add docs

OnRunning.md

@@ -0,0 +1,27 @@
+# Running Experiments with Lambda:
+This is not supposed to be a instruction on how to do it properly, but it is a writeup on how i did it.
+If you want to do it properly, extend the command line Arguments for haga-lambda and allow runtime tweaking of Hyperparams and Datasets. While at it, generalizing LamdaCalculusV1 would be smart, too. You can use LamdaCalculusV2 as a template on how to do it more properly. (I wrote that later, and was IMO quite a bit smarter about it. I sadly didn't have time to fix up V1...)
+
+You just want to do the same hack i did or know about it?
+
+create a branch for each Dataset-experiment pair. e.g. iris_1 ... iris_9
+
+here git is your friend, especially if you inevitably screw up.
+e.g. echo git\ checkout\ iris_{1..9}\;\ git\ cherry-pick\ 7ced1e1\; will create a command for applying the commit 7ced1e1 to every iris branch.
+
+Adapt the build.sbatch and run.sbatch and **commit them**!
+clone the branch you committed to on the cluster.
+create the required folders! If you forget the output one, slurm will fail silently!
+
+Make sure to sbatch an adapted **build.sbatch before run.sbatch**!
+build.sbatch will need to be adapted for and run on every node you will use!
+Otherwise stuff WILL break!
+
+sbatch run.sbatch
+
+You can use squeue to monitor progress.
+
+A huge slew of raw data will be dumped into the output Folder. The error files contain results, the output files stats during training.
+
+On how to process these results, see: https://merl.dnshome.de/git/Hans/haga-graphics
+

build.sbatch

@@ -1,9 +1,28 @@
 #!/usr/bin/env bash
 #SBATCH --time=00:10:00
 #SBATCH --partition=cpu
-#SBATCH --output=./output/output_build.txt
+# 9 Experiments * 3 Datasets
-#SBATCH --error=./output/error_build.txt
+#SBATCH --array=0-27
+# ensure output exists, is a folder and is writable in your working directory
+#SBATCH --output=./output/output_run_%a.txt
+#SBATCH --error=./output/error_run_%a.txt
+# run once for every node you plan to use
 #SBATCH --nodelist=oc-compute02
-#SBATCH --mem=4G
+#SBATCH --mem=2G
-#SBATCH -c16
+
-srun nix develop --command stack --no-nix --system-ghc --no-install-ghc build
+
+# list your branches
+problems=("iris" "nurse" "german")
+
+#9 Experiments
+current_problem=${problems[(${SLURM_ARRAY_TASK_ID}/9)]}
+#9 Experiments
+current_variant=$(((${SLURM_ARRAY_TASK_ID}) % 9 + 1))
+current_branch="${current_problem}_${current_variant}"
+
+# ensure [full path to writable folder on node *] exists
+git clone -b $current_branch --single-branch "[your git repo]" [full path to writable folder on node 1]/$current_branch
+git clone -b $current_branch --single-branch "[your git repo]" [full path to writable folder on node 1]/$current_branch
+#... for every node
+
+srun bash -c "cd /data/$SLURMD_NODENAME/merljoha/$current_branch; nix develop --command stack --no-nix --system-ghc --no-install-ghc build"

lambda/src/LambdaDatasets/GermanDataset.hs

@@ -86,13 +86,13 @@ lE =
             ((Ref.SomeTypeRep (Ref.TypeRep @(Job))), [(fmap show (enumUniform UnemployedOrUnskilledNonResident HighlySkilled ))])
           ],
       targetType = (Ref.SomeTypeRep (Ref.TypeRep @(AccountStatus -> Int -> CreditHistory -> Purpose -> Int -> Savings -> EmploymentStatus -> Int -> StatusAndSex -> OtherDebtors -> Int -> Property -> Int -> OtherPlans -> Housing -> Int -> Job -> Int -> Bool -> Bool -> GermanClass))),
-      maxDepth = 8,
+      maxDepth = 9,
       weights =
         ExpressionWeights
-          { lambdaSpucker = 10,
+          { lambdaSpucker = 0,
-            lambdaSchlucker = 1,
+            lambdaSchlucker = 10,
-            symbol = 20,
+            symbol = 100,
-            variable = 100,
+            variable = 5,
             constant = 5
           }
     }
@@ -151,7 +151,6 @@ data LamdaExecutionEnv = LamdaExecutionEnv
 data FittnesRes = FittnesRes
   { total :: R,
     fitnessTotal :: R,
-    costAccordingToDataset :: N,
     fitnessGeoMean :: R,
     fitnessMean :: R,
     accuracy :: R,
@@ -190,9 +189,8 @@ evalResults ex trs = do
 evalResult :: LamdaExecutionEnv -> TypeRequester -> (AccountStatus -> Int -> CreditHistory -> Purpose -> Int -> Savings -> EmploymentStatus -> Int -> StatusAndSex -> OtherDebtors -> Int -> Property -> Int -> OtherPlans -> Housing -> Int -> Job -> Int -> Bool -> Bool -> GermanClass) -> (TypeRequester, FittnesRes)
 evalResult ex tr result = ( tr,
       FittnesRes
-        { total = (biasSmall - 1) - (fromIntegral costAccordingToDS),
+        { total = score,
           fitnessTotal = fitness',
-          costAccordingToDataset = costAccordingToDS,
           fitnessMean = meanOfAccuricyPerClass resAndTarget,
           fitnessGeoMean = geomeanOfDistributionAccuracy resAndTarget,
           accuracy = acc,
@@ -203,8 +201,7 @@ evalResult ex tr result = ( tr,
     where
     res = map (\(a, b, c, d, e, f, g, h, i, j, k, l, m, n, o, p, q, r, s, t) -> result a b c d e f g h i j k l m n o p q r s t) (fst (dset ex))
     resAndTarget = (zip (snd (dset ex)) res)
-    acc = (foldr (\(actual,predicted) s -> if (actual == predicted) then s + 1 else s) 0 resAndTarget) / fromIntegral (length resAndTarget)
+    acc = (foldr (\ts s -> if ((fst ts) == (snd ts)) then s + 1 else s) 0 resAndTarget) / fromIntegral (length resAndTarget)
-    costAccordingToDS = (foldr (\(actual,predicted) s -> if ((actual) == (predicted)) then s else (if actual == Deny then s+5 else s+1)) 0 resAndTarget)
     biasSmall = exp ((-(fromIntegral (countTrsR tr))) / 1000) -- 0 (schlecht) bis 1 (gut)
     fitness' = meanOfAccuricyPerClass resAndTarget
     score = fitness' + (biasSmall - 1)

