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
|
function [W, bias, Y0] = dnn_builtin2gb (W, bias, Y0)
%DNN_BUILTIN2GB convert sparse deep neural network from built-in to GraphBLAS
%
% Usage:
%
% [W, bias, Y0] = dnn_builtin2gb (W, bias, Y0) ;
%
% This is mostly optional since GrB.* methods can take inputs that are either
% GrB objects or built-in sparse matrices. The problem is converted to single
% precision since it gives the same result. This is a bit faster, but built-in
% sparse single precision matrices don't exist, so a conversion from one to the
% other needs to be made.
%
% The bias{i} matrix differs, and this needs to be modified here (or in
% dnn_builtin.m). For dnn_builtin.m, bias{i} is a 1-by-n row vector. For the
% GraphBLAS semiring, it is an n-by-n diagonal matrix. When comparing GrB.dnn
% and dnn_builtin.m, this code should not be considered extra work, since the
% problem could be generated in GraphBLAS format to begin with. In that case,
% dnn_builtin.m would include this conversion code, to convert the problem from
% GraphBLAS format to built-in sparse matrices.
%
% In any case, the setup time is very low.
%
% See also GrB.dnn, dnn_builtin.
% SuiteSparse:GraphBLAS, Timothy A. Davis, (c) 2017-2022, All Rights Reserved.
% SPDX-License-Identifier: Apache-2.0
fmt = 'by row' ;
prec = 'single' ;
d = struct ('format', fmt) ;
n = size (Y0, 2) ;
Y0 = GrB (Y0, prec, fmt) ;
for k=1:length(W)
W {k} = GrB (W {k}, prec, fmt) ;
bias {k} = GrB.build (1:n, 1:n, bias {k}, n, n, '+', prec, d) ;
end
|
