https://github.com/pierric/neural-network

Neural network framework in Haskell
https://github.com/pierric/neural-network

haskell neural-network

Last synced: 2 months ago
JSON representation

Neural network framework in Haskell

• Host: GitHub
• URL: https://github.com/pierric/neural-network
• Owner: pierric
• License: bsd-3-clause
• Created: 2016-11-28T10:25:22.000Z (about 8 years ago)
• Default Branch: master
• Last Pushed: 2017-10-22T09:20:12.000Z (about 7 years ago)
• Last Synced: 2024-10-28T14:18:44.388Z (3 months ago)
• Topics: haskell, neural-network
• Language: Haskell
• Size: 279 KB
• Stars: 5
• Watchers: 4
• Forks: 0
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# 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.

```yaml

    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