owl 0.3.0 · OCaml Package

README

%%VERSION%%

Owl is an emerging numerical library for scientific computing and engineering. The library is developed in the OCaml language and inherits all its powerful features such as static type checking, powerful module system, and superior runtime efficiency. Owl allows you to write succinct type-safe numerical applications in functional language without sacrificing performance, significantly reduces the cost from prototype to production use.

Behind the scene, Owl builds the advanced numerical functions atop of its solid implementation of n-dimensional arrays. Quite different from other libraries, algorithmic differentiation and distributed computing have been included as integral components in the core library to maximise developers' productivity. Owl is young but grows fast, the current features include:

N-dimensional array (both dense and sparse)
Various number types: float32, float64, complex32, complex64, int8, int16, int32 ...
Linear algebra and full interface to CBLAS and LAPACKE
Algorithmic differentiation (or automatic differentiation)
Neural network module for deep learning applications
Parallel and Distributed engine (map-reduce, parameter server, etc.)
Advanced math and stats functions (e.g., hypothesis tests, MCMC, etc.)
Zoo system for efficient scripting and code sharing
GPU support (work in progress ...)

You can reach us in the following ways, looking forward to hearing from you!

Build status:

Learning

Here are some learning material to help you start with Owl. The full API documentation is recently built on 26th September, 2017:

on cl.cam.ac.uk

This How-To? maintains a list of practical tips. The series of tutorials is here (more are coming):

The Presentation Slides in OCaml Workshop 2017 can be viewed with this link.
The tutorial material of Owl in CUFP 2017 can be accessed with this link.

Installation

Owl requires OCaml >=4.04.0. The installation is rather trivial. There are four possible ways to try out Owl, as shown below, from the most straightforward one to the least one.

Option 1: Install from OPAM

You can simply type the following in the command line to start.

opam install owl

Owl's current version on OPAM is 0.2.9, and it lags behind the master branch and misses many new features. If you want to try the newest version, I recommend the other three ways to install.

Option 2: Pull from Docker Hub

Owl's docker image is perfectly synced with master branch and is always automatically built whenever there are new commits. You can check the building history on Docker Hub.

You only need to pull the image then start a container, then play with it in utop. The source code is stored in /root/owl directory.

docker pull ryanrhymes/owl
docker run -t -i ryanrhymes/owl

Option 3: Pin the Dev-Repo

opam pin allows us to pin the local code to Owl's development repository on Github. The first command opam depext installs all the dependencies Owl needs.

opam depext owl
opam pin add owl --dev-repo

Option 4: Compile from Source

First, you need to clone the repository.

git clone git@github.com:ryanrhymes/owl.git

Second, you need to figure out the missing dependencies and install them.

jbuilder external-lib-deps --missing @install

Last, this is perhaps the most classic step.

make && make install

CBLAS/LAPACKE Dependency

The most important dependency is OpenBLAS. Linking to the correct OpenBLAS is the key to achieve the best performance. Depending on the specific platform, you can use yum, apt-get, brew to install the binary format. For example on Mac OSX,

brew install homebrew/science/openblas

However, installing from OpenBLAS source code leads to way better performance in my own experiment. In future, the dependency on OpenBLAS should also be resolved by opam automatically.

Integration to Toplevel

Owl is well integrated with utop. You can use utop to try out the experiments in our tutorials. If you want utop to automatically load Owl for you, you can also edit .ocamlinit file in your home folder by adding the following lines. (Note that the library name is owl with lowercase o.)

#require "owl_top"

The owl_top is the toplevel library of Owl, it automatically loads owl core library and installs the corresponding pretty printers of various data types.

Access Modules

Owl currently has the following core modules and their names all start with Owl_ to avoid name conflicts, e.g., Owl_dense, Owl_sparse, Owl_maths, Owl_stats, Owl_const, Owl_fft, Owl_plot and etc. After utop successfully loads Owl library, you can access the module functions using aforementioned names.

