Documentation

Define Custom Deep Learning Layer with Learnable Parameters

If Deep Learning Toolbox™ does not provide the layer you require for your classification or regression problem, then you can define your own custom layer using this example as a guide. For a list of built-in layers, see List of Deep Learning Layers.

To define a custom deep learning layer, you can use the template provided in this example, which takes you through the following steps:

1. Name the layer – give the layer a name so that it can be used in MATLAB®.

2. Declare the layer properties – specify the properties of the layer and which parameters are learned during training.

3. Create a constructor function (optional) – specify how to construct the layer and initialize its properties. If you do not specify a constructor function, then at creation, the software initializes the Name, Description, and Type properties with [] and sets the number of layer inputs and outputs to 1.

4. Create forward functions – specify how data passes forward through the layer (forward propagation) at prediction time and at training time.

5. Create a backward function – specify the derivatives of the loss with respect to the input data and the learnable parameters (backward propagation).

This example shows how to create a PReLU layer, which is a layer with a learnable parameter and use it in a convolutional neural network. A PReLU layer performs a threshold operation, where for each channel, any input value less than zero is multiplied by a scalar learned at training time.[1] For values less than zero, a PReLU layer applies scaling coefficients ${\alpha }_{i}$ to each channel of the input. These coefficients form a learnable parameter, which the layer learns during training.

This figure from [1] compares the ReLU and PReLU layer functions.

Layer with Learnable Parameters Template

Copy the layer with learnable parameters template into a new file in MATLAB. This template outlines the structure of a layer with learnable parameters and includes the functions that define the layer behavior.

classdef myLayer < nnet.layer.Layer

properties
% (Optional) Layer properties.

% Layer properties go here.
end

properties (Learnable)
% (Optional) Layer learnable parameters.

% Layer learnable parameters go here.
end

methods
function layer = myLayer()
% (Optional) Create a myLayer.
% This function must have the same name as the class.

% Layer constructor function goes here.
end

function [Z1, …, Zm] = predict(layer, X1, …, Xn)
% Forward input data through the layer at prediction time and
% output the result.
%
% Inputs:
%         layer       - Layer to forward propagate through
%         X1, ..., Xn - Input data
% Outputs:
%         Z1, ..., Zm - Outputs of layer forward function

% Layer forward function for prediction goes here.
end

function [Z1, …, Zm, memory] = forward(layer, X1, …, Xn)
% (Optional) Forward input data through the layer at training
% time and output the result and a memory value.
%
% Inputs:
%         layer       - Layer to forward propagate through
%         X1, ..., Xn - Input data
% Outputs:
%         Z1, ..., Zm - Outputs of layer forward function
%         memory      - Memory value for backward propagation

% Layer forward function for training goes here.
end

function [dLdX1, …, dLdXn, dLdW1, …, dLdWk] = ...
backward(layer, X1, …, Xn, Z1, …, Zm, dLdZ1, …, dLdZm, memory)
% Backward propagate the derivative of the loss function through
% the layer.
%
% Inputs:
%         layer             - Layer to backward propagate through
%         X1, ..., Xn       - Input data
%         Z1, ..., Zm       - Outputs of layer forward function
%         dLdZ1, ..., dLdZm - Gradients propagated from the next layers
%         memory            - Memory value from forward function
% Outputs:
%         dLdX1, ..., dLdXn - Derivatives of the loss with respect to the
%                             inputs
%         dLdW1, ..., dLdWk - Derivatives of the loss with respect to each
%                             learnable parameter

% Layer backward function goes here.
end
end
end

Name the Layer

First, give the layer a name. In the first line of the class file, replace the existing name myLayer with preluLayer.

classdef preluLayer < nnet.layer.Layer
...
end

Next, rename the myLayer constructor function (the first function in the methods section) so that it has the same name as the layer.

methods
function layer = preluLayer()
...
end

...
end

Save the Layer

Save the layer class file in a new file named preluLayer.m. The file name must match the layer name. To use the layer, you must save the file in the current folder or in a folder on the MATLAB path.

Declare Properties and Learnable Parameters

Declare the layer properties in the properties section and declare learnable parameters by listing them in the properties (Learnable) section.

By default, custom intermediate layers have these properties:

PropertyDescription
Name Layer name, specified as a character vector or a string scalar. To include a layer in a layer graph, you must specify a nonempty unique layer name. If you train a series network with the layer and Name is set to '', then the software automatically assigns a name to the layer at training time.
Description

One-line description of the layer, specified as a character vector or a string scalar. This description appears when the layer is displayed in a Layer array. If you do not specify a layer description, then the software displays the layer class name.

