Swap .lhs modules for .hs modules

COMP4027.cabal

@@ -30,6 +30,7 @@ library
   build-depends:
       ad ==4.5.2
     , base ^>=4.14.1.0
+    , extra ==1.7.11
     , random ==1.2.0
     , JuicyPixels ==3.3.5
     , JuicyPixels-util ==0.2

src/MLP_testing_MNIST.hs

@@ -0,0 +1,52 @@
+module MLP_testing_MNIST where
+
+{- |
+  In this module are a number of functions that have been used to test models
+  that have been trained on the MNIST dataset.
+
+  The functions below have been defined to test models that have been trained
+  on the MNIST dataset. The testMNIST function obtains the MNIST test dataset
+  and uses this to test the model. In testModel, each sample is tested to see
+  if the one hot encoded prediction is the same as the target, if so then 1
+  is added to a sum, otherwise 0 is added. The final sum tells us how many
+  samples (out of 10000) the model predicted correctly.
+-}
+
+import Data.Bool (bool)
+import Data.Tuple.Extra (both)
+import Data.Time.Clock -- clock to measure convergence time
+import Types -- all custom types/classes for parallel multi-layer perceptrons
+import MNIST -- MNIST dataset for target function. Source: https://git.thm.de/akwl20/neural-network-haskell
+import Data.Matrix as Matrix -- toList operation for matrices
+import MLP_utils -- for feedFrwd and toLoss
+import qualified Data.Text.IO as TXTIO
+import qualified Data.Text as TXT
+
+-- Returns bool indicating whether the sample was predicted correctly.
+testSample :: [UnactivatedLayer] -> ([Double],[Double]) -> Bool
+testSample model (input,target) =
+  target == (predictClass . outputs . head $ feedFrwd model input)
+  where
+  predictClass :: [Double] -> [Double]
+  predictClass xs = [if x == maximum xs then 1.0 else 0.0 | x <- xs]
+  -- ^ Converts a softmax output to a one hot vector prediction.
+
+-- | Returns the number of samples predicted correctly.
+testModel :: [UnactivatedLayer] -> [([Double], [Double])] -> Int
+testModel model testData = sum $ map (bool 0 1 . testSample model) testData
+
+-- | Measures the time taken to test a model that has been trained on the MNIST
+-- dataset. Also prints how many samples out of the 10000-sample MNIST test-set
+-- the model predicted correctly.
+timedRun :: IO ()
+timedRun = do
+  let preprocess :: (Real a, Fractional b) => [(Matrix a, Matrix a)] -> [([b], [b])]
+      preprocess = map $ both $ map realToFrac . Matrix.toList
+  model <- read . TXT.unpack <$> TXTIO.readFile "trained_model" :: IO [UnactivatedLayer]
+  testData <- preprocess <$> getTestSamples
+  t0 <- getCurrentTime
+  let numTruePositives = testModel model testData
+  -- Force evaluation by printing the # of true positives before taking the time:
+  putStrLn ("Accuracy: " ++ show numTruePositives ++ " out of " ++ show (length testData))
+  t1 <- getCurrentTime
+  putStrLn ("Time to test: " ++ show (diffUTCTime t1 t0))