') && isvalid(h0), delete(h0);end
h0 = plotFeatMat( TruePsi.F);
%% 2b) Toy 2D dataset, 2 Unique Emission models transformed, 4 time-series
clc; clear all; close all;
[data, TruePsi, Data, True_states] = genToySeqData_TR_Gaussian(4, 2, 3, 500, 0.5 );
dataset_name = '2D Transformed';
super_states = 1;
% Feat matrix F (binary 4 x 4 matrix )
if exist('h0','var') && isvalid(h0), delete(h0);end
h0 = plotFeatMat( TruePsi.F);
% Similarity matrix S (4 x 4 matrix)
if exist('h1','var') && isvalid(h1), delete(h1);end
h1 = plotSimMat( TruePsi.S );
%% 5) Real 'Peeling' (max) 32-D dataset, 5 Unique Emission models, 3 time-series
% Demonstration of a Bimanual Peeling Task consisting of
% 3 (32-d) time-series X = {x_1,..,x_T} with variable length T.
% Dimensions:
% x_a = {pos_x, pos_y, pos_z, q_i, q_j, q_k, q_w, f_x, f_y, f_z, tau_x, tau_y, tau_z}
% - positions: Data{i}(1:3,:) (3-d: x, y, z)
% - orientations: Data{i}(4:7,:) (4-d: q_i, q_j, q_k, q_w)
% - forces: Data{i}(8:10,:) (3-d: f_x, f_y, f_z)
% - torques: Data{i}(11:13,:) (3-d: tau_x, tau_y, tau_z)
% x_p = {pos_x, pos_y, pos_z, q_i, q_j, q_k, q_w, f_x, f_y, f_z, tau_x, tau_y, tau_z}
% - same as above Data{i}(14:26,:)
% x_o = {mu_r, mu_g, mu_b, sigma_r, sigma_g, sigma_b}
% - rate_mean: Data{i}(27:29,:) (3-d: mu_r, mu_g, mu_b)
% - rate_variance: Data{i}(30:32,:) (3-d: sigma_r, sigma_g, sigma_b)
% Dimension type:
% dim: 'all', include all 32 dimensions (active + passive robots + object)
% dim: 'robots', include only 26-d from measurements from active + passive robots
% dim: 'act+obj', include only 19-d from measurements from active robot + object
% dim: 'active', include only 13-d from measurements from active robot
% Dataset type:
% sub-sampled to 100 Hz (from 500 Hz), smoothed f/t trajectories, fixed rotation
% discontinuities.
clc;
clear all; close all
data_path = './test-data/'; display = 1;
% Type of data processing
% O: no data manipulation -- 1: zero-mean -- 2: scaled by range * weights
normalize = 2;
% Select dimensions to use
dim = 'active';
% Define weights for dimensionality scaling
weights = [3*ones(1,3) 1/2*ones(1,4) 1/10*ones(1,3) 1/20*ones(1,3)]';
switch dim
case 'active'
case 'robots'
weights = [weights' 1/7*ones(1,3) 1/2*ones(1,4) 1/20*ones(1,3) 1/10*ones(1,3)]';
end
% Select dimensions to use
dim = 'robots';
% Define weights for dimensionality scaling
switch dim
case 'active'
weights = [5*ones(1,3) 2*ones(1,4) 1/10*ones(1,6)]'; % active
case 'act+obj'
weights = [2*ones(1,7) 1/10*ones(1,6) 2*ones(1,6)]'; % act+obj
case 'robots'
% weights = [5*ones(1,3) ones(1,4) 1/10*ones(1,6) ones(1,3) 2*ones(1,4) 1/20*ones(1,6) ]'; % robots(velocities)
weights = [5*ones(1,3) 1/2*ones(1,4) 1/10*ones(1,6) ones(1,3) 1/2*ones(1,4) 1/20*ones(1,6) ]'; % robots(position)
case 'all'
weights = [5*ones(1,3) ones(1,4) 1/10*ones(1,6) ones(1,3) 2*ones(1,4) 1/20*ones(1,6) 2*ones(1,6) ]'; % all
end
% Define if using first derivative of pos/orient
use_vel = 0;
[data, TruePsi, Data, True_states, Data_] = load_peeling_dataset( data_path, dim, display, normalize, weights, use_vel);
dataset_name = 'Peeling';
%% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% Step 2: Run Collapsed IBP-HMM Sampler T times for good statistics %%
%% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%% Define Settings for IBP-HMM %%%
% IBP hyper-parametrs
gamma = 2*length(Data); % length(Data)
alpha = 1; % typically 1.. could change
kappa = 10; % sticky parameter
% Model Setting (IBP mass, IBP concentration, HMM alpha, HMM sticky)
modelP = {'bpM.gamma', gamma, 'bpM.c', 1, 'hmmM.alpha', alpha, 'hmmM.kappa', kappa,'obsM.Scoef',1};
% Sampler Settings
algP = {'Niter', 1000, 'HMM.doSampleHypers', 1, 'BP.doSampleMass',1,'BP.doSampleConc', 0, ...