lambda/src/LambdaDatasets/IrisDataset.hs

@@ -53,13 +53,13 @@ lE =
             ((Ref.SomeTypeRep (Ref.TypeRep @(IrisClass))), [(fmap show (enumUniform Setosa Versicolor :: RVar IrisClass))])
           ],
       targetType = (Ref.SomeTypeRep (Ref.TypeRep @(Float -> Float -> Float -> Float -> IrisClass))),
-      maxDepth = 8,
+      maxDepth = 9,
       weights =
         ExpressionWeights
-          { lambdaSpucker = 10,
+          { lambdaSpucker = 0,
-            lambdaSchlucker = 1,
+            lambdaSchlucker = 10,
-            symbol = 20,
+            symbol = 100,
-            variable = 100,
+            variable = 5,
             constant = 5
           }
     }
@@ -155,7 +155,7 @@ evalResults ex trs = do
 evalResult :: LamdaExecutionEnv -> TypeRequester -> (Float -> Float -> Float -> Float -> IrisClass) -> (TypeRequester, FittnesRes)
 evalResult ex tr result = ( tr,
       FittnesRes
-        { total = acc * 100 + (biasSmall - 1),
+        { total = score,
           fitnessTotal = fitness',
           fitnessMean = meanOfAccuricyPerClass resAndTarget,
           fitnessGeoMean = geomeanOfDistributionAccuracy resAndTarget,

lambda/src/LambdaDatasets/NurseryDataset.hs

@@ -74,13 +74,13 @@ lE =
             ((Ref.SomeTypeRep (Ref.TypeRep @(Health))), [(fmap show (enumUniform NotRecommendHealth PriorityHealth ))])
           ],
       targetType = (Ref.SomeTypeRep (Ref.TypeRep @(Parents -> HasNurs -> Form -> Children -> Housing -> Finance -> Social -> Health -> NurseryClass))),
-      maxDepth = 8,
+      maxDepth = 9,
       weights =
         ExpressionWeights
-          { lambdaSpucker = 10,
+          { lambdaSpucker = 0,
-            lambdaSchlucker = 1,
+            lambdaSchlucker = 10,
-            symbol = 20,
+            symbol = 100,
-            variable = 100,
+            variable = 5,
             constant = 5
           }
     }

lambda/src/Main.hs

@@ -8,9 +8,9 @@ import Pipes
 import Pretty
 import Protolude hiding (for)
 import System.IO
--- import LambdaDatasets.IrisDataset
+import LambdaDatasets.IrisDataset
 -- import LambdaDatasets.NurseryDataset
-import LambdaDatasets.GermanDataset
+-- import LambdaDatasets.GermanDataset
 import Debug.Trace as DB
 import qualified Data.Map.Strict as Map
 

run.sbatch

@@ -1,9 +1,31 @@
 #!/usr/bin/env bash
-#SBATCH --time=18:00:00
+# test this timing, it scales with result sizes
+#SBATCH --time=12:00:00
 #SBATCH --partition=cpu
-#SBATCH --array=0-30
+# 30 Runs * 9 Experiments * 3 Datasets
-#SBATCH --output=./output/output_run_%j.txt
+#SBATCH --array=0-809
-#SBATCH --error=./output/error_run_%j.txt
+# ensure output exists, is a folder and is writable in your working directory
-#SBATCH --nodelist=oc-compute02
+#SBATCH --output=./output/output_run_%a.txt
-#SBATCH --mem=3G
+#SBATCH --error=./output/error_run_%a.txt
-srun nix develop --command stack --no-nix --system-ghc --no-install-ghc run haga-lambda
+# exclude nodes with weaker CPUs
+#SBATCH --exclude=oc222
+# test memory usage, it scales **Exponentially** with max Depth. Implement some countermeasures if that's a problem, e.g. raise max depth over time.
+#SBATCH --mem=6G
+#SBATCH --nodes=1
+
+
+# list your branches
+problems=("iris" "nurse" "german")
+
+# 30 Runs * 9 Experiments
+current_problem=${problems[(${SLURM_ARRAY_TASK_ID}/270)]}
+# 30 Runs, 9 Experiments
+current_variant=$(((${SLURM_ARRAY_TASK_ID} / 30) % 9 + 1))
+current_branch="${current_problem}_${current_variant}"
+
+# ensure [full path to writable folder on node *] exists
+git clone -b $current_branch --single-branch "[your git repo]" [full path to writable folder on node 1]/$current_branch
+git clone -b $current_branch --single-branch "[your git repo]" [full path to writable folder on node 2]/$current_branch
+#... for every node
+
+srun bash -c "cd /data/$SLURMD_NODENAME/merljoha/$current_branch; nix develop --command stack --no-nix --system-ghc --no-install-ghc run haga-lambda"