TypeType of the layer, specified as a character vector or a string scalar. The value of Type appears when the layer is displayed in a Layer array. If you do not specify a layer type, then the software displays the layer class name.
NumInputsNumber of inputs of the layer specified as a positive integer. If you do not specify this value, then the software automatically sets NumInputs to the number of names in InputNames. The default value is 1.
InputNamesThe input names of the layer specified as a cell array of character vectors. If you do not specify this value and NumInputs is greater than 1, then the software automatically sets InputNames to {'in1',...,'inN'}, where N is equal to NumInputs. The default value is {'in'}.
NumOutputsNumber of outputs of the layer specified as a positive integer. If you do not specify this value, then the software automatically sets NumOutputs to the number of names in OutputNames. The default value is 1.
OutputNamesThe output names of the layer specified as a cell array of character vectors. If you do not specify this value and NumOutputs is greater than 1, then the software automatically sets OutputNames to {'out1',...,'outM'}, where M is equal to NumOutputs. The default value is {'out'}.

If the layer has no other properties, then you can omit the properties section.

Tip

If you are creating a layer with multiple inputs, then you must set either the NumInputs or InputNames in the layer constructor. If you are creating a layer with multiple outputs, then you must set either the NumOutputs or OutputNames in the layer constructor. For an example, see Define Custom Deep Learning Layer with Multiple Inputs.

A PReLU layer does not require any additional properties, so you can remove the properties section.

A PReLU layer has only one learnable parameter, the scaling coefficient a. Declare this learnable parameter in the properties (Learnable) section and call the parameter Alpha.

properties (Learnable)
% Layer learnable parameters

% Scaling coefficient
Alpha
end

Create Constructor Function

Create the function that constructs the layer and initializes the layer properties. Specify any variables required to create the layer as inputs to the constructor function.

The PReLU layer constructor function requires two inputs arguments: the number of channels of the expected input data and the layer name. The number of channels specifies the size of the learnable parameter Alpha. Specify two input arguments named numChannels and name in the preluLayer function. Add a comment to the top of the function that explains the syntax of the function.

function layer = preluLayer(numChannels, name)
% layer = preluLayer(numChannels) creates a PReLU layer with
% numChannels channels and specifies the layer name.

...
end

Initialize Layer Properties

Initialize the layer properties, including learnable parameters in the constructor function. Replace the comment % Layer constructor function goes here with code that initializes the layer properties.

Set the Name property to the input argument name.

% Set layer name.
layer.Name = name;

Give the layer a one-line description by setting the Description property of the layer. Set the description to describe the type of layer and its size.

% Set layer description.
layer.Description = "PReLU with " + numChannels + " channels";

For a PReLU layer, when the input values are negative, the layer multiplies each channel of the input by the corresponding channel of Alpha. Initialize the learnable parameter Alpha to be a random vector of size 1-by-1-by-numChannels. With the third dimension specified as size numChannels, the layer can use element-wise multiplication of the input in the forward function. Alpha is a property of the layer object, so you must assign the vector to layer.Alpha.

% Initialize scaling coefficient.
layer.Alpha = rand([1 1 numChannels]);

View the completed constructor function.

function layer = preluLayer(numChannels, name)
% layer = preluLayer(numChannels, name) creates a PReLU layer
% with numChannels channels and specifies the layer name.

% Set layer name.
layer.Name = name;

% Set layer description.
layer.Description = "PReLU with " + numChannels + " channels";

% Initialize scaling coefficient.
layer.Alpha = rand([1 1 numChannels]);
end

With this constructor function, the command preluLayer(3,'prelu') creates a PReLU layer with three channels and the name 'prelu'.

Create Forward Functions

Create the layer forward functions to use at prediction time and training time.

Create a function named predict that propagates the data forward through the layer at prediction time and outputs the result.

The syntax for predict is

[Z1,…,Zm] = predict(layer,X1,…,Xn)
where X1,…,Xn are the n layer inputs and Z1,…,Zm are the m layer outputs. The values n and m must correspond to the NumInputs and NumOutputs properties of the layer.

Tip

If the number of inputs to predict can vary, then use varargin instead of X1,…,Xn. In this case, varargin is a cell array of the inputs, where varargin{i} corresponds to Xi. If the number of outputs can vary, then use varargout instead of Z1,…,Zm. In this case, varargout is a cell array of the outputs, where varargout{j} corresponds to Zj.

