Friday, April 30, 2010
The other day I noted that I couldn't install the darcs version of Happstack on my windows machine. There was a build failure that was reported on the mailing list. So I waited a few days, and tried again. But then, before I started building Happstack, something happened. I don't exactly remember what, I think I was a bit to eager with cabal upgrade, and I upgraded some base packages that I shouldn't have. I remember directory-184.108.40.206 failing to build and having to go through MSYS to get it to build. Then more and more things needed to be upgraded and ghc-pkg check was reporting more and more weird things. When I thought I had upgraded everything, I tried happstack. happstack-util built, but then happstack-data reported that happstack-util couldn't be built because it required both time-1.1.4 and time-2-something. Oh oh. I screwed around for a while. Then I saw some posts on the web talking about removing package.conf.d. In my random experiments with cabal upgrade I had managed to install everything in my user profile and not inside my Haskell Platform directory, so there was hope!
I deleted package.conf.d. ghc-pkg list reported a saner list of packages. I tried building happstack again. Then something "funny" happened: every cabal build I was trying to do first tried to reinstall process-220.127.116.11, and failed every time!! And of course process-18.104.22.168 appeared in the ouput of ghc-pkg list... WTF? Ah, ok, directory-22.214.171.124 is still installed, let's kill the beast! And then everything worked! Happstack built, happstack "hello world" server even runs, life is happy again!
I didn't try to investigate why Cabal or directory-126.96.36.199 were insisting on reinstalling something that I have already. I've seen enough weird things and spent enough time this week away from the nice world of pure lazy functional programming and in dirty low-level configuration and build issues to not care too much anymore...
Wednesday, April 28, 2010
Despite my earlier problems, I finally managed to install WXWidgets and WxHaskell on my Windows machine. Pfiu! I started again from scratch, and things went much better when I installed, as recommended, MinGW and MSYS in different folders. When I need to compile stuff, I only put MinGW in my path, and when I need Unixy tools like make I add MSYS. Of course MinGW has its own version of make, mingw32-make, that works when MSYS make doesn't...
So I installed MinGW in c:\dev\MinGW and:
Then I modified the WxWidgets config.gcc file to set SHARED=1 UNICODE=1 MONOLITHIC=1.
mingw32-make -f makefile.gcc BUILD=debug
Setting new environment variables (wxWidgets is installed in c:\ui):
Then I could install then wxdirect, wxcore and wx packages using a few more flags than what a default cabal install would do, for example:
runhaskell Setup configure --extra-lib-dirs=c:\dev\MinGW\lib --extra-include-dirs=c:\dev\MinGW\include --extra-include-dirs=c:\dev\MinGW\include\c++\3.4.5\mingw32 --extra-include-dirs=c:\ui\wxWidgets-2.8.10\include --extra-include-dirs=c:\dev\MinGW\include\c++\3.4.5
And it install, and works!
Thursday, April 22, 2010
Having secured fame and fortune with MazesOfMonad, my console based RPG, and having some ideas for some games that would look even better with a graphical front end, I decided it was time to take the bulls by the horn and go seriously into Haskell GUI programming. Well, the bull won and I feel truly gored...
So what's wrong? Well, after fooling around on my Windows 7 computer for a few nights I've achieved nothing. I now have three different copies of MinGW installed in different places. I have tried to installed SDL, GTK2HS, WXWidgets and Happstack and I have failed at all of them. I know WXWidgets is manageable because I've done it before, I only seem to have some include paths issues. libSDL itself installed but not the Haskell bindings. I found somebody with the same problem on the web, that then told me he had given up. I even managed to get Darcs errors with Gtk2Hs. In despair I thought I could develop a web based game but even the darcs version of happstack failed to compile, I remembered then some bug mentioned on the mailing list, so I should probably just download the latest stable version instead of the development branch.
Some problems are down to some silly things I did. For example installing the Haskell Platform in a directory with spaces is not a great idea (I don't remember if it was the default), because then a lot of paths referencing its MinGW install are wrong in the MSYS shell...
But most are due to the Unixy nature of a lot of the GUI components. Now I see a configure script and my blood freezes. I run "cabal install foo" and I get the dreaded message "this package has a configure script, you need to dive into Unix hell". Can you tell me why I have three versions of GCC, and two make a configure script crash with "c compiler cannot create executables" and the third one works? Where can I download one single file that will allow me to have a full MSYS/MinGW setup with most of what I need? I'm sure there are solutions to a lot of the problems I have, and maybe I need to spend more time reading the INSTALL files and such, and download everything from the MinGW files page. But why can't it Just Work? I mean, I work with Java at work, we do GUI work in SWT now and before that in Swing, we use a lot of libraries, and we never run into any installation problem. Maybe once or twice over the past couple of years we had a cross-platform compatibility issue or a library conflict. I know, I know, apples and oranges, but I'm now thinking that writing my own lazy, purely functional language with monads on top of the JVM with Java integration is going to be easier than getting a GUI library with Haskell bindings to work.
I like to write applications, not spend my free time trying to get libraries to install.
Friday, April 09, 2010
After the university of Waterloo AI Challenge I went back to a bit of AI programming in Haskell. Three years ago I had written some code to implement a neural network, and it didn't seem to learn as fast as the book said it should. I had not worked on that any more, but I decided to look at it again. And I found a bug!! I mixed up the deltas to pass on from one run to the next. Changing two lines of code makes my network learning much faster.
So now onto something a bit more meaty than exclusive or: very simple character recognition. I want to be able to recognize numbers written as a digital alarm clock would display them (at least mine): you have 7 lights, four vertical and three horizontal, and various patterns make the numbers. If all lights are on, it's 8, if only the middle horizontal light is off we have zero, etc...
I just test things with a network that has 7 input neurons (one for each light), 10 output neurons (one for each number) and 8 hidden neurons (the average between input and output)
n<-network 7 8 10
([1,1,1,-1,1,1,1],[1,0,0,0,0,0,0,0,0,0]), -- 0
([-1,-1,1,-1,-1,1,-1],[0,1,0,0,0,0,0,0,0,0]), -- 1
([1,-1,1,1,1,-1,1],[0,0,1,0,0,0,0,0,0,0]), -- 2
([1,-1,1,1,-1,1,1],[0,0,0,1,0,0,0,0,0,0]), -- 3
([-1,1,1,1,-1,1,-1],[0,0,0,0,1,0,0,0,0,0]), -- 4
([1,1,-1,1,-1,1,1],[0,0,0,0,0,1,0,0,0,0]), -- 5
([1,1,-1,1,1,1,1],[0,0,0,0,0,0,1,0,0,0]), -- 6
([1,1,1,-1,-1,1,-1],[0,0,0,0,0,0,0,1,0,0]), -- 7
([1,1,1,1,1,1,1],[0,0,0,0,0,0,0,0,1,0]), -- 8
([1,1,1,1,-1,1,1],[0,0,0,0,0,0,0,0,0,1]) -- 9
let (LNS(n',err,ds),e) = run (initialState n) defaultLearningRate trainingSet (1,1000)
And after something like 150 iterations I get a network that has a low error rate. A little helper method to massage the results:
runNumber n inputs=do
(_, output)=exec n inputs
normalized=map (\v->if (1-v)>0.5 then 0 else 1) output
return $ elemIndex 1 normalized
Main> runNumber nn [1,-1,1,1,-1,1,1]
Now, things are starting to get interesting! Hopefully I won't wait another three years before trying maybe more complex character recognition.