However, a more convenient way is to use Owl module as an entry point which provides aliases of the core module names for easy access, e.g., Owl.Dense is the same as Owl_dense, and Owl.Regression is the same as Owl_regression. Given no name conflicts, you can simply open the whole Owl module for convenience as I will do in the rest of this tutorial.

open Owl;;

Create Matrices

Dense.Matrix module supports dense matrix operations while Sparse.Matrix module supports sparse ones. There are five submodules in Dense.Matrix:

Dense.Matrix.S module supports single precision float numbers float32;
Dense.Matrix.D module supports double precision float numbers float64;
Dense.Matrix.C module supports single precision complex numbers complex32;
Dense.Matrix.Z module supports double precision complex numbers complex64;
Dense.Matrix.Generic module supports all aforementioned number types via GADT.

To start, we can use Dense.Matrix.D.uniform_int to create a 5x5 random dense matrix.

let x = Dense.Matrix.D.uniform_int 5 5;;

You should see the following output in utop.

   C0 C1 C2 C3 C4
R0 25  2 77 85 72
R1 71 92 98 87 53
R2 35 29 82 65 20
R3  2 29 66 42 12
R4 99 72 78 30 11
val x : Owl_dense_matrix_d.mat =

To save some typing efforts, we have made Mat as an alias of Dense.Matrix.D by assuming 64-bit float numbers are commonly used. Therefore, we can use Mat directly after open Owl instead of using Dense.Matrix.D. Similarly, there are also aliases for 64-bit float vectors and ndarrays (i.e., Vec and Arr) but we will talk about them later. Mat module also provides other functions to create various matrices, e.g., as below.

let x = Mat.eye 5;;             (* identity matrix *)
let x = Mat.zeros 5 5;;         (* all elements are zeros *)
let x = Mat.ones 5 5;;          (* all elements are ones *)
let x = Mat.uniform 5 5;;       (* random matrix of uniform distribution *)
let x = Mat.gaussian 5 5;;      (* random matrix of gaussian distribution *)
let x = Mat.triu x;;            (* Upper triangular matrix *)
let x = Mat.toeplitz v;;        (* Toeplitz matrix *)
let x = Mat.hankel v;;          (* Hankel matrix *)
let x = Mat.hadamard 8;;        (* Hadamard matrix *)
let x = Mat.magic 8;;           (* Magic square matrix *)
...

Combined with Stats module, you can also create any matrices of many distributions. E.g., the following code first creates an empty dense matrix, then initialise the elements with Bernoulli distribution. Test it in utop, you should get a dense matrix where half of the elements are zeros.

let x = Mat.empty 8 8 |> Mat.map (fun _ -> Stats.Rnd.bernoulli 0.5 |> float_of_int);;

Or create a matrix where the elements follow Laplace distribution.

let x = Mat.empty 8 8 |> Mat.map (fun _ -> Stats.Rnd.laplace 0.2);;

With Dense module, you can also generate linearly spaced interval and meshgrids, e.g.,

let x = Mat.linspace 0. 5. 6;;

which will return a 1x5 row vector as below

   C0  C1  C2  C3  C4  C5
R0  0   1   2   3   4   5
val x : Owl_dense_matrix_d.mat =

The created matrices can be casted into other number types easily. For example the following code casts a float32 matrix x into complex64 matrix y.

let x = Dense.Matrix.S.uniform 3 3;;
let y = Dense.Matrix.Generic.cast_s2z x;;

Matrices can be saved to and loaded from a file.

Mat.save x "matrix_01.data";;  (* save the matrix to a file *)
Mat.load "matrix_01.data";;    (* load the matrix from a file *)

Access Elements, Rows, and Columns

Both Dense.Matrix and Sparse.Matrix modules provide a wide range of operations to access the elements, rows, and columns of a matrix. You can refer to the full document in Dense.Matrix.Generic. Here we just gave some simple examples briefly.

You can use Mat.set and Mat.get to manipulate individual element.

Mat.set x 0 1 2.5;;
Mat.get x 0 1;;

Equivalently, there are shorthands for Mat.get and Mat.set.

x.{0,1} <- 2.5;;  (* Mat.set x 0 1 2.5 *)
x.{0,1};;         (* Mat.get x 0 1 *)

We can use Mat.row and Mat.col to retrieve a specific row or column of a matrix, or use Mat.rows and Mat.cols to retrieve multiple of them.

