Inanna / News: Recent posts

The updated versions have many improvements and compatibility fixes.

The network training architecture is still very immature, and many object-oriented features have been compromized for the sake of training speed (vanilla backprop and RProp). However, most important features work quite nicely now.

Saving and loading networks works now, together with equalization data(!!!). The file format is roughly SNNS compatible, with a few additions. It's also possible to save multiple networks in the same file, although that definitely breaks the SNNS compatibility....... read more

The CVS versions of MagiClib and Inanna seem to work fine now. I managed to cvs checkout them, and compile (on Linux-i386).

Initial 0.2.alpha2 release seems to compile fine on Linux/i386. I believe there might be some trouble with other architectures.

The proj/backprop example seems to work. It requires the Proben1 data set, which is not available from here. You might try the XOR and Letters problems though.

The documentation is now downloadable, but it's very incomplete, erroneus and obsolete. Do not trust it.

Inanna requires MagiClib 1.0.1, included in the Files section. It's a personal base class library, and has some features which I like over any other base class libraries I've seen.... read more