{-# LANGUAGE FlexibleContexts #-} ----------------------------------------------------------------------------- -- | -- Module : Data.Packed.Array.Decomposition -- Copyright : (c) Alberto Ruiz 2009 -- License : BSD3 -- Maintainer : Alberto Ruiz -- Stability : experimental -- -- Common multidimensional array decompositions. See the paper by Kolda & Balder. -- ----------------------------------------------------------------------------- module Numeric.LinearAlgebra.Array.Decomposition ( -- * HOSVD hosvd, hosvd', truncateFactors, -- * CP cpAuto, cpRun, cpInitRandom, cpInitSvd, -- * Utilities ALSParam(..), defaultParameters ) where import Numeric.LinearAlgebra.Array import Numeric.LinearAlgebra.Array.Internal(seqIdx,namesR,sizesR,renameRaw) import Numeric.LinearAlgebra.Array.Util import Numeric.LinearAlgebra.Array.Solve import Numeric.LinearAlgebra.HMatrix hiding (scalar) import Data.List import System.Random --import Control.Parallel.Strategies {- | Full version of 'hosvd'. The first element in the result pair is a list with the core (head) and rotations so that t == product (fst (hsvd' t)). The second element is a list of rank and singular values along each mode, to give some idea about core structure. -} hosvd' :: Array Double -> ([Array Double],[(Int,Vector Double)]) hosvd' t = (factors,ss) where (rs,ss) = unzip $ map usOfSVD $ flats t n = length rs dummies = take n $ seqIdx (2*n) "" \\ (namesR t) axs = zipWith (\a b->[a,b]) dummies (namesR t) factors = renameRaw core dummies : zipWith renameRaw (map (fromMatrix None None . tr) rs) axs core = product $ renameRaw t dummies : zipWith renameRaw (map (fromMatrix None None) rs) axs {- | Multilinear Singular Value Decomposition (or Tucker's method, see Lathauwer et al.). The result is a list with the core (head) and rotations so that t == product (hsvd t). The core and the rotations are truncated to the rank of each mode. Use 'hosvd'' to get full transformations and rank information about each mode. -} hosvd :: Array Double -> [Array Double] hosvd a = truncateFactors rs h where (h,info) = hosvd' a rs = map fst info -- get the matrices of the flattened tensor for all dimensions flats t = map (flip fibers t) (namesR t) --check trans/ctrans usOfSVD m = if rows m < cols m then let (s2,u) = eigSH' $ m <> tr m s = sqrt (abs s2) in (u,r s) else let (s2,v) = eigSH' $ tr m <> m s = sqrt (abs s2) u = m <> v <> pinv (diag s) in (u,r s) where r s = (ranksv (sqrt peps) (max (rows m) (cols m)) (toList s), s) -- (rank m, sv m) where sv m = s where (_,s,_) = svd m ttake ns t = (foldl1' (.) $ zipWith (onIndex.take) ns (namesR t)) t -- | Truncate a 'hosvd' decomposition from the desired number of principal components in each dimension. truncateFactors :: [Int] -> [Array Double] -> [Array Double] truncateFactors _ [] = [] truncateFactors ns (c:rs) = ttake ns c : zipWith f rs ns where f r n = onIndex (take n) (head (namesR r)) r ------------------------------------------------------------------------ frobT = norm_2 . coords ------------------------------------------------------------------------ unitRows [] = error "unitRows []" unitRows (c:as) = foldl1' (.*) (c:xs) : as' where (xs,as') = unzip (map g as) g a = (x,a') where n = head (namesR a) -- hmmm rs = parts a n scs = map frobT rs x = diagT scs (order c) `renameRaw` (namesR c) a' = (zipWith (.*) (map (scalar.recip) scs)) `onIndex` n $ a {- | Basic CP optimization for a given rank. The result includes the obtained sequence of errors. For example, a rank 3 approximation can be obtained as follows, where initialization is based on the hosvd: @ (y,errs) = cpRank 3 t where cpRank r t = cpRun (cpInitSvd (fst $ hosvd' t) r) defaultParameters t @ -} cpRun :: [Array Double] -- ^ starting point -> ALSParam None Double -- ^ optimization parameters -> Array Double -- ^ input array -> ([Array Double], [Double]) -- ^ factors and error history cpRun s0 params t = (unitRows $ head s0 : sol, errs) where (sol,errs) = mlSolve params [head s0] (tail s0) t {- | Experimental implementation of the CP decomposition, based on alternating least squares. We try approximations of increasing rank, until the relative reconstruction error is below a desired percent of Frobenius norm (epsilon). The approximation of rank k is abandoned if the error does not decrease at least delta% in an iteration. Practical usage can be based on something like this: @ cp finit d e t = cpAuto (finit t) defaultParameters {delta = d, epsilon = e} t cpS = cp (InitSvd . fst . hosvd') cpR s = cp (cpInitRandom s) @ So we can write @ \-\- initialization based on hosvd y = cpS 0.01 1E-6 t \-\- (pseudo)random initialization z = cpR seed 0.1 0.1 t @ -} cpAuto :: (Int -> [Array Double]) -- ^ Initialization function for each rank -> ALSParam None Double -- ^ optimization parameters -> Array Double -- ^ input array -> [Array Double] -- ^ factors cpAuto finit params t = fst . head . filter ((<epsilon params). head . snd) . map (\r->cpRun (finit r) params t) $ [1 ..] ---------------------- -- | cp initialization based on the hosvd cpInitSvd :: [NArray None Double] -- ^ hosvd decomposition of the target array -> Int -- ^ rank -> [NArray None Double] -- ^ starting point cpInitSvd (hos) k = d:as where c:rs = hos as = trunc (replicate (order c) k) rs d = diagT (replicate k 1) (order c) `renameO` (namesR c) trunc ns xs = zipWith f xs ns where f r n = onIndex (take n . cycle) (head (namesR r)) r cpInitSeq rs t k = ones:as where n = order t auxIndx = take n $ seqIdx (2*n) "" \\ namesR t --take (order t) $ map return ['a'..] \\ namesR t ones = diagT (replicate k 1) (order t) `renameO` auxIndx ts = takes (map (*k) (sizesR t)) rs as = zipWith4 f ts auxIndx (namesR t) (sizesR t) f c n1 n2 p = (listArray [k,p] c) `renameO` [n1,n2] takes [] _ = [] takes (n:ns) xs = take n xs : takes ns (drop n xs) -- | pseudorandom cp initialization from a given seed cpInitRandom :: Int -- ^ seed -> NArray i t -- ^ target array to decompose -> Int -- ^ rank -> [NArray None Double] -- ^ random starting point cpInitRandom seed = cpInitSeq (randomRs (-1,1) (mkStdGen seed)) ----------------------------------------------------------------------