|
a |
|
b/combinedDeepLearningActiveContour/functions/stack2params.m |
|
|
1 |
function [params, netconfig] = stack2params(stack) |
|
|
2 |
|
|
|
3 |
% Converts a "stack" structure into a flattened parameter vector and also |
|
|
4 |
% stores the network configuration. This is useful when working with |
|
|
5 |
% optimization toolboxes such as minFunc. |
|
|
6 |
% |
|
|
7 |
% [params, netconfig] = stack2params(stack) |
|
|
8 |
% |
|
|
9 |
% stack - the stack structure, where stack{1}.w = weights of first layer |
|
|
10 |
% stack{1}.b = weights of first layer |
|
|
11 |
% stack{2}.w = weights of second layer |
|
|
12 |
% stack{2}.b = weights of second layer |
|
|
13 |
% ... etc. |
|
|
14 |
|
|
|
15 |
|
|
|
16 |
% Setup the compressed param vector |
|
|
17 |
params = []; |
|
|
18 |
for d = 1:numel(stack) |
|
|
19 |
|
|
|
20 |
% This can be optimized. But since our stacks are relatively short, it |
|
|
21 |
% is okay |
|
|
22 |
params = [params ; stack{d}.w(:) ; stack{d}.b(:) ]; |
|
|
23 |
|
|
|
24 |
% Check that stack is of the correct form |
|
|
25 |
assert(size(stack{d}.w, 1) == size(stack{d}.b, 1), ... |
|
|
26 |
['The bias should be a *column* vector of ' ... |
|
|
27 |
int2str(size(stack{d}.w, 1)) 'x1']); |
|
|
28 |
if d < numel(stack) |
|
|
29 |
assert(size(stack{d}.w, 1) == size(stack{d+1}.w, 2), ... |
|
|
30 |
['The adjacent layers L' int2str(d) ' and L' int2str(d+1) ... |
|
|
31 |
' should have matching sizes.']); |
|
|
32 |
end |
|
|
33 |
|
|
|
34 |
end |
|
|
35 |
|
|
|
36 |
if nargout > 1 |
|
|
37 |
% Setup netconfig |
|
|
38 |
if numel(stack) == 0 |
|
|
39 |
netconfig.inputsize = 0; |
|
|
40 |
netconfig.layersizes = {}; |
|
|
41 |
else |
|
|
42 |
netconfig.inputsize = size(stack{1}.w, 2); |
|
|
43 |
netconfig.layersizes = {}; |
|
|
44 |
for d = 1:numel(stack) |
|
|
45 |
netconfig.layersizes = [netconfig.layersizes ; size(stack{d}.w,1)]; |
|
|
46 |
end |
|
|
47 |
end |
|
|
48 |
end |
|
|
49 |
|
|
|
50 |
end |