This is an Automatic Differentiation library based on the paper "Provably correct, asymptotically efficient, higher-order reverse-mode automatic differentiation" by Krawiec, Krishnaswami, Peyton Jones, Ellis, Fitzgibbon, and Eisenberg, developed in collaboration with the paper's authors.
This is an early prototype, both in terms of the engine performance and the API augmented with a set of tools and examples. The user should be ready to add missing primitives, as well as any obvious tools that should be predefined but aren't. You can already differentiate all basic neural network architectures, such as fully connected, recurrent, convolutional and residual networks. The library, due to its loose coupling of differentiation and data containers, can naturally handle exotic variants of such networks that wouldn't express well in a limited language of matrices or tensors.
Applications outside machine learning are very much in scope, given that the notion of a neural network is not hardwired into the formalism, but ad hoc redefined from basic building blocks of general automatic differentiation whenever it's needed.
Mature Haskell libraries with similar capabilities, but varying efficiency, are https://hackage.haskell.org/package/ad and https://hackage.haskell.org/package/backprop. See also https://github.com/Mikolaj/horde-ad/blob/master/CREDITS.md. Benchmarks suggest that horde-ad has competitive performance.
It is hoped that the separation of AD logic from matrix and tensor manipulation (deferred to hmatrix and orthotope, respectively) will enable similar speedups on numerical accelerators.
Here is an example of a Haskell function to be differentiated:
-- A function that goes from R^3 to R.
foo :: RealFloat a => (a,a,a) -> a
foo (x,y,z) =
let w = x * sin y
in atan2 z w + z * w -- note that w appears twice
The gradient of foo
is:
grad_foo :: forall r. (HasDelta r, AdaptableScalar 'ADModeGradient r)
=> (r, r, r) -> (r, r, r)
grad_foo = rev @r foo
As can be verified by computing the gradient at (1.1, 2.2, 3.3)
:
>>> grad_foo (1.1 :: Double, 2.2, 3.3)
(2.4396285219055063, -1.953374825727421, 0.9654825811012627)
As a side note, w
is processed only once during gradient computation and this property of sharing preservation is guaranteed universally by horde-ad without any action required from the user. The property holds not only for scalar values, but for arbitrary tensors, e.g., those in further examples. We won't mention the property further.
An additional feature of this library is a type system for tensor shape arithmetic. Look at this beauty.
convMnistTwoS
kh@MkSN kw@MkSN h@MkSN w@MkSN c_in@MkSN c_out@MkSN _n_hidden@MkSN batch_size@MkSN
-- integer parameters denoting basic dimensions, with some notational noise
input -- input images, shape (batch_size, c_in, h, w)
(ker1, bias1) -- layer1 kernel, shape (c_out, c_in, kh+1, kw+1); and bias, shape (c_out)
(ker2, bias2) -- layer2 kernel, shape (c_out, c_out, kh+1, kw+1); and bias, shape (c_out)
( weigthsDense -- dense layer weights,
-- shape (n_hidden, c_out * ((h+kh)/2 + kh)/2, ((w+kw)/2 + kw)/2)
, biasesDense ) -- dense layer biases, shape (n_hidden)
( weigthsReadout -- readout layer weights, shape (10, n_hidden)
, biasesReadout ) -- readout layer biases (10)
= -- -> output classification, shape (10, batch_size)
let t1 = convMnistLayerS kh kw
h w
c_in c_out batch_size
ker1 (constant input) bias1
t2 = convMnistLayerS kh kw
(MkSN @((h + kh) `Div` 2)) (MkSN @((w + kw) `Div` 2))
c_out c_out batch_size
ker2 t1 bias2
m1 = mapOuterS reshapeS t2
m2 = transpose2S m1
denseLayer = weigthsDense <>$ m2 + asColumnS biasesDense
denseRelu = relu denseLayer
in weigthsReadout <>$ denseRelu + asColumnS biasesReadout
But we don't just want the shapes in comments and in runtime expressions; we want them in the type system to be checked at compile time:
convMnistTwoS
:: forall kh kw h w c_in c_out n_hidden batch_size d r.
