variation 4

OnRunning.md

@@ -1,27 +0,0 @@
-# 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

@@ -1,28 +1,9 @@
 #!/usr/bin/env bash
 #SBATCH --time=00:10:00
 #SBATCH --partition=cpu
-# 9 Experiments * 3 Datasets
+#SBATCH --output=./output/output_build.txt
-#SBATCH --array=0-27
+#SBATCH --error=./output/error_build.txt
-# 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=2G
+#SBATCH --mem=4G
-
+#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 = 9,
+      maxDepth = 8,
       weights =
         ExpressionWeights
-          { lambdaSpucker = 0,
+          { lambdaSpucker = 10,
-            lambdaSchlucker = 10,
+            lambdaSchlucker = 1,
-            symbol = 100,
+            symbol = 20,
-            variable = 5,
+            variable = 100,
             constant = 5
           }
     }

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 = 9,
+      maxDepth = 8,
       weights =
         ExpressionWeights
-          { lambdaSpucker = 0,
+          { lambdaSpucker = 10,
-            lambdaSchlucker = 10,
+            lambdaSchlucker = 1,
-            symbol = 100,
+            symbol = 20,
-            variable = 5,
+            variable = 100,
             constant = 5
           }
     }
@@ -68,7 +68,7 @@ lEE :: LamdaExecutionEnv
 lEE =
   LamdaExecutionEnv
     { -- For now these need to define all available functions and types. Generic functions can be used.
-      imports = ["LambdaDatasets.IrisDefinition"],
+      imports = ["LambdaDatasets.IrisDataset"],
       training = True,
       trainingData =
         ( map fst (takeFraktion 0.8 irisTrainingData),
@@ -89,7 +89,7 @@ shuffledLEE = do
   itD <- smpl $ shuffle irisTrainingData
   return  LamdaExecutionEnv
     { -- For now these need to define all available functions and types. Generic functions can be used.
-      imports = ["LambdaDatasets.IrisDefinition"],
+      imports = ["LambdaDatasets.IrisDataset"],
       training = True,
       trainingData =
         ( map fst (takeFraktion 0.8 itD),

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 = 9,
+      maxDepth = 8,
       weights =
         ExpressionWeights
-          { lambdaSpucker = 0,
+          { lambdaSpucker = 10,
-            lambdaSchlucker = 10,
+            lambdaSchlucker = 1,
-            symbol = 100,
+            symbol = 20,
-            variable = 5,
+            variable = 100,
             constant = 5
           }
     }

lambda/src/Main.hs

@@ -35,7 +35,7 @@ options =
       ( long "population-size"
           <> short 'p'
           <> metavar "N"
-          <> value 100
+          <> value 400
           <> help "Population size"
       )
 
@@ -59,7 +59,7 @@ main =
       selectionType = Tournament 3,
       termination = (steps (iterations opts)),
       poulationSize = (populationSize opts),
-      stepSize = 90,
+      stepSize = 120,
       elitismRatio = 5/100
     }
     pop' <- runEffect (for (run cfg) logCsv)

run.sbatch

@@ -1,31 +1,9 @@
 #!/usr/bin/env bash
-# test this timing, it scales with result sizes
+#SBATCH --time=18:00:00
-#SBATCH --time=12:00:00
 #SBATCH --partition=cpu
-# 30 Runs * 9 Experiments * 3 Datasets
+#SBATCH --array=0-30
-#SBATCH --array=0-809
+#SBATCH --output=./output/output_run_%j.txt
-# ensure output exists, is a folder and is writable in your working directory
+#SBATCH --error=./output/error_run_%j.txt
-#SBATCH --output=./output/output_run_%a.txt
+#SBATCH --nodelist=oc-compute02
-#SBATCH --error=./output/error_run_%a.txt
+#SBATCH --mem=3G
-# exclude nodes with weaker CPUs
+srun nix develop --command stack --no-nix --system-ghc --no-install-ghc run haga-lambda
-#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"