Showing posts with label SMT. Show all posts
Showing posts with label SMT. Show all posts
March 31, 2009
Who Says You Can't Swear in C++?
I currently have the pleasure of using Och's YASMET tool for maximum entropy model optimization. (Quick background primer: maximum entropy is really a simple concept - assume nothing about the data beyond what you're given. In slightly more technical terms, you want the most uniform distribution that fits the constraints.) Looking at the source code is akin to reading one too many JAPH snippets or sifting through Google's obfuscated JavaScript - it's an experience both awe-inspiring and twitch-inducing. Of course, it's GNU GPL'd, so anyone who can navigate the mess of two-letter variable names, magic numbers, and minimally-whitespaced text is free to modify it! (A re-indenting utility might help here - though a quick run against the K&R style expands it to 362 lines and reveals the terrifying depth of the conditional statements lurking within. Fair warning.)
Labels:
c++,
machine learning,
optimization,
SMT
January 25, 2009
Level 3, Almost
93 problems and counting, including two of the 25 most recent (which means I get a nice red-coloured progress bar under Current Performance!) I'm shooting for 100 by the end of the week - enough for Level 3 status and a permanent place on the Project Euler scoreboard.
Aside from that, I'm currently optimizing my first project at work. After all, a new statistical machine translation feature function isn't much use unless you can run it across a cluster on corpora of at least 10 million phrase pairs within your lifetime. (In my defense, the goal of the first implementation was to get something working; judging from the unit tests, this objective seems to have been met.) Following discussion with co-workers at NRC-IIT, some intense whiteboard sessions, and profuse streams of invective launched (unfairly, perhaps) at valgrind, I'm confident that it can easily be improved to run within a day.
(Now that would be a good use of Python generators - (theoretically) infinite streams of invective, scraped from the bowels of our beloved Internet via httplib. Some would argue that the Internet's capacity to provide said invective is necessarily finite. I (and Bucket - NSFW) respectfully disagree.)
Aside from that, I'm currently optimizing my first project at work. After all, a new statistical machine translation feature function isn't much use unless you can run it across a cluster on corpora of at least 10 million phrase pairs within your lifetime. (In my defense, the goal of the first implementation was to get something working; judging from the unit tests, this objective seems to have been met.) Following discussion with co-workers at NRC-IIT, some intense whiteboard sessions, and profuse streams of invective launched (unfairly, perhaps) at valgrind, I'm confident that it can easily be improved to run within a day.
(Now that would be a good use of Python generators - (theoretically) infinite streams of invective, scraped from the bowels of our beloved Internet via httplib. Some would argue that the Internet's capacity to provide said invective is necessarily finite. I (and Bucket - NSFW) respectfully disagree.)
Labels:
bucket,
euler,
optimization,
programming,
SMT
Subscribe to:
Posts (Atom)