( 1 <= kh -- kernel height is large enough
, 1 <= kw -- kernel width is large enough
, ADModeAndNum d r ) -- differentiation mode and numeric type are known to the engine
=> -- The two boilerplate lines below tie type parameters to the corresponding
-- value parameters (built with MkSN) denoting basic dimensions.
StaticNat kh -> StaticNat kw -> StaticNat h -> StaticNat w
-> StaticNat c_in -> StaticNat c_out -> StaticNat n_hidden -> StaticNat batch_size
-> OS.Array '[batch_size, c_in, h, w] r
-> ( ADVal d (OS.Array '[c_out, c_in, kh + 1, kw + 1] r)
, ADVal d (OS.Array '[c_out] r ) )
-> ( ADVal d (OS.Array '[c_out, c_out, kh + 1, kw + 1] r)
, ADVal d (OS.Array '[c_out] r) )
-> ( ADVal d (OS.Array '[ n_hidden
, c_out * (((h + kh) `Div` 2 + kh) `Div` 2)
* (((w + kw) `Div` 2 + kw) `Div` 2)
] r)
, ADVal d (OS.Array '[n_hidden] r) )
-> ( ADVal d (OS.Array '[10, n_hidden] r)
, ADVal d (OS.Array '[10] r) )
-> ADVal d (OS.Array '[10, batch_size] r)
The full neural network definition from which this function is taken can be found at
https://github.com/Mikolaj/horde-ad/tree/master/example
in file MnistCnnShaped.hs
and the directory contains several other sample neural networks for MNIST digit classification. Among them are recurrent, convolutional and fully connected networks based on fully typed tensors (sizes of all dimensions are tracked in the types, as above) as well as weakly typed fully connected networks built with, respectively, matrices, vectors and raw scalars (working with scalars is the most flexible but slowest; all others have comparable performance on CPU).
Because we use hmatrix the OS needs libraries that on Ubuntu/Debian are called libgsl0-dev, liblapack-dev and libatlas-base-dev. See https://github.com/haskell-numerics/hmatrix/blob/master/INSTALL.md for information about other OSes. Other Haskell packages need their usual C library dependencies, as well, e.g., package zlib needs C library zlib1g-dev.
For development, copying the included cabal.project.local.development
to cabal.project.local
provides a sensible default to run cabal build
with.
Then a command like
cabal test minimalTest --enable-optimization --test-options='-p "Simple QuickCheck of gradient vs derivative vs perturbation"'
ensures that the code is compiled with optimization and so executes the rather computation-intensive testsuite in reasonable time.
The test suite is run in parallel mode by default:
cabal test shortTestForCI --enable-optimization
However, this may cause extra printf messages from within the tests to be out of order. To keep your screen tidy, simply redirect stderr
, e.g.: 2>/dev/null
:
cabal test shortTestForCI --enable-optimization 2>/dev/null
You can also run the test suite sequentially:
cabal test shortTestForCI --enable-optimization -f test_seq
Stylish Haskell is used for slight auto-formatting at buffer save; see .stylish-haskell.yaml. As defined in the file, indentation is 2 spaces wide and screen is 80-columns wide. Spaces are used, not tabs. Spurious whitespace avoided. Spaces around arithmetic operators encouraged. Generally, relax and try to stick to the style apparent in a file you are editing. Put big formatting changes in separate commits.
Haddocks should be provided for all module headers and for all functions and types from most important modules. Apart of that, only particularly significant functions and types are distinguished by having a haddock. If minor ones have comments, they should not be haddocks and they are permitted to describe implementation details and be out of date. Prefer assertions instead of comments, unless too verbose.
Copyright 2022 Mikolaj Konarski and others (see git history)
License: BSD-3-Clause (see file LICENSE)