Mat.row x 5;;            (* retrieve the fifth row *)
Mat.cols x [|1;3;2|];;   (* retrieve the column 1, 3, and 2 *)

E.g., the following code generates a random matrix, then scales up each element by a factor of 10 using Mat.map function.

let x = Mat.(uniform 6 6 |> map (fun x -> x *. 10.));;

We can iterate a matrix row by row, or column by column. The following code calculates the sum of each row by calling Mat.map_rows function.

let x = Mat.(uniform 6 6 |> map_rows sum);;

We can fold elements by calling Mat.fold, fold rows by calling Mat.fold_rows. Similarly, there are also functions for filter operations. The following code filters out the elements not greater than 0.1 in x.

Mat.filter ((>) 0.1) x;;    (* not greater than 0.1 in x *)

We can also do something more complicated, e.g., by filtering out the rows whose summation is greater than 3.

Mat.filter_rows (fun r -> Mat.sum r > 3.) x;;

Shuffle the rows and columns, or draw some of them from a matrix.

Mat.shuffle_rows x;;     (* shuffle the rows in x *)
Mat.draw_cols x 3;;      (* draw 3 columns from x with replacement *)
...

Practically, Sparse.Matrix module provides a subset of the similar operations for sparse matrices. In addition, Sparse.Matrix module also has extra functions such as only iterating non-zero elements Sparse.Matrix.Generic.iter_nz, and etc. Please read the full documentation for Sparse.Matrix.Generic for details.

Linear Algebra

Simple matrix mathematics like add, sub, multiplication, and division are included in Dense module. Moreover, there are predefined shorthands for such operations. E.g., the following code creates two random matrices then compare which is greater.

let x = Mat.uniform 6 6;;
let y = Mat.uniform 6 6;;
Mat.(x > y);;                (* is x greater than y? return boolean *)
Mat.(x = y);;                (* is x equal to y? *)
Mat.(x =~ y);;               (* is x approximately equal to y? *)
Mat.(x <. y);;               (* is x smaller than y? return 0/1 matrix *)
...

Owl natively supports broadcast operation similar to other numerical libraries. Some basic math operations includes:

Mat.(x + y);;                (* addition of two matrices *)
Mat.(x * y);;                (* element-wise multiplication *)
Mat.(x *@ y);;               (* matrix multiplication of two matrices *)
Mat.(x +$ 2.);;              (* add a scalar to all elements in x *)
...

Apply various functions in Maths module to every element in x

Mat.(Maths.atanh @@ x);;        (* apply atanh function *)
Mat.(Maths.airy_Ai @@ x);;      (* apply Airy function *)
...

However, it is worth pointing out that Mat already implements many useful math functions. These functions are vectorised and are much faster than the example above which actually calls Mat.map for transformation.

Mat.sin x;;         (* call sine function *)
Mat.erfc x;;        (* call erfc function *)
Mat.round x;;       (* call round function *)
Mat.signum x;;      (* call signum function *)
Mat.sigmoid x;;     (* apply sigmoid function *)
...

Concatenate two matrices, vertically or horizontally by

Mat.(x @= y);;              (* equivalent to Mat.concat_vertical *)
Mat.(x @|| y);;             (* equivalent to Mat.concat_horizontal *)
Mat.concatenate [|x;...|];; (* concatenate a list of matrices along rows *)

More advanced linear algebra operations such as svd, qr, and cholesky decomposition are included in Linalg module. Linalg module also supports both real and complex number of single and double precision.

let u,s,v = Linalg.D.svd x;;     (* singular value decomposition *)
let q,r = Linalg.D.qr x;;        (* QR decomposition *)
let l,u,_ = Linalg.D.lu x;;      (* LU decomposition *)
let l = Linalg.D.chol x;;        (* cholesky decomposition *)
...
let e, v = Linalg.D.eig x;;      (* Eigenvectors and eigenvalues *)
let x = Linalg.D.null a;;        (* Solve A*x = 0, null space *)
let x = Linalg.D.linsolve a b;;  (* Solve A*x = B *)
let a, b = Linalg.D.linreg x y;; (* Simple linear regression y = a*x + b *)
...

Linalg module offers additional functions to check the properties of a matrix. For example,

Linalg.D.cond x;;          (* condition number of x *)
Linalg.D.rank x;;          (* rank of x *)
Linalg.D.is_diag x;;       (* is it diagonal *)
Linalg.D.is_triu x;;       (* is it upper triangular *)
Linalg.D.is_tril x;;       (* is it lower triangular *)
Linalg.D.is_posdef x;;     (* is it positive definite *)
Linalg.D.is_symmetric x;;  (* is it symmetric *)
...

