Tutorial: Neural Network

I will cover the neural network module in this tutorial. My original purpose of introducing neural network module into Owl is two-fold:

• Test the expressiveness of Owl. Neural network is a useful and complex tool for building modern analytical applications so I chose it.

• To validate my research argument on how to structure modern (distributed) analytical libraries. Namely, the high-level analytical functionality (ML, DNN, optimisation, regression, and etc.) should be "glued" to the classic numerical functions via algorithmic differentiation, and the computation should be distributed via a specialised engine providing several well-defined distribution abstractions.

In the end, I only used less than 3.5k lines of code to implement a quite full-featured neural network module. Now let's go through what Neural module offers.

Module Structure

The Owl.Neural provides two submodules S and D for both single precision and double precision neural networks. In each submodule, it contains the following modules to allow you to work with the structure of the network and fine-tune the training.

• Graph : create and manipulate the neural network structure.
• Parallel : provide parallel computation capability, need to compose with Actor engine.
• Init : control the initialisation of the weights in the network.
• Activation : provide a set of frequently used activation functions.
• Params : maintains a set of training parameters.
• Batch : the batch parameter of training.
• Learning_Rate : the learning rate parameter of training.
• Loss : the loss function parameter of training.
• Gradient : the gradient method parameter of training.
• Momentum : the momentum parameter of training.
• Regularisation : the regularisation parameter of training.
• Clipping : the gradient clipping parameter of training.
• Checkpoint : the checkpoint parameter of training.

Types of Neuron

I have implemented a set of commonly used neurons in Owl.Neural.Neuron. Each neuron is a standalong module and adding a new type of neuron is much easier than adding a new one in Tensorflow or other framework thanks to Owl's Algodiff module.

Algodiff is the most powerful part of Owl and offers great benefits to the modules built atop of it. In neural network case, we only need to describe the logic of the forward pass without worrying about the backward propagation at all, because the Algodiff figures it out automatically for us thus reduces the potential errors. This explains why a full-featured neural network module only requires less than 3.5k lines of code. Actually, if you are really interested, you can have a look at Owl's Feedforward which only uses a couple of hundreds lines of code to implement a complete Feedforward network.

In practice, you do not need to use the modules defined in Owl.Neural.Neuron directly. Instead, you should call the functions in Graph module to create a new neuron and add it to the network. Currently, Graph contains the following neurons.

• input
• activation
• linear
• linear_nobias
• embedding
• recurrent
• lstm
• gru
• conv1d
• conv2d
• conv3d
• max_pool1d
• max_pool2d
• avg_pool1d
• avg_pool2d
• global_max_pool1d
• global_max_pool2d
• global_avg_pool1d
• global_avg_pool2d
• fully_connected
• dropout
• gaussian_noise
• gaussian_dropout
• alpha_dropout
• normalisation
• reshape
• flatten
• lambda
• add
• mul
• dot
• max
• average
• concatenate

These neurons should be sufficient for creating from simple MLP to the most complicated Google's Inception network.

Training & Inference

Owl provides a very functional way to construct a neural network. You only need to provide the shape of the date in the first node (often input neuron), then Owl will automatically infer the shape for you in the downstream nodes which saves us a lot of efforts and significantly reduces the potential bugs.

Let's use the single precision neural network as an example. To work with single precision networks, you need to use/open the following modules

open Owl
open Neural.S
open Neural.S.Graph
open Algodiff.S

The code below creates a small convolutional neural network of six layers. Usually, the network definition always starts with input neuron and ends with get_network function which finalises and returns the constructed network. We can also see the input shape is reserved as a passed in parameter so the shape of the data and the parameters will be inferred later whenever the input_shape is determined.

let make_network input_shape =
  input input_shape
  |> lambda (fun x -> Maths.(x / F 256.))
  |> conv2d [|5;5;1;32|] [|1;1|] ~act_typ:Activation.Relu
  |> max_pool2d [|2;2|] [|2;2|]
  |> dropout 0.1
  |> fully_connected 1024 ~act_typ:Activation.Relu
  |> linear 10 ~act_typ:Activation.Softmax
  |> get_network

Next, I will show you how the train function looks like. The first three lines in the train function is for loading the MNIST dataset and print out the network structure on the terminal. The rest lines defines a params which contains the training parameters such as batch size, learning rate, number of epochs to run. In the end, we call Graph.train_cnn to kick off the training process.

let train () =
  let x, _, y = Dataset.load_mnist_train_data_arr () in
  let network = make_network [|28;28;1|] in
  Graph.print network;

  let params = Params.config
    ~batch:(Batch.Mini 100) ~learning_rate:(Learning_Rate.Adagrad 0.005) 2.
  in
  Graph.train_cnn ~params network x y |> ignore

After the training is finished, you can call Graph.model_cnn to generate a functional model to perform inference. Moreover, Graph module also provides functions such as save, load, print, to_string and so on to help you in manipulating the neural network.

let model = Graph.model_cnn network;;
let predication = model data;;
...

You can have a look at Owl's MNIST CNN example for more details and run the code by yourself.

Examples

In the following, I will present several neural networks defined in Owl. All have been included in Owl's examples and can be run separately. If you are interested in the computation graph Owl generated for these networks, you can also have a look at this tutorial on Algodiff.

Multilayer Perceptron (MLP) for MNIST

let make_network input_shape =
  input input_shape
  |> linear 300 ~act_typ:Activation.Tanh
  |> linear 10 ~act_typ:Activation.Softmax
  |> get_network

Convolutional Neural Network for MNIST

let make_network input_shape =
  input input_shape
  |> lambda (fun x -> Maths.(x / F 256.))
  |> conv2d [|5;5;1;32|] [|1;1|] ~act_typ:Activation.Relu
  |> max_pool2d [|2;2|] [|2;2|]
  |> dropout 0.1
  |> fully_connected 1024 ~act_typ:Activation.Relu
  |> linear 10 ~act_typ:Activation.Softmax
  |> get_network

VGG-like Neural Network for CIFAR10

let make_network input_shape =
  input input_shape
  |> normalisation ~decay:0.9
  |> conv2d [|3;3;3;32|] [|1;1|] ~act_typ:Activation.Relu
  |> conv2d [|3;3;32;32|] [|1;1|] ~act_typ:Activation.Relu ~padding:VALID
  |> max_pool2d [|2;2|] [|2;2|] ~padding:VALID
  |> dropout 0.1
  |> conv2d [|3;3;32;64|] [|1;1|] ~act_typ:Activation.Relu
  |> conv2d [|3;3;64;64|] [|1;1|] ~act_typ:Activation.Relu ~padding:VALID
  |> max_pool2d [|2;2|] [|2;2|] ~padding:VALID
  |> dropout 0.1
  |> fully_connected 512 ~act_typ:Activation.Relu
  |> linear 10 ~act_typ:Activation.Softmax
  |> get_network

LSTM Network for Text Generation

let make_network wndsz vocabsz =
  input [|wndsz|]
  |> embedding vocabsz 40
  |> lstm 128
  |> linear 512 ~act_typ:Activation.Relu
  |> linear vocabsz ~act_typ:Activation.Softmax
  |> get_network

Google's Inception for Image Classification

Coming very soon ...

Future Work

There is a great space for optimisation. There are also some new neurons need to be added, e.g., upsampling, transposed convolution, and etc. Anyway, things will get better and better.

Enjoy Owl! Happy hacking!