mom_ann module reference

Implements the general purpose Artificial Neural Network (ANN).

Data Types

`layer_type`	Type for a single Linear layer of ANN, i.e.
`ann_cs`	Control structure/type for ANN.

Functions/Subroutines

`ann_init()`	Initialization of ANN.
`ann_allocate()`	Allocate an ANN.
`ann_test()`	Test ANN by comparing the prediction with the test data.
`ann_end()`	Deallocates memory of ANN.
`activation_fn()`	The default activation function.
`ann_apply_vector_orig()`	Single application of ANN inference using vector input and output.
`ann_apply_vector_oi()`	Single application of ANN inference using vector input and output.
`ann_apply_array_sio()`	Single application of ANN inference using array input and output with (space,feature) indexing.
`set_layer()`	Sets weights and bias for a single layer.
`set_input_normalization()`	Sets input normalization.
`set_output_normalization()`	Sets output normalization.
`ann_random()`	Create a random ANN.
`randomize_layer()`	Fill a layer with random numbers.
`ann_unit_tests()`	Runs unit tests on ANN functions.

Detailed Description

The mom_ann() module is a pure fortran implementation of fully-connected feed-forward networks to facilitate easy evaluation of data-driven functions in MOM6. For performant implementations or for novel architectires, using machine-learning libraries (e.g. via module is a pure fortran implementation of fully-connected feed-forward networks to facilitate easy evaluation of data-driven functions in MOM6. For performant implementations or for novel architectires, using machine-learning libraries (e.g. via mom_database_comms()) are necessary, or at least likely to be more efficient.) are necessary, or at least likely to be more efficient.

The artificial neural network (ANN) understood by this MOM6 module has \(N\) layers, including the input-layer and output-layer, thus requireing \(N \geq 2\).

The output values (neurons or nodes) of any layer other than the input layer (i.e. \(l>1\)) are

\[y_{l,j} = f_l( b_{l,j} + A_{l,j,i} x_{l-1,i} )\]

where \(f(x) = max(0, x)\) is the ReLU activation function, \(b_{l,j}\) is a bias for each neuron, \(A_{l,j,i}\) are a rectangular matrix of weights for each layer, and \(x_{l-1,i}\) are the outputs of the previous layer, \(l-1\). The subscript on \(f_l()\) indicates the activation function is optional for each layer.

Currently, the performance of various implementations is dependent on the shape/size of the network and the size of input data. For this reason we provide several versions that all yield the same result but for differently shaped inputs.

/images/https://upload.wikimedia.org/wikipedia/commons/4/46/Colored_neural_network.svgFig:Athreelayernetworkwith3inputs,2outputs,and1hiddenlayer.Therearetworectanglarmatricesofweights(blackarrows).Thebiasforeachneuronisimplied."

Type Documentation

type mom_ann/layer_type

Type for a single Linear layer of ANN, i.e. stores the matrix A and bias b for matrix-vector multiplication y = A*x + b.

Type fields:

% output_width :: integer, private Number of rows in matrix A.
% input_width :: integer, private Number of columns in matrix A.
% activation :: logical, private If true, apply the default activation function.
% a :: real, dimension(:,:), allocatable, private Matrix in column-major order of size A(output_width, input_width) [nondim].
% b :: real, dimension(:), allocatable, private bias vector of size output_width [nondim]

mom_ann module reference

Data Types

Functions/Subroutines

Detailed Description

Type Documentation

Function/Subroutine Documentation