neural-network

Neural network framework in Haskell

Packages included

• Base. This package defines the abstract neural-network, and a extendable specification of layers.
• Backend-hmatrix. This package implements the full-connect layer and convolution layer based on hmatrix library. It has a simple and plain representation but some issues in both time and space efficiency.
• Backend-blashs. This package implements the full-connect layer and convolution layer based on blas-hs library. A imperative interface for manipulating dense vector and matrix is devised for better storage utilization.

Build with stack tool

• Please see https://docs.haskellstack.org/en/stable/README/

Additional notes on build

Linux

• with openblas flag true in the flags section, please install the openblas by the official package management.
  • or else, install blas/lapack package.

Windows

• Download OpenBLAS from http://www.openblas.net/
• Modify the following fields in the stack.yaml
  • extra-include-dirs: path-to-include-dir-of-openblas
  • extra-lib-dirs: path-to-lib-dir-of-openblas

Utilizing SIMD

• The vec128 flag for neural-network-blashs can be turned on, and many operations will utilize SIMD for better performance.

    neural-network-blashs:
      vec128: true

• The flags vec256 and vec512 cause segment-fault for the moment.

• A known bug on windows. vec128 implies compiler option -fllvm for ghc. However due to a known bug of binutils on mingw-w64, this option leads to a segment fault

  • mingw-w64-x86_64-binutils < 2.27-2
  • ghc <= 8.0.1 (because it is bundled with old binutils)
  • stack resolver <= lts-7.14 (because it imples ghc <= 8.0.1)
  • See bug report here https://ghc.haskell.org/trac/ghc/ticket/8974