focalLossLayer

(To be removed) Create focal loss layer using focal loss function

Since R2020a

The FocalLossLayer object will be removed in a future release. Use the trainnet (Deep Learning Toolbox) function and specify the loss as the focalCrossEntropy function. For more information, see Version History.

Description

A focal loss layer predicts object classes using focal loss. Add the focal loss layer to train an object detection, semantic segmentation, or a classification network when imbalance exists between foreground and background classes. To compensate for class imbalance, the focal loss function multiplies the cross entropy function with a modulating factor that increases the sensitivity of the network to misclassified observations.

Creation

Syntax

layer = focalLossLayer

layer = focalLossLayer(Name,Value)

Description

layer = focalLossLayer creates a focal loss layer for deep learning networks.

layer = focalLossLayer(Name,Value) sets properties of the focal loss layer by using one or more name-value pair arguments. Enclose each property name in quotes.

For example, focalLossLayer('Name','focalloss') creates a focal loss layer with the name 'focalloss' and the specified balancing and focusing parameters.

Properties

Alpha — Balancing parameter

0.25 (default) | positive real number

Balancing parameter of the focal loss function, specified as a positive real number. The Alpha value scales the loss function linearly and is typically set to 0.25. If you decrease Alpha, increase Gamma.

Gamma — Focusing parameter

2.0 (default) | positive real number

Focusing parameter of the focal loss function, specified as a positive real number. Increasing the value of Gamma increases the sensitivity of the network to misclassified observations.

Classes — Object classes to detect

'auto' (default) | string vector | categorical vector | cell array of character vectors

Object classes that the network is trained to detect, specified as a string vector, categorical vector, cell array of character vectors, or 'auto'. When you set Classes to 'auto', the classes are automatically set at training time. When you specify a string vector or cell array of character vectors, then the elements of Classes are sorted according to the output of the categories function.

Data Types: string | categorical | cell | char

Name — Layer name

"" (default) | character vector | string scalar

Layer name, specified as a character vector or string scalar. For Layer array input, the trainnet (Deep Learning Toolbox) and dlnetwork (Deep Learning Toolbox) functions automatically assign names to layers with the name "".

The FocalLossLayer object stores this property as a character vector.

Data Types: char | string

Examples

Create Focal Loss Layer

Specify class names.

classes = ["Vehicle","Background"];

Specify the balancing parameter and focusing parameter of the focal loss function. Create a focal loss layer named "focallosslayer" for the two classes, displaying results.

layer = focalLossLayer(Classes=classes,Name="focallosslayer")

layer = 
  FocalLossLayer with properties:

            Name: 'focallosslayer'

   Hyperparameters
           Gamma: 2
           Alpha: 0.2500
         Classes: [2x1 categorical]
    LossFunction: 'focalLoss'

Create 2-D Semantic Segmentation Network with Focal Loss Layer

Create a DeepLab v3+ network based on ResNet-18.

imageSize = [480 640 3];
numClasses = 5;
network = "resnet18";
lgraph = deeplabv3plusLayers(imageSize,numClasses,network,DownsamplingFactor=16)

lgraph = 
  LayerGraph with properties:

     InputNames: {'data'}
    OutputNames: {'classification'}
         Layers: [100x1 nnet.cnn.layer.Layer]
    Connections: [113x2 table]

Display the output layer of the network. The output layer of the DeepLab v3+ network is a PixelClassificationLayer that uses cross-entropy loss to predict the categorical label for every pixel in an input 2-D image.

lgraph.Layers(end)

ans = 
  PixelClassificationLayer with properties:

            Name: 'classification'
         Classes: 'auto'
    ClassWeights: 'none'
      OutputSize: 'auto'

   Hyperparameters
    LossFunction: 'crossentropyex'

Replace the output PixelClassificationLayer with the FocalLossLayer to handle imbalanced classes in data.

layer = focalLossLayer("Name","focalloss");
lgraph = replaceLayer(lgraph,"classification",layer);

Display the network.

analyzeNetwork(lgraph);

Create 3-D Semantic Segmentation Network with Focal Loss Layer

Create a 3-D U-Net network for semantic segmentation by using the unet3dLayers function. Set the encoder-decoder depth to 2 and specify the number of output channels for the first convolution layer as 16.

imageSize = [128 128 128 3];
numClasses = 5;
lgraph = unet3dLayers(imageSize,numClasses, ...
    EncoderDepth=2,NumFirstEncoderFilters=16);
figure
plot(lgraph)

Create a focal loss layer and replace the Segmentation-Layer in the network with the focal loss layer. The layer predicts the categorical label for every voxel in an input 3-D volume.

layer = focalLossLayer(Name="focalloss");
lgraph = replaceLayer(lgraph,"Segmentation-Layer",layer)

lgraph = 
  LayerGraph with properties:

     InputNames: {'ImageInputLayer'}
    OutputNames: {'focalloss'}
         Layers: [40x1 nnet.cnn.layer.Layer]
    Connections: [41x2 table]

Display the network.

analyzeNetwork(lgraph);

More About

Focal Loss

The focal loss function is based on cross-entropy loss. Focal loss compensates for class imbalance by using a modulating factor that emphasizes hard negatives during training.

The focal loss function, L, used by the focalLossLayer object for the loss between one image Y and the corresponding ground truth T is given by:

$$L=-\frac{1}{M}{\displaystyle \sum _{m=1}^M{\displaystyle \sum _{k=1}^K{T}_{mk}\alpha {\left(1-Y_{mk}\right)}^{\gamma }\ln \left(Y_{mk}\right)}}$$

K is the number of classes. M is the number of observations, which is typically the number of elements along the first two dimensions of Y multiplied by the number of anchor boxes. The function used by the focalLossLayer object ensures that the input to the natural logarithm is greater than 0.

References

[1] Lin, Tsung-Yi, Priya Goyal, Ross Girshick, Kaiming He, and Piotr Dollar. "Focal Loss for Dense Object Detection." In 2017 IEEE^® International Conference on Computer Vision (ICCV), 2999–3007. Venice: IEEE, 2017. https://doi.org/10.1109/ICCV.2017.324.

Extended Capabilities

C/C++ Code Generation

Generate C and C++ code using MATLAB® Coder™.

Usage notes and limitations:

The code generator represents characters in an 8-bit ASCII codeset that the locale setting determines. Therefore, the use of non-ASCII characters in class names, layer names, layer description, or network names might result in errors. For more information, see Encoding of Characters in Code Generation (MATLAB Coder).

GPU Code Generation

Generate CUDA® code for NVIDIA® GPUs using GPU Coder™.

Usage notes and limitations:

• To generate CUDA^® or C++ code by using GPU Coder™, you must first construct and train a deep neural network. Once the network is trained and evaluated, you can configure the code generator to generate code and deploy the convolutional neural network on platforms that use NVIDIA^® or ARM^® GPU processors. For more information, see Deep Learning with GPU Coder (GPU Coder).

• For this layer, you can generate code that takes advantage of the NVIDIA CUDA deep neural network library (cuDNN), NVIDIA TensorRT™ high performance inference library, or the ARM Compute Library for Mali GPU.

• The code generator represents characters in an 8-bit ASCII codeset that the locale setting determines. Therefore, the use of non-ASCII characters in class names, layer names, layer description, or network names might result in errors. For more information, see Encoding of Characters in Code Generation (MATLAB Coder).

Version History

Introduced in R2020a

R2024a: focalLossLayer will be removed

The focalLossLayer object will be removed in a future release. Follow these steps to update your code:

• Define your network as a dlnetwork (Deep Learning Toolbox) object. You can