fix(notmnist): notmnist sigmoid chart with 10 epochs

examples/notmnist.hs

@@ -43,7 +43,7 @@ module Main where
 
     let session = def { learningRate = 0.5
                       , batchSize    = 32
-                      , epochs       = 9
+                      , epochs       = 10
                       , network      = rnetwork
                       , training     = zip trinputs trlabels
                       , test         = zip teinputs telabels

src/Sibe.hs

@@ -84,6 +84,7 @@ module Sibe
                               , batchSize    :: Int
                               , chart        :: [(Int, Double, Double)]
                               , momentum     :: Double
+                              , biases       :: Bool
                               } deriving (Show)
 
       emptyNetwork = randomNetwork 0 (0, 0) 0 [] (0, (id, id))
@@ -97,6 +98,7 @@ module Sibe
                       , batchSize    = 0
                       , chart        = []
                       , momentum     = 0
+                      , biases       = True
                       }
 
       saveNetwork :: Network -> String -> IO ()
@@ -122,6 +124,9 @@ module Sibe
       runLayer :: Input -> Layer -> Output
       runLayer input (Layer !biases !weights _) = input <# weights + biases
 
+      runLayer' :: Input -> Layer -> Output
+      runLayer' input (Layer !biases !weights _) = input <# weights
+
       forward :: Input -> Session -> Output
       forward input session = compute input (network session)
         where

src/Sibe/Word2Vec.hs

@@ -26,14 +26,16 @@ module Sibe.Word2Vec
       return trainingData
       let s = session { training = trainingData
                       , network = buildNetwork 0 (-1, 1) v [(v, 25, (id, one))] (20, v, (softmax, crossEntropy'))
+                      , biases = False
                       }
       print trainingData
-      newses <- run (gd . learningRateDecay (1.1, 0.1) . ignoreBiases) s
+      newses <- run (gd . learningRateDecay (1.1, 0.1)) s
 
       let (hidden@(Layer biases nodes _) :- _) = network newses
       {-let computedVocVec = map (\(w, v) -> (w, forward v newses)) vocvec-}
-      let computedVocVec = map (\(w, v) -> (w, (fromRows [v]) <> nodes)) vocvec
-      print computedVocVec
+      print biases
+      let computedVocVec = map (\(w, v) -> (w, v <# nodes)) vocvec
+      {-print computedVocVec-}
 
       {-mapM_ (\(w, v) -> do
                       putStr $ w ++ ": "