/
ChangeLog
68 lines (51 loc) · 3.1 KB
/
ChangeLog
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
version 0.4-3 (09-1-2012):
- included a citation to the JSS article about RSNNS in the documentation and the CITATION file
- recompressed example data to fix a warning in CRAN checks
- fixed the jordan and elman examples: "decay=0.1" was given as parameter, instead of the correct "learnFuncParams=c(0.1)"
- added an example for training an MLP for the iris data, using the low-level api (mlp_irisSnnsR.R)
- fixed the decodeClassLables() function, which didn't correctly translate factor level names to column names
version 0.4-2 (29-9-2011):
- removed original SNNS manual, to avoid license confusions, as it states an outdated license
version 0.4-1 (28-7-2011):
- fixed a bug that prevented savePatFile to work
- fixed a bug that prevented compilation on Mac OSX
- fixed the "variable set but not used" warnings that occur in the most recent gcc versions
version 0.4-0 (21-6-2011):
- The art2 demo was fixed
- changed package encoding to UTF-8
- wrapping of some of the missing functions was added
- artmap was added in the high-level interface
- initialization to zero was added for all former static variables
(this was not always performed before, but it turned out that some
SNNS functions rely on this and crash otherwise, e.g. dlvq and artmap)
- segfaults related to Rcpp wrapping of (const char*) NULL were fixed
- an object serializaton mechanism was implemented that now allows for rsnns
objects being saved and loaded through R's normal save/load mechanism
- new normalization functions were added that allow for denormalization,
and norm of the test set with the parameters obtained on the training set
- some new high-level functions were implemented to extract internals of
the neural networks (getWeights, getUnitDefinitions)
- documentation of the high-level functions was greatly improved with references
to original literature and some descriptions
- original SNNS 4.2 User Manual was added to inst/doc
version 0.3-1 (17-11-2010):
fixes to make it run on Solaris
version 0.3 (15-11-2010):
port of the relevant SNNS parts to C++,
resulting in an SNNS fork named SnnsCLib.
version 0.1 and 0.2 were unpublished.
version 0.2:
used nearly unchanged SNNS code. In order to train various networks and
use them later for prediction, different ways seem possible: After training,
the model could be saved as a temporary .net file to disk or even to memory
using e.g. the C function fmemopen. For prediction, the net could be loaded
into the SNNS kernel again. Another potential possibility is to load the SNNS
library various times to memory. However, all these methods have mayor drawbacks
in stability, performance, and are not parallelizable.
version 0.1:
used swig instead of Rcpp to wrap the SNNS code. This has the advantage,
that the wrapper functions are automatically generated. However, it turned
out that some of the interfaces of SNNS functions are quite complicated
and the support of swig for R is not as advanced as for other programming
languages. Extending swig in order to get the needed functionality seemed
quite difficult, so that we chose to implement the wrapping manually using Rcpp.