Owl has implemented a complete set of OCaml interface to CBLAS and LAPACKE libraries. You can utilise these highly optimised functions to achieve the best performance. However in most cases, you should only use the high-level functions in Linalg module rather than dealing with these low-level interface.

Regression

Regression module currently includes linear, exponential, nonlinear, ols, ridge, lasso, svm, and etc. Most of them are based on a stochastic gradient descent algorithm implemented in Optimise module.

In the following, let's use an example to illustrate the simplest linear regression in Regression module. First, let's generate the measurement x which is a 1000 x 3 matrix. Each row of x is an independent measurement.

let x = Mat.uniform 1000 3;;

Next let's define the parameter of a linear model, namely p, a 3 x 1 matrix.

let p = Mat.of_array [|0.2;0.4;0.8|] 3 1;;

Then we generate the observations y from x and p by

let y = Mat.(x *@ p);;

Now, assume we only know x and y, how can we fit x and y into a linear model? It is very simple.

let p' = Regression.linear x y;;

From utop, you can see that p' equals [|0.2; 0.4; 0.8|] which is exactly the same as p. For other regression such as lasso and svm, the operation is more or less the same, please read Owl document for details.

Plotting

There is another separate Tutorial on Plotting in Owl.
Herein, let's use an example to briefly show how to plot the result using Plot module. We first generate two mesh grids then apply sine function to them by using the operations introduced before.

let x, y = Mat.meshgrid (-2.5) 2.5 (-2.5) 2.5 100 100 in
let z = Mat.(sin ((x * x) + (y * y))) in
Plot.mesh x y z;;

No matter what plot terminal you use, you should end up with a figure as below.

Besides Plot.mesh, there are several other basic plotting functions in Plot. Even though the module is still immature and under active development, it can already do some fairly complicated plots with minimal coding efforts. E.g., the following code will generate a 2 x 2 subplot.

let f p i = match i with
  | 0 -> Stats.Rnd.gaussian ~sigma:0.5 () +. p.(1)
  | _ -> Stats.Rnd.gaussian ~sigma:0.1 () *. p.(0)
in
let y = Stats.gibbs_sampling f [|0.1;0.1|] 5_000 |> Mat.of_arrays in
let h = Plot.create ~m:2 ~n:2 "" in
Plot.set_background_color h 255 255 255;

Plot.subplot h 0 0;
Plot.set_title h "Bivariate model";
Plot.scatter ~h (Mat.col y 0) (Mat.col y 1);

Plot.subplot h 0 1;
Plot.set_title h "Distribution of y";
Plot.set_xlabel h "y";
Plot.set_ylabel h "Frequency";
Plot.histogram ~h ~bin:50 (Mat.col y 1);

Plot.subplot h 1 0;
Plot.set_title h "Distribution of x";
Plot.set_ylabel h "Frequency";
Plot.histogram ~h ~bin:50 (Mat.col y 0);

Plot.subplot h 1 1;
Plot.set_foreground_color h 0 50 255;
Plot.set_title h "Sine function";
Plot.(plot_fun ~h ~spec:[ LineStyle 2 ] Maths.sin 0. 28.);
Plot.autocorr ~h (Mat.sequential 1 28);

Plot.output h;;

The end result is as follows. You probably have already grasped the idea of how to plot in Owl. But I promise to write another separate post to introduce plotting in more details.

Maths and Stats

There are a lot of basic and advanced mathematical and statistical functions in Maths and Stats modules. Most of them are interfaced to Gsl directly, so you may want to read GSL Manual carefully before using the module. In the future, Owl will also supports other math library as optional backend in case you need different licence.

Stats has three submodules: Stats.Rnd for random numbers, Stats.Pdf for probability dense functions, and Stats.Cdf for cumulative distribution functions. In addition, I have implemented extra functions such as two ranking correlations: Stats.kendall_tau and Stats.spearman_rho); two MCMC (Markov Chain Monte Carlo) functions in Stats module: Metropolis-Hastings (Stats.metropolis_hastings) and Gibbs sampling (Stats.gibbs_sampling) algorithms.

