diff --git a/ChangeLog.md b/ChangeLog.md new file mode 100644 --- /dev/null +++ b/ChangeLog.md @@ -0,0 +1,12 @@ +0.0.1.1 +------- + +- added simple demo +- renamed statistic TaskCount to Count +- fixed nondescript error message on empty input +- relaxed version bounds on dependencies + +0.0.1.0 +------- + +- initial version diff --git a/Driver.hs b/Driver.hs --- a/Driver.hs +++ b/Driver.hs @@ -86,6 +86,7 @@ istate = Engine.init responses engine = setupEngine args in do printStats resp responses + when (length resp > 0) $ do maybeWriteFile oResponses (formatResp oRespFormat responses) putStrLn "Reading initial task parameters..." istate <- maybe return readParams iTaskParams $ istate @@ -96,7 +97,7 @@ thetas results `seq` params results `seq` return () putStrLn "Calculating bayes probabilities..." let (bayesBounds, bayesValues) = calcBayes results in do - putStrLn "Writing bayes probability values..." + putStrLn "Saving bayes probability values..." maybeWriteFile oBayesPlot . buildTable $ bayesValues putStrLn "Writing task parameters..." let taskBase = tableTaskParams . getTaskParamsList $ results in do diff --git a/Engine.hs b/Engine.hs --- a/Engine.hs +++ b/Engine.hs @@ -35,7 +35,7 @@ where [dt] = diff (thetas old) (thetas new) dp = diff (params old) (params new) - diff xs ys = (max' . trans . sub xs) ys + diff xs ys = (max' . trans) (sub xs ys) where max' :: [[Double]] -> [Double] max' = map (maximum . map abs) diff --git a/Statistics.hs b/Statistics.hs --- a/Statistics.hs +++ b/Statistics.hs @@ -20,7 +20,7 @@ import System.Random.MWC statTheta :: StatisticType -> ContestantsData -> IO Statistic -statTheta TaskCount xs = return . SingleStatistic . map (fromIntegral . V.length . snd) $ xs +statTheta Count xs = return . SingleStatistic . map (fromIntegral . V.length . snd) $ xs statTheta SolvedProp xs = return . SingleStatistic . map (prop . snd) $ xs where prop = mean . V.map (ok . snd) @@ -75,7 +75,7 @@ statTask :: StatisticType -> TasksData -> IO Statistic -statTask TaskCount xs = return . SingleStatistic . map (fromIntegral . V.length . snd) $ xs +statTask Count xs = return . SingleStatistic . map (fromIntegral . V.length . snd) $ xs statTask SolvedProp xs = return . SingleStatistic . map (prop . snd) $ xs where prop = mean . V.map (ok . snd) diff --git a/Types.hs b/Types.hs --- a/Types.hs +++ b/Types.hs @@ -69,7 +69,7 @@ type TaskParams = UV.Vector TaskParam type Thetas = UV.Vector Theta -data StatisticType = TaskCount | SolvedProp | LogLikelihood +data StatisticType = Count | SolvedProp | LogLikelihood | DLogLikelihood | FisherSEM | Bootstrap deriving (Eq, Show, Data, Typeable) diff --git a/demo.sh b/demo.sh new file mode 100644 --- /dev/null +++ b/demo.sh @@ -0,0 +1,85 @@ +#!/bin/sh -v + +function find_exe() { + HIRT= + [ -z "$HIRT" ] && type hirt 2>/dev/null && HIRT=hirt + [ -z "$HIRT" ] && [ -x hirt ] && HIRT=./hirt + [ -z "$HIRT" ] && [ -x dist/build/hirt/hirt ] && HIRT=dist/build/hirt/hirt +} + +find_exe + +if [ -z "$HIRT" ]; then + echo "Could not find executable..." + echo "Perhaps build it with:" + echo " cabal configure && cabal build" + echo " or cabal install hirt" + exit 1 +fi >&2 + + +# a very simple sample +RESPONSES=/tmp/responses +TASKPARAM=/tmp/params +BAYES=/tmp/bayes + +cat > "$RESPONSES" << EOF +contestant task result +c1 t1 0 +c1 t2 0 +c2 t1 0 +c2 t2 1 +c3 t2 1 +c3 t1 1 +EOF + +# fit model with default options +"$HIRT" "$RESPONSES" + +# the differences in loglikehood on this very small example are tiny, +# throwing off the JML estimate, let's try BFGS + +"$HIRT" "$RESPONSES" --algo lbfgsb + +# Better, but not quite there yet. +# We will increase the precision. + +"$HIRT" "$RESPONSES" --algo lbfgsb --prec 1e-30 + +# Let's save task parameters for later use +# and show contestant ability estimates. + +"$HIRT" "$RESPONSES" --algo lbfgsb --prec 1e-30 --otaskparam "$TASKPARAM" --otheta /dev/stdout + +# Now, let's use the task parameter estimates +# to estimate contestant abilities. +# We will use one round of JML to leave task parameters fixed. + +"$HIRT" "$RESPONSES" --algo jml -n 1 --itaskparam "$TASKPARAM" --otheta /dev/stdout + +# Now we will show some statistics + +"$HIRT" "$RESPONSES" --algo lbfgsb --prec 1e-30 --otaskparam /dev/stdout --otheta /dev/stdout \ + --taskstats count --taskstats solvedprop --taskstats loglikelihood --taskstats dloglikelihood \ + --thetastats count --thetastats solvedprop --thetastats loglikelihood --thetastats dloglikelihood \ + --thetastats fishersem --thetastats bootstrap + +# Finally, we will plot a graph of the bayes expected a posteriori probability + +"$HIRT" "$RESPONSES" --algo lbfgsb --prec 1e-30 --otaskparam /dev/stdout --otheta /dev/stdout \ + --obayesplot /tmp/bayes + +cat > /tmp/plot.R << EOF +library(ggplot2) + +x <- read.table(file="$BAYES", header=T) +p <- ggplot(data=x, aes(x=x, y=p, color=contestant)) + + geom_line(size=1) + + scale_color_brewer(palette="Set1",name="") + + scale_x_continuous(expression(theta)) + + scale_y_continuous("y") + +ggsave(file="/tmp/plot.pdf", w=10, h=7, plot=p) +EOF + +Rscript /tmp/plot.R diff --git a/hirt.cabal b/hirt.cabal --- a/hirt.cabal +++ b/hirt.cabal @@ -1,5 +1,5 @@ Name: hirt -Version: 0.0.1.0 +Version: 0.0.1.1 Synopsis: Calculates IRT 2PL and 3PL models Description: Program for fitting Item Response Theory (IRT) two (2PL) and @@ -30,12 +30,15 @@ Build-type: Simple Cabal-version: >=1.6 +Extra-source-files: + ChangeLog.md + demo.sh + Flag PL3 Description: Compile for 3PL model, doesn't support JML yet. Model needs to be selected at compile time. Default: False --- Extra-source-files: Executable hirt Main-is: Main.hs @@ -45,12 +48,12 @@ Build-depends: base >= 4 && < 5, vector >= 0.9 && < 0.10, - containers >= 0.4 && < 0.5, + containers >= 0.4 && < 0.6, text >= 0.11.1.13 && < 0.12, attoparsec >= 0.10.1 && < 0.11, text-format >= 0.3.0.7 && < 0.4, csv >= 0.1.2 && < 0.2, - hmatrix >= 0.13.1.0 && < 0.14, + hmatrix >= 0.13.1.0 && < 0.15, numeric-extras >= 0.0.2.2 && < 0.1, cmdargs >= 0.9.3 && < 0.10, random >= 1.0.1.1 && < 1.1,