diff options
| -rw-r--r-- | mnist/Neuronet.hs | 33 | ||||
| -rw-r--r-- | mnist/main.hs | 4 |
2 files changed, 11 insertions, 26 deletions
diff --git a/mnist/Neuronet.hs b/mnist/Neuronet.hs index ece288a..6c3ea32 100644 --- a/mnist/Neuronet.hs +++ b/mnist/Neuronet.hs @@ -1,18 +1,3 @@ -{- | - - N E U R O T I C U S - - A small and straightforward neural network coded in Haskell - from scratch. It uses gradient descent and the beauty of - backpropagation for learning. - - TODOS: improve performance, parallelize, CUDA, FFI - - Michal Idziorek <m.i@gmx.at> - March 2019 - --} - module Neuronet ( Neuronet -- the neuronet ,neuronet -- initalize neuronet @@ -20,15 +5,16 @@ module Neuronet ,trainBatch -- train with batch ,asknet -- ask the neuroal net - ,wghtact + ,wghtact ,backprop )where -import Numeric.LinearAlgebra (Matrix,Vector,tr,scale,cmap,(#>),randn,toList,fromList,toLists,fromLists,outer) import Data.List +import Numeric.LinearAlgebra (Matrix,Vector,tr,scale,cmap,(#>),randn, + toList,fromList,toLists,fromLists,outer) --- | A layer of our network consists of a weight matrix with input weights --- and a vector holding the bias at each neuron. +-- | A layer of our network consists of a weight matrix with input +-- weights and a vector holding the bias at each neuron. type Layer = (Matrix Double,Vector Double) -- | Our neural network is simply a list of layers @@ -39,15 +25,15 @@ type Neuronet = [Layer] -- using gaussian distribution with mean 0 and standard deviation 1. neuronet :: [Int] -> IO Neuronet neuronet l = mapM nl $ zip l (tail l) - where nl (i,l) = (,) <$> randn l i <*> (randn 1 l >>= return.fromList.head.toLists) + where nl (i,l) = (,) <$> randn l i <*> + (randn 1 l >>= return.fromList.head.toLists) -- | Given the input vector calculate the `weighted inputs` and -- `activations` for all layers of our network. wghtact :: Neuronet -> Vector Double -> [(Vector Double,Vector Double)] wghtact [] _ = [] -wghtact ((w,b):lx) x = (z,a):wghtact lx a - where z = w #> x + b - a = cmap sigmoid z +wghtact ((w,b):lx) x = (z,a):wghtact lx a where z = w #> x + b + a = cmap sigmoid z -- | Given the input vector calculate the final output asknet :: Neuronet -> Vector Double -> Vector Double @@ -90,4 +76,3 @@ trainBatch r net xs ys = zipWith (upd r) net bp -- | Update a single Layer given the `direction` and `training rate` upd :: Double -> Layer -> (Matrix Double,Vector Double) -> Layer upd r (a,b) (c,d) = (a-scale r c,b-scale r d) - diff --git a/mnist/main.hs b/mnist/main.hs index 2810e34..02ba6f7 100644 --- a/mnist/main.hs +++ b/mnist/main.hs @@ -59,9 +59,9 @@ read_samples f1 f2 = do mainMNIST :: IO () mainMNIST =do - let ep = 20 -- number of epochs + let ep = 10 -- number of epochs let mbs = 10 -- mini-batch size - let lr = 2 -- learning rate + let lr = 3 -- learning rate let cap = 999999 -- cap number of training samples putStrLn "= Init =" |