Because a PReLU layer has only one input and one output, the syntax for predict for a PReLU layer is Z = predict(layer,X).

By default, the layer uses predict as the forward function at training time. To use a different forward function at training time, or retain a value required for the backward function, you must also create a function named forward.

The dimensions of the inputs depend on the type of data and the output of the connected layers:

Layer InputInput SizeObservation Dimension
2-D imagesh-by-w-by-c-by-N, where h, w, and c correspond to the height, width, and number of channels of the images respectively, and N is the number of observations.4
3-D imagesh-by-w-by-D-by-c-by-N, where h, w, D, and c correspond to the height, width, depth, and number of channels of the 3-D images respectively, and N is the number of observations.5
Vector sequencesc-by-N-by-S, where c is the number of features of the sequences, N is the number of observations, and S is the sequence length.2
2-D image sequencesh-by-w-by-c-by-N-by-S, where h, w, and c correspond to the height, width, and number of channels of the images respectively, N is the number of observations, and S is the sequence length.4
3-D image sequencesh-by-w-by-d-by-c-by-N-by-S, where h, w, d, and c correspond to the height, width, depth, and number of channels of the 3-D images respectively, N is the number of observations, and S is the sequence length.5

The forward function propagates the data forward through the layer at training time and also outputs a memory value.

The syntax for forward is

[Z1,…,Zm,memory] = forward(layer,X1,…,Xn)
where X1,…,Xn are the n layer inputs, Z1,…,Zm are the m layer outputs, and memory is the memory of the layer.

Tip

If the number of inputs to forward can vary, then use varargin instead of X1,…,Xn. In this case, varargin is a cell array of the inputs, where varargin{i} corresponds to Xi. If the number of outputs can vary, then use varargout instead of Z1,…,Zm. In this case, varargout is a cell array of the outputs, where varargout{j} corresponds to Zj for j=1,…,NumOutputs and varargout{NumOutputs+1} corresponds to memory.

The PReLU operation is given by

where ${x}_{i}$ is the input of the nonlinear activation f on channel i, and ${\alpha }_{i}$ is the coefficient controlling the slope of the negative part. The subscript i in ${\alpha }_{i}$ indicates that the nonlinear activation can vary on different channels.

Implement this operation in predict. In predict, the input X corresponds to x in the equation. The output Z corresponds to $f\left({x}_{i}\right)$. The PReLU layer does not require memory or a different forward function for training, so you can remove the forward function from the class file. Add a comment to the top of the function that explains the syntaxes of the function.

function Z = predict(layer, X)
% Z = predict(layer, X) forwards the input data X through the
% layer and outputs the result Z.

Z = max(0, X) + layer.Alpha .* min(0, X);
end

Create Backward Function

Implement the derivatives of the loss with respect to the input data and the learnable parameters in the backward function.

The syntax for backward is

[dLdX1,…,dLdXn,dLdW1,…,dLdWk] = backward(layer,X1,…,Xn,Z1,…,Zm,dLdZ1,…,dLdZm,memory)
where X1,…,Xn are the n layer inputs, Z1,…,Zm are the m outputs of forward, dLdZ1,…,dLdZm are the gradients backward propagated from the next layer, and memory is the memory output of forward. For the outputs, dLdX1,…,dLdXn are the derivatives of the loss with respect to the layer inputs and dLdW1,…,dLdWk are the derivatives of the loss with respect to the k learnable parameters. To reduce memory usage by preventing unused variables being saved between the forward and backward pass, replace the corresponding input arguments with ~.

Tip

If the number of inputs to backward can vary, then use varargin instead of the input arguments after layer. In this case, varargin is a cell array of the inputs, where varargin{i} corresponds to Xi for i=1,…,NumInputs, varargin{NumInputs+j} and varargin{NumInputs+NumOutputs+j} correspond to Zj and dLdZj, respectively, for j=1,…,NumOutputs, and varargin{end} corresponds to memory.

If the number of outputs can vary, then use varargout instead of the output arguments. In this case, varargout is a cell array of the outputs, where varargout{i} corresponds to dLdXi for i=1,…,NumInputs and varargout{NumInputs+t} corresponds to dLdWt for t=1,…,k, where k is the number of learnable parameters.