E.g., the following code first defines a probability density function f for a mixture Gaussian model. Then we use Stats.metropolis_hastings to draw 100_000 samples based on the given pdf f, and the initial point is 0.1. In the end, we call Plot.histogram to plot the distribution of the samples, from which we can clearly see they are from a mixture Gaussian model.

let f p = Stats.Pdf.((gaussian p.(0) 0.5) +. (gaussian (p.(0) -. 3.5) 1.)) in
let y = Stats.metropolis_hastings f [|0.1|] 100_000 |>  Mat.of_arrays in
Plot.histogram ~bin:100 y;;

The histogram below shows the distribution of the samples.

Here is another example using Stats.gibbs_sampling to sample a bivariate distribution. Gibbs sampling requires the full conditional probability function so we defined its corresponding random number generator in f p i where p is the parameter vector and i indicates which parameter to sample.

let f p i = match i with
  | 0 -> Stats.Rnd.gaussian ~sigma:0.5 () +. p.(1)
  | _ -> Stats.Rnd.gaussian ~sigma:0.1 () *. p.(0)
in
let y = Stats.gibbs_sampling f [|0.1;0.1|] 5_000 |> Mat.of_arrays in
Plot.scatter (Mat.col y 0) (Mat.col y 1);;

We take 5000 samples from the defined distribution and plot them as a scatter plot, as below.

The future plan is to embed a small PPL (Probabilistic Programming Language) in Stats module.

N-dimensional Array

Owl has a very powerful module to manipulate dense N-dimensional arrays, i.e., Dense.Ndarray. Ndarray is very similar to the corresponding modules in Numpy and Julia. For sparse N-dimensional arrays, you can use Sparse.Ndarray which provides a similar set of APIs as aforementioned Ndarray. Here is an initial evaluation on the performance of Ndarray.

Similar to Matrix module, Ndarray also has five submodules S (for float32), D (for float32), C (for complex32), Z (for complex64), and Generic (for all types) to handle different number types. There is an alias in Owl for double precision float ndarray (i.e., Dense.Ndarray.D) which is Arr. Ndarray also natively supports broadcast operations

In the following, I will present a couple of examples using Dense.Ndarray module. First, we can create empty ndarrays of shape [|3;4;5|].

let x0 = Dense.Ndarray.S.empty [|3;4;5|];;
let x1 = Dense.Ndarray.D.empty [|3;4;5|];;
let x2 = Dense.Ndarray.C.empty [|3;4;5|];;
let x3 = Dense.Ndarray.Z.empty [|3;4;5|];;

You can also assign the initial values to the elements, generate a zero/one ndarray, or even a random ndarray.

Dense.Ndarray.C.zeros [|3;4;5|];;
Dense.Ndarray.D.ones [|3;4;5|];;
Dense.Ndarray.S.create [|3;4;5|] 1.5;;
Dense.Ndarray.Z.create [|3;4;5|] Complex.({im=1.5; re=2.5});;
Dense.Ndarray.D.uniform [|3;4;5|];;

With these created ndarray, you can do some math operation as below. Now, let's use shortcut Arr module to make examples.

let x = Arr.uniform [|3;4;5|];;
let y = Arr.uniform [|3;4;5|];;
let z = Arr.add x y;;
Arr.print z;;

Owl supports many math operations and these operations have been well vectorised so they are very fast.

Arr.sin x;;
Arr.tan x;;
Arr.exp x;;
Arr.log x;;
Arr.min x;;
Arr.add_scalar x 2.;;
Arr.mul_scalar x 2.;;
...

Examining elements and comparing two ndarrays are also very easy.

Arr.is_zero x;;
Arr.is_positive x;;
Arr.is_nonnegative x;;
...
Arr.equal x y;;
Arr.greater x y;;
Arr.less_equal x y;;
...

You can certainly plugin your own functions to check each elements.

