hasktorch-naperian

This README and further docs are in progress.

Goal

To provide a finite Naperian functor based interface to hasktorch.

Install and build

To build, clone Hasktorch 0.2.0.0 into the same parent directory and build. Then run stack build in the top level of this project. The resulting binary requires some shared objects. One way to run the binary is to copy it to the top-level hasktorch directory and execute source setenv while in bash.

Status

The official Google Summer of Code project has ended.

Finite Naperian functors

The module Data.Naperian contains:

• Instances of Enum and Bounded for (a, b) and Either a b.
• Naperian instances for Product and Compose.
• A newtype wrapper around Naperian to derive Applicative.
• The definition of FiniteNaperian.
• FiniteNaperian instances for Product and Compose.
• A newtype wrapper around FiniteNaperian to derive Foldable, Dimension, Eq1, Ord1, and Show1. Also included is a default definition of traverse to define Traversable (which cannot be derived with DerivingVia).
• The definition of FiniteHyper multidimensional version of FiniteNaperian, which is indexed by a heterogeneous list HList [a1, ..., an].
• Various instances for HList [a1, ..., an] including Bounded and Enum.
• Instances for FiniteHyper including FiniteNaperian.

The module Data.Naperian.Tree implements some type indexed trees and gives their finite Naperian instances. Additionally, it uses the category of indexed functors to define a form of catamorphisms. Finally, it uses some extra structure of the trees to allow a mapAccum style function to be defined.

The module Data.Naperian.Examples shows how to define your own finite Naperian functor.

Working with Torch

The module Torch.Naperian shows some functions which are similar to Torch functions, and are implemented in pure Haskell. The main portion of the is the Dim data type and functions for interacting with it.

The module Torch.Naperian.Examples shows how to perform a simple linear regression using finite Naperian functors and the ad library for it's gradient descent function. It also uses indexed trees to create a Tree LSTM.