Because a PReLU layer has only one input, one output, and one learnable parameter, the syntax for backward for a PReLU layer is [dLdX,dLdAlpha] = backward(layer,X,Z,dLdZ,memory). The dimensions of X and Z are the same as in the forward functions. The dimensions of dLdZ are the same as the dimensions of Z. The dimensions and data type of dLdX are the same as the dimensions and data type of X. The dimension and data type of dLdAlpha is the same as the dimension and data type of the learnable parameter Alpha.

During the backward pass, the layer automatically updates the learnable parameters using the corresponding derivatives.

To include a custom layer in a network, the layer forward functions must accept the outputs of the previous layer and forward propagate arrays with the size expected by the next layer. Similarly, backward must accept inputs with the same size as the corresponding output of the forward function and backward propagate derivatives with the same size.

The derivative of the loss with respect to the input data is

$\frac{\partial L}{\partial {x}_{i}}=\frac{\partial L}{\partial f\left({x}_{i}\right)}\frac{\partial f\left({x}_{i}\right)}{\partial {x}_{i}}$

where $\partial L/\partial f\left({x}_{i}\right)$ is the gradient propagated from the next layer, and the derivative of the activation is

The derivative of the loss with respect to the learnable parameters is

$\frac{\partial L}{\partial {\alpha }_{i}}=\sum _{j}^{}\frac{\partial L}{\partial f\left({x}_{ij}\right)}\frac{\partial f\left({x}_{ij}\right)}{\partial {\alpha }_{i}}$

where i indexes the channels, j indexes the elements over height, width, and observations, and $\partial L/\partial f\left({x}_{i}\right)$ is the gradient propagated from the deeper layer, and the gradient of the activation is

In backward of the layer template, replace the output dLdW with the output dLdAlpha, where dLdAlpha corresponds to $\partial L/\partial {\alpha }_{i}$. In backward, the input X corresponds to x. The input Z corresponds to $f\left({x}_{i}\right)$. The input dLdZ corresponds to $\partial L/\partial f\left({x}_{i}\right)$. The output dLdX corresponds to $\partial L/\partial {x}_{i}$.

Add a comment to the top of the function that explains the syntaxes of the function. To reduce memory usage by preventing unused variables being saved between the forward and backward pass, replace the corresponding input arguments with ~. Because the layer function does not require the input arguments Z and memory, replace these arguments with ~.

function [dLdX, dLdAlpha] = backward(layer, X, ~, dLdZ, ~)
% [dLdX, dLdAlpha] = backward(layer, X, ~, dLdZ, ~)
% backward propagates the derivative of the loss function
% through the layer.
% Inputs:
%         layer    - Layer to backward propagate through
%         X        - Input data
%         dLdZ     - Gradient propagated from the deeper layer
% Outputs:
%         dLdX     - Derivative of the loss with respect to the
%                    input data
%         dLdAlpha - Derivative of the loss with respect to the
%                    learnable parameter Alpha

dLdX = layer.Alpha .* dLdZ;
dLdX(X>0) = dLdZ(X>0);
dLdAlpha = min(0,X) .* dLdZ;
dLdAlpha = sum(sum(dLdAlpha,1),2);

% Sum over all observations in mini-batch.
dLdAlpha = sum(dLdAlpha,4);
end

Completed Layer

View the completed layer class file.

classdef preluLayer < nnet.layer.Layer
% Example custom PReLU layer.

properties (Learnable)
% Layer learnable parameters

% Scaling coefficient
Alpha
end

methods
function layer = preluLayer(numChannels, name)
% layer = preluLayer(numChannels, name) creates a PReLU layer
% with numChannels channels and specifies the layer name.

% Set layer name.
layer.Name = name;

% Set layer description.
layer.Description = "PReLU with " + numChannels + " channels";

% Initialize scaling coefficient.
layer.Alpha = rand([1 1 numChannels]);
end

function Z = predict(layer, X)
% Z = predict(layer, X) forwards the input data X through the
% layer and outputs the result Z.