Arr.exists ((>) 2.) x;;
Arr.not_exists ((<) 2.) x;;
Arr.for_all ((=) 2.) x;;

Most importantly, you can use Owl to iterate a ndarray in various ways. Owl provides a simple but flexible and powerful way to define a "slice" in ndarray. Comparing to the "Bigarray.slice_left" function, the slice in Owl does not have to start from the left-most axis. E.g., for the previously defined [|3;4;5|] ndarray, you can define a slice in the following ways:

let s0 = [ []; []; [] ]      (* (*,*,*), essentially the whole ndarray as one slice *)
let s1 = [ [0]; []; [] ]     (* (0,*,*) *)
let s2 = [ []; [2]; [] ]     (* (*,2,*) *)
let s3 = [ []; []; [1] ]     (* (*,*,1) *)
let s4 = [ [1]; []; [2] ]    (* (1,*,2) *)
...

slice function is very flexible, it basically has the same semantic as that in numpy. So you know how to index ndarray in numpy, you should be able to do the same thing in Owl. For advanced use of slice function, please refer to my separate tutorial. Some examples as as below.

let s = [ [1]; []; [-1;0;-1]; ];;
let s = [ [1]; [0]; [-1;0;-1]; ];;
let s = [ [1]; [0]; [-2;0]; ];;
let s = [ [0]; [0;1]; [-2;0;-2]; ];;
...

With the slice definition above, we can iterate and map the elements in a slice. E.g., we add one to all the elements in slice (0,*,*).

Arr.map ~axis:[ [0]; []; [] ] (fun a -> a +. 1.) x;;

There are more functions to help you to iterate elements and slices in a ndarray: iteri, iter, mapi, map, filteri, filter, foldi, fold, iteri_slice, iter_slice, iter2i, iter2. Please refer to the documentation for their details.

Algorithmic Differentiation

Algorithmic differentiation (AD) is another key component in Owl which can make many analytical tasks so easy to perform. It is also often referred to as Automatic differentiation. Here is a Wikipedia article to help you understand the topic if you are interested in.

The AD support is provided Algodiff module. More precisely, Algodiff.Numerical provides numerical differentiation whilst Algodiff.S and Algodiff.D provides algorithmic differentiation for single and double precision float numbers respectively. For the detailed differences between the two, please read the wiki article as your starting point. Simply put, Algodiff.S/D is able to provide exact result of the derivative whereas Algodiff.Numerical is just approximation which is subject to round and truncate errors.

Algodiff.S supports higher-order derivatives. Here is an example which calculates till the fourth derivative of tanh function.

open Algodiff.S;;

(* calculate derivatives of f0 *)
let f0 x = Maths.(tanh x);;
let f1 = f0 |> diff;;
let f2 = f0 |> diff |> diff;;
let f3 = f0 |> diff |> diff |> diff;;
let f4 = f0 |> diff |> diff |> diff |> diff;;

Quite easy, isn't it? Then we can plot the values of tanh and its four derivatives between interval [-4, 4].

let map f x = Vec.map (fun a -> a |> pack_flt |> f |> unpack_flt) x;;

(* calculate point-wise values *)
let x = Vec.linspace (-4.) 4. 200;;
let y0 = map f0 x;;
let y1 = map f1 x;;
let y2 = map f2 x;;
let y3 = map f3 x;;
let y4 = map f4 x;;

(* plot the values of all functions *)
let h = Plot.create "plot_021.png" in
Plot.set_foreground_color h 0 0 0;
Plot.set_background_color h 255 255 255;
Plot.plot ~h x y0;
Plot.plot ~h x y1;
Plot.plot ~h x y2;
Plot.plot ~h x y3;
Plot.plot ~h x y4;
Plot.output h;;

Then you should be able to see a figure like this one below. For more advanced use, please see my separate tutorial.

Machine Learning and Neural Network

Even though this is still work in progress, I find it necessary to present a small neural network example to show how necessary it is to have a comprehensive numerical infrastructure. The illustration in the following is of course the classic MNIST example wherein we will train a two-layer network that can recognise hand-written digits.

Neural module has been included in Owl main library. To define a feedforward network, you can first open Neural.S.Graph in utop. Neural.S.Graph contains many easy-to-use wrappers to create different types of neuron node, S indicates the network is for single precision.

open Neural.S;;
open Neural.S.Graph;;

Now, let's see how to define a two-layer neural network.

let nn = input [|784|]
  |> linear 300 ~act_typ:Activation.Tanh
  |> linear 10  ~act_typ:Activation.Softmax
  |> get_network
;;

Done! Only three lines of code, that's easy, isn't it? Owl's Neural module is built atop of its Algodiff module. I am often amazed by the power of algorithmic differentiation while developing the neural network module, it just simplifies the design so much and makes life so easy.