'doSampleFUnique', 1, 'doSplitMerge', 0};
% Number of Repetitions
T = 10;
% Run MCMC Sampler for T times
Sampler_Stats = [];
jobID = ceil(rand*1000);
for run=1:T
% Run Gibbs Sampler for Niter once.
clear CH
% Start out with random number of features
initP = {'F.nTotal', randsample(10,1)};
CH = runBPHMM( data, modelP, {jobID, run}, algP, initP, './ibp-Results' );
Sampler_Stats(run).CH = CH;
end
%% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% Extract Estimation Statistics %%
%% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%% Visualize Sampler Convergence/Metrics and extract Best Psi/run %%%%%%%%%%
if exist('h1','var') && isvalid(h1), delete(h1);end
if exist('h1b','var') && isvalid(h1b), delete(h1b);end
% [h1, h1b, Best_Psi] = plotSamplerStatsBestPsi(Sampler_Stats);
[h1, h1b, Best_Psi] = plotSamplerStatsBestPsi(Sampler_Stats, 'hist');
%%%%%% Compute Clustering/Segmentation Metrics vs Ground Truth %%%%%%
if isfield(TruePsi, 'sTrueAll')
true_states_all = TruePsi.sTrueAll;
else
true_states_all = data.zTrueAll;
end
% Compute metrics for IBP-HMM
ibp_stats = computeSegmClustmetrics(true_states_all, Best_Psi);
%% %%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% Extract Best Run and Plot Results %%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%% Choose best run
log_probs = zeros(1,T);
for ii=1:T; log_probs(ii) = Best_Psi(ii).logPr; end
[val_max, id_max] = sort(log_probs,'descend');
bestPsi = Best_Psi(id_max(1));
% Plot Segmentation with Chosen Run
if exist('h2','var') && isvalid(h2), delete(h2);end
title_name = strcat(dataset_name, ': $\mathcal{IBP}$-HMM Segmentation');
[ h2 ] = plotSingleLabelSegmentation(data, bestPsi, data_legends, title_name);
% Plot Estimated Feature Matrix
if exist('h3','var') && isvalid(h3), delete(h3);end
[ h3 ] = plotFeatMat( bestPsi.Psi.F);
% Plot Estimated Transition Matrices
if exist('h4','var') && isvalid(h4), delete(h4);end
[h4, bestPsi] = plotTransitionMatrices(bestPsi);
% Compute Segmentation and State Clustering Metrics
results = computeSegmClustmetrics(true_states_all, bestPsi);
% Plot Segmentated 3D Trajectories
est_states = [];
for e=1:length(bestPsi.Psi.stateSeq)
est_states{e} = bestPsi.Psi.stateSeq(e).z';
end
N = size(Data{1},2);
if N > 2
% Plot Ground Segmentations
if exist('h5','var') && isvalid(h5), delete(h5);end
title_name = {dataset_name,'Ground Truth'};
title_edge = [0 0 0];
h5 = plot3D_labeledTrajectories(Data,True_states,plot_subsample,title_name, title_edge, dataset_view);
% Plot Estimated Segmentations
if exist('h6','var') && isvalid(h6), delete(h6);end
title_name = {dataset_name,' $\mathcal{IBP}$-HMM Segmentation'};
title_edge = [0.470588235294118 0.670588235294118 0.188235294117647];
h6 = plot3D_labeledTrajectories(Data,est_states,plot_subsample,title_name, title_edge, dataset_view);
end