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
|
function [new_bel, niter, new_msg, edge_id, nstates] = bp_mrf2_general(adj_mat, pot, local_evidence, varargin)
% BP_MRF2_GENERAL Belief propagation on an MRF with pairwise potentials
% function [bel, niter] = bp_mrf2_general(adj_mat, pot, local_evidence, varargin)
%
% Input:
% adj_mat(i,j) = 1 iff there is an edge between nodes i and j
% pot(ki,kj,i,j) or pot{i,j}(ki,kj) = potential on edge between nodes i,j
% If the potentials on all edges are the same,
% you can just pass in 1 array, pot(ki,kj)
% local_evidence(state, node) or local_evidence{i}(k) = Pr(observation at node i | Xi=k)
%
% Use cell arrays if the hidden nodes do not all have the same number of values.
%
% Output:
% bel(k,i) or bel{i}(k) = P(Xi=k|evidence)
% niter contains the number of iterations used
%
% [ ... ] = bp_mrf2(..., 'param1',val1, 'param2',val2, ...)
% allows you to specify optional parameters as name/value pairs.
% Parameters names are below [default value in brackets]
%
% max_iter - max. num. iterations [ 5*nnodes]
% momentum - weight assigned to old message in convex combination
% (useful for damping oscillations) - currently ignored i[0]
% tol - tolerance used to assess convergence [1e-3]
% maximize - 1 means use max-product, 0 means use sum-product [0]
% verbose - 1 means print error at every iteration [0]
%
% fn - name of function to call at end of every iteration [ [] ]
% fnargs - we call feval(fn, bel, iter, fnargs{:}) [ [] ]
nnodes = length(adj_mat);
[max_iter, momentum, tol, maximize, verbose, fn, fnargs] = ...
process_options(varargin, 'max_iter', 5*nnodes, 'momentum', 0, ...
'tol', 1e-3, 'maximize', 0, 'verbose', 0, ...
'fn', [], 'fnargs', []);
if iscell(local_evidence)
use_cell = 1;
else
use_cell = 0;
[nstates nnodes] = size(local_evidence);
end
if iscell(pot)
tied_pot = 0;
else
tied_pot = (ndims(pot)==2);
end
% give each edge a unique number
ndx = find(adj_mat);
nedges = length(ndx);
edge_id = zeros(1, nnodes*nnodes);
edge_id(ndx) = 1:nedges;
edge_id = reshape(edge_id, nnodes, nnodes);
% initialise messages
if use_cell
prod_of_msgs = cell(1, nnodes);
old_bel = cell(1, nnodes);
nstates = zeros(1, nnodes);
old_msg = cell(1, nedges);
for i=1:nnodes
nstates(i) = length(local_evidence{i});
prod_of_msgs{i} = local_evidence{i};
old_bel{i} = local_evidence{i};
end
for i=1:nnodes
nbrs = find(adj_mat(:,i));
for j=nbrs(:)'
old_msg{edge_id(i,j)} = normalise(ones(nstates(j),1));
end
end
else
prod_of_msgs = local_evidence;
old_bel = local_evidence;
%old_msg = zeros(nstates, nnodes, nnodes);
old_msg = zeros(nstates, nedges);
m = normalise(ones(nstates,1));
for i=1:nnodes
nbrs = find(adj_mat(:,i));
for j=nbrs(:)'
old_msg(:, edge_id(i,j)) = m;
%old_msg(:,i,j) = m;
end
end
end
converged = 0;
iter = 1;
while ~converged & (iter <= max_iter)
% each node sends a msg to each of its neighbors
for i=1:nnodes
nbrs = find(adj_mat(i,:));
for j=nbrs(:)'
if tied_pot
pot_ij = pot;
else
if iscell(pot)
pot_ij = pot{i,j};
else
pot_ij = pot(:,:,i,j);
end
end
pot_ij = pot_ij'; % now pot_ij(xj, xi)
% so pot_ij * msg(xi) = sum_xi pot(xj,xi) msg(xi) = f(xj)
if 1
% Compute temp = product of all incoming msgs except from j
% by dividing out old msg from j from the product of all msgs sent to i
if use_cell
temp = prod_of_msgs{i};
m = old_msg{edge_id(j,i)};
else
temp = prod_of_msgs(:,i);
m = old_msg(:, edge_id(j,i));
end
if any(m==0)
fprintf('iter=%d, send from i=%d to j=%d\n', iter, i, j);
keyboard
end
m = m + (m==0); % valid since m(k)=0 => temp(k)=0, so can replace 0's with anything
temp = temp ./ m;
temp_div = temp;
end
if 1
% Compute temp = product of all incoming msgs except from j in obvious way
if use_cell
%temp = ones(nstates(i),1);
temp = local_evidence{i};
for k=nbrs(:)'
if k==j, continue, end;
temp = temp .* old_msg{edge_id(k,i)};
end
else
%temp = ones(nstates,1);
temp = local_evidence(:,i);
for k=nbrs(:)'
if k==j, continue, end;
temp = temp .* old_msg(:, edge_id(k,i));
end
end
end
%assert(approxeq(temp, temp_div))
assert(approxeq(normalise(pot_ij * temp), normalise(pot_ij * temp_div)))
if maximize
newm = max_mult(pot_ij, temp); % bottleneck
else
newm = pot_ij * temp;
end
newm = normalise(newm);
if use_cell
new_msg{edge_id(i,j)} = newm;
else
new_msg(:, edge_id(i,j)) = newm;
end
end % for j
end % for i
old_prod_of_msgs = prod_of_msgs;
% each node multiplies all its incoming msgs and computes its local belief
if use_cell
for i=1:nnodes
nbrs = find(adj_mat(:,i));
prod_of_msgs{i} = local_evidence{i};
for j=nbrs(:)'
prod_of_msgs{i} = prod_of_msgs{i} .* new_msg{edge_id(j,i)};
end
new_bel{i} = normalise(prod_of_msgs{i});
end
err = abs(cat(1,new_bel{:}) - cat(1, old_bel{:}));
else
for i=1:nnodes
nbrs = find(adj_mat(:,i));
prod_of_msgs(:,i) = local_evidence(:,i);
for j=nbrs(:)'
prod_of_msgs(:,i) = prod_of_msgs(:,i) .* new_msg(:,edge_id(j,i));
end
new_bel(:,i) = normalise(prod_of_msgs(:,i));
end
err = abs(new_bel(:) - old_bel(:));
end
converged = all(err < tol);
if verbose, fprintf('error at iter %d = %f\n', iter, sum(err)); end
if ~isempty(fn)
if isempty(fnargs)
feval(fn, new_bel);
else
feval(fn, new_bel, iter, fnargs{:});
end
end
iter = iter + 1;
old_msg = new_msg;
old_bel = new_bel;
end % while
niter = iter-1;
fprintf('converged in %d iterations\n', niter);
|