Let's look closer at what the code does: the first line defines the input shape of the neural network; the second line adds a linear layer (of shape 784 x 300) with Tanh activation; the third line does the similar thing by adding another linear layer with Softmax activation. The input shape of each layer is automatically inferred for you. The last line get_network returns the created network for training.

You can print out the summary of the neural network by calling Graph.print nn, then you see the following output.

Graphical network

[ Node input_0 ]:
    Input : in/out:[*,784]
    prev:[] next:[linear_1]

[ Node linear_1 ]:
    Linear : matrix in:(*,784) out:(*,300)
    init   : standard
    params : 235500
    w      : 784 x 300
    b      : 1 x 300
    prev:[input_0] next:[activation_2]

[ Node activation_2 ]:
    Activation : tanh in/out:[*,300]
    prev:[linear_1] next:[linear_3]

[ Node linear_3 ]:
    Linear : matrix in:(*,300) out:(*,10)
    init   : standard
    params : 3010
    w      : 300 x 10
    b      : 1 x 10
    prev:[activation_2] next:[activation_4]

[ Node activation_4 ]:
    Activation : softmax in/out:[*,10]
    prev:[linear_3] next:[]

How to train the defined network now? You only need two lines of code to load the dataset and start training. By the way, calling Dataset.download_all () will download all the data sets used in Owl (about 1GB uncompressed data).

let x, _, y = Dataset.load_mnist_train_data () in
train nn x y;;

You may ask "what if I want different training configuration?" Well, the training and network module is actually very flexible and highly configurable. But I will talk about these details in another separate tutorial.

Distributed & Parallel Computing

Owl's distributed and parallel computing relies on my another research prototype - Actor System. Actor is a specialised distributed data processing framework. Please do not get confused with Actor Model since Owl's Actor system actually implements three engines: MapReduce, Parameter Server, and Peer-to-Peer.

My design principle of distributed analytics is: Owl handles "analytics" whilst Actor deals with "distribution" with a suitable engine. Two systems can be composed through functors just like we play LEGO. This composition includes both low-level data structures and high-level models.

For example, the following one-line code composes Owl's Ndarray with Actor's MapReduce engine to provide us a distributed Ndarray module M1.

module M1 = Owl_parallel.Make_Distributed (Dense.Ndarray.D) (Actor_mapre)

Similarly, for high-level neural network models, we only need to add one extra line of code to transform a single-node training model to a distributed training model. Note that we have composed Owl's Feedforward neural network with Actor's Parameter Server engine.

module M2 = Owl_neural_parallel.Make (Owl_neural_feedforward) (Actor_param)

Actor system is currently in a closed repository (due to my techreport writing). I will introduce this exciting feature very soon (in September).

Run Owl on Different Platforms

If you want to try Owl on ARM based platforms such as Raspberry Pi rather than x86 ones, the installation are similar. Just note that Owl requires OCaml 4.04, which might not be supported on your platform's binary distribution system yet, so you might consider compiling OCaml sources. Besides, to solve a potential conflict with gsl package, after running ./configure in the top directory, you should run:

sed -i -e 's/#define ARCH_ALIGN_DOUBLE/#undef ARCH_ALIGN_DOUBLE/g' config/m.h config/m-templ.h

before running make world.opt.

A Docker image is also provided on Docker Hub specifically for ARM platform. Just pull the image, start a container, then play with it in utop.

docker run --name owl -it matrixanger/owl:arm

Note that after starting a new container you need to run

eval `opam config env`

for once before starting utop.

How To Contribute

Owl is under active development, and I really look forward to your comments and contributions. Besides setting up a complete development environment on your native system, the easiest way to contribute is to use the Owl Docker Image. Moreover, we have also built a docker image for ARM-based platform so that you can run Owl on Raspberry PI and Cubietruck (see the section above).

Just pull the image and dig into code saved in /root/owl, then have fun!

Student Project: If you happen to be a student in the Computer Lab and want to do some challenging development and design, here are some Part II Projects. If you are interested in more researchy topics, I also offer Part III Projects and please contact me directly via Email.

Acknowledgement: Funded in part by EPSRC project - Contrive (EP/N028422/1). Please refer to the full acknowledgement for more details.