Z = max(0, X) + layer.Alpha .* min(0, X);
end

function [dLdX, dLdAlpha] = backward(layer, X, ~, dLdZ, ~)
% [dLdX, dLdAlpha] = backward(layer, X, ~, dLdZ, ~)
% backward propagates the derivative of the loss function
% through the layer.
% Inputs:
%         layer    - Layer to backward propagate through
%         X        - Input data
%         dLdZ     - Gradient propagated from the deeper layer
% Outputs:
%         dLdX     - Derivative of the loss with respect to the
%                    input data
%         dLdAlpha - Derivative of the loss with respect to the
%                    learnable parameter Alpha

dLdX = layer.Alpha .* dLdZ;
dLdX(X>0) = dLdZ(X>0);
dLdAlpha = min(0,X) .* dLdZ;
dLdAlpha = sum(sum(dLdAlpha,1),2);

% Sum over all observations in mini-batch.
dLdAlpha = sum(dLdAlpha,4);
end
end
end

GPU Compatibility

For GPU compatibility, the layer functions must support inputs and return outputs of type gpuArray. Any other functions the layer uses must do the same. Many MATLAB built-in functions support gpuArray input arguments. If you call any of these functions with at least one gpuArray input, then the function executes on the GPU and returns a gpuArray output. For a list of functions that execute on a GPU, see Run MATLAB Functions on a GPU (Parallel Computing Toolbox). To use a GPU for deep learning, you must also have a CUDA® enabled NVIDIA® GPU with compute capability 3.0 or higher. For more information on working with GPUs in MATLAB, see GPU Computing in MATLAB (Parallel Computing Toolbox).

The MATLAB functions used in predict, forward, and backward all support gpuArray inputs, so the layer is GPU compatible.

Check Validity of Layer Using checkLayer

Check the layer validity of the custom layer preluLayer.

Define a custom PReLU layer. To create this layer, save the file preluLayer.m in the current folder.

Create an instance of the layer and check its validity using checkLayer. Specify the valid input size to be the size of a single observation of typical input to the layer. The layer expects 4-D array inputs, where the first three dimensions correspond to the height, width, and number of channels of the previous layer output, and the fourth dimension corresponds to the observations.

Specify the typical size of the input of an observation and set 'ObservationDimension' to 4.

layer = preluLayer(20,'prelu');
validInputSize = [24 24 20];
checkLayer(layer,validInputSize,'ObservationDimension',4)
Skipping GPU tests. No compatible GPU device found.

Running nnet.checklayer.TestCase
.......... ........
Done nnet.checklayer.TestCase
__________

Test Summary:
18 Passed, 0 Failed, 0 Incomplete, 6 Skipped.
Time elapsed: 106.6761 seconds.

Here, the function does not detect any issues with the layer.

Include Custom Layer in Network

You can use a custom layer in the same way as any other layer in Deep Learning Toolbox. This section shows how to create and train a network for digit classification using the PReLU layer you created earlier.

Load the example training data.

[XTrain,YTrain] = digitTrain4DArrayData;

Define a custom PReLU layer. To create this layer, save the file preluLayer.m in the current folder. Create a layer array including the custom layer preluLayer.

layers = [
imageInputLayer([28 28 1])
convolution2dLayer(5,20)
batchNormalizationLayer
preluLayer(20,'prelu')
fullyConnectedLayer(10)
softmaxLayer
classificationLayer];

Set the training options and train the network.

net = trainNetwork(XTrain,YTrain,layers,options);
Training on single CPU.
Initializing input data normalization.
|========================================================================================|
|  Epoch  |  Iteration  |  Time Elapsed  |  Mini-batch  |  Mini-batch  |  Base Learning  |
|         |             |   (hh:mm:ss)   |   Accuracy   |     Loss     |      Rate       |
|========================================================================================|
|       1 |           1 |       00:00:00 |        7.03% |       3.3828 |          0.0010 |
|       2 |          50 |       00:00:11 |       74.22% |       0.7206 |          0.0010 |
|       3 |         100 |       00:00:21 |       89.84% |       0.3583 |          0.0010 |
|       4 |         150 |       00:00:32 |       88.28% |       0.4037 |          0.0010 |
|       6 |         200 |       00:00:45 |       96.88% |       0.2034 |          0.0010 |
|       7 |         250 |       00:00:57 |       96.88% |       0.1370 |          0.0010 |
|       8 |         300 |       00:01:09 |      100.00% |       0.0609 |          0.0010 |
|       9 |         350 |       00:01:22 |      100.00% |       0.0534 |          0.0010 |
|      10 |         390 |       00:01:30 |       99.22% |       0.0527 |          0.0010 |
|========================================================================================|

Evaluate the network performance by predicting on new data and calculating the accuracy.

[XTest,YTest] = digitTest4DArrayData;
YPred = classify(net,XTest);
accuracy = sum(YTest==YPred)/numel(YTest)
accuracy = 0.9194

References

[1] He, Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. "Delving deep into rectifiers: Surpassing human-level performance on ImageNet classification." In Proceedings of the IEEE international conference on computer vision, pp. 1026-1034. 2015.