chore: Update California housing example to work with hasktorch.

Author: mchav

dataframe.cabal

@@ -141,6 +141,7 @@ executable california_housing
                       bytestring >= 0.11 && <= 0.12.2.0,
                       containers >= 0.6.7 && < 0.8,
                       directory >= 1.3.0.0 && <= 1.3.9.0,
+                      hasktorch,
                       hashable >= 1.2 && <= 1.5.0.0,
                       statistics >= 0.16.2.1 && <= 0.16.3.0,
                       template-haskell >= 2.0 && <= 2.30,

examples/CaliforniaHousing.hs

@@ -1,14 +1,67 @@
-{-# LANGUAGE OverloadedStrings #-}
+{-# LANGUAGE OverloadedStrings   #-}
+{-# LANGUAGE TypeApplications    #-}
+{-# LANGUAGE ScopedTypeVariables #-}
+{-# LANGUAGE NumericUnderscores #-}
 module Main where
 
+import Control.Monad (when)
+import Data.Maybe
+import qualified Data.Map as M
+import qualified Data.Text as T
 import qualified DataFrame as D
+import qualified DataFrame.Functions as F
+import qualified Data.Vector.Unboxed as VU
+import qualified Data.Vector as V
+import Torch
+
+import DataFrame ((|>))
 
 main :: IO ()
 main = do
-  parsed <- D.readCsv "./data/housing.csv"
+    {- Feature ingestion and engineering -}
+    df <- fmap (D.apply (\(op :: T.Text) -> oceanProximity M.! op) "ocean_proximity") (D.readCsv "./data/housing.csv")
+    -- This column has nulls so we:
+    --  * Remove all nulls with filterJust
+    --  * Calculate the mean of total_bedrooms
+    --  * impute the mean.
+    -- This could probably be a utility function.
+    let meanTotalBedrooms = fromMaybe 0 $ df |> D.filterJust "total_bedrooms"
+                                             |> D.mean "total_bedrooms"
+        imputed = df |> D.impute "total_bedrooms" meanTotalBedrooms
+                     |> D.exclude ["median_house_value"]
+                     |> normalizeFeatures
+        (r, c) = D.dimensions imputed
+        features = reshape [r,c] $ asTensor (flattenFeatures imputed)
+        labels = asTensor ((VU.map realToFrac . VU.convert) (D.columnAsVector @Double "median_house_value" df) :: VU.Vector Float)
+    
+    {- Train the model -}
+    putStrLn "Training linear regression model..."
+    init <- sample $ LinearSpec{in_features = (snd (D.dimensions df) - 1), out_features = 1}
+    trained <- foldLoop init 100_000 $ \state i -> do
+        let labels' = model state features
+            loss = mseLoss labels labels'
+        when (i `mod` 10_000 == 0) $ do
+            putStrLn $ "Iteration: " ++ show i ++ " | Loss: " ++ show loss
+        (state', _) <- runStep state GD loss 0.1
+        pure state'
+    
+    {- Show predictions -}
+    let predictions = D.insertUnboxedVector "predicted_house_value" (asValue @(VU.Vector Float) (model trained features)) df
+    print $ D.select ["median_house_value", "predicted_house_value"] predictions |> D.take 10
+
+normalizeFeatures :: D.DataFrame -> D.DataFrame
+normalizeFeatures df = df |> D.fold (\name d -> let
+    m      = fromMaybe 0 (D.mean name d)
+    stdDev = fromMaybe 0.01 (D.standardDeviation name d)
+    col    = F.col @Double name
+  in D.derive name ((col - (F.minimum col)) / (F.maximum col - F.minimum col)) d) (D.columnNames df)
+
 
-  print $ D.describeColumns parsed
+model :: Linear -> Tensor -> Tensor
+model state input = squeezeAll $ linear state input
 
