I hate developing over someone else's code... and that particular code is Torch (www.torch.ch) - a machine learning library. Well, kind of.
I hate re-writing code if I can help it, but sometimes... bah... there is no way around and Torch has its own fast-but-not-very-versatile linked list class which is more like an extended-array kind of thing. This are not made much easier, considering that the linked-list access code is replicated across ALL of the Torch source. How about using a macro or something?
So this week, I had to do the following things, which... unfortunately I have been unable to integrate into the main library yet:
- Completed an experimental set on the letters database using the online RL-gated MLP mixture.
- Compiled necessary bibliographical references for report.
- Created a bufferred data container class and a FIFO (linked-list) class, with tags, hashing and data IDs.
- Used the linked list classes to create a data-format conversion program from arbitrary tag-named arrays to torch-formated numerical arrays.
- Fortunately the new classes were quite useful in creating a margin-distribution estimator for classifiers... although the measurer class still does not integrate nicely with online learning
Hm... hope I have everything working by monday... got a NIPS2002 deadline on 1st of July. And I have to go swimming at some nice baeach.,