-  print $ D.take 5 parsed
+oceanProximity :: M.Map T.Text Double
+oceanProximity = M.fromList [("ISLAND", 0), ("NEAR OCEAN", 1), ("NEAR BAY", 2), ("<1H OCEAN", 3), ("INLAND", 4)] 
 
-  D.plotHistograms D.PlotAll D.VerticalHistogram parsed
+flattenFeatures :: D.DataFrame -> VU.Vector Float
+flattenFeatures df = V.foldl' (\acc v -> acc VU.++ v) VU.empty (D.toMatrix df)

examples/README.md

@@ -0,0 +1,42 @@
+# Running the examples
+
+## California housing
+
+Preparation:
+This has a hasktorch integration that requires some setup. Copy the [`get-deps.sh`](https://github.com/hasktorch/hasktorch/blob/master/deps/get-deps.sh) file into the dataframe home directory. This will download and link some pytorch files required to run the examples.
+
+After this is done you'll need to run `./set_hasktorch_env` to put the hasktorch libraries in your `LD_LIBRARY_PATH`.
+
+Running:
+`cabal run california_housing`.
+
+Expected output:
+
+```
+Training linear regression model...
+Iteration: 10000 | Loss: Tensor Float []  5.0225e9   
+Iteration: 20000 | Loss: Tensor Float []  4.9093e9   
+Iteration: 30000 | Loss: Tensor Float []  4.8576e9   
+Iteration: 40000 | Loss: Tensor Float []  4.8333e9   
+Iteration: 50000 | Loss: Tensor Float []  4.8217e9   
+Iteration: 60000 | Loss: Tensor Float []  4.8160e9   
+Iteration: 70000 | Loss: Tensor Float []  4.8130e9   
+Iteration: 80000 | Loss: Tensor Float []  4.8114e9   
+Iteration: 90000 | Loss: Tensor Float []  4.8105e9   
+Iteration: 100000 | Loss: Tensor Float []  4.8099e9   
+--------------------------------------------------
+index | median_house_value | predicted_house_value
+------|--------------------|----------------------
+ Int  |       Double       |         Float        
+------|--------------------|----------------------
+0     | 452600.0           | 414079.94            
+1     | 358500.0           | 423011.94            
+2     | 352100.0           | 383239.06            
+3     | 341300.0           | 324928.94            
+4     | 342200.0           | 256934.23            
+5     | 269700.0           | 264944.84            
+6     | 299200.0           | 259094.13            
+7     | 241400.0           | 257224.55            
+8     | 226700.0           | 201753.69            
+9     | 261100.0           | 268698.7
+```

set_hasktorch_env

@@ -0,0 +1,28 @@
+#!/usr/bin/env bash
+
+if [ -z "${XDG_CACHE_HOME:-}" ]; then
+  export XDG_CACHE_HOME="$HOME/.cache"
+fi
+
+# execute this command if needed when building on OSX if there are linker errors.
+# dylib files in extra-lib-dirs  don't get forwarded to ghc
+# in some versions of OSX. See https://github.com/commercialhaskell/stack/issues/1826
+HASKTORCH_LIB_PATH="$XDG_CACHE_HOME/libtorch/lib:$XDG_CACHE_HOME/mklml/lib/:$XDG_CACHE_HOME/libtokenizers/lib"
+
+function add_vendor_lib_path {
+    case "$(uname)" in
+      "Darwin")
+        DYLD_LIBRARY_PATH=/opt/homebrew/lib:/opt/homebrew/opt/libomp/lib:$HASKTORCH_LIB_PATH:$DYLD_LIBRARY_PATH
+        export DYLD_LIBRARY_PATH
+        ;;
+      "Linux"|"FreeBSD")
+        LD_LIBRARY_PATH=$HASKTORCH_LIB_PATH:$LD_LIBRARY_PATH
+        export LD_LIBRARY_PATH
+        ;;
+      *)
+        echo "OS doesn't have known environment variable hacks to set"
+        ;;
+    esac
+}
+
+add_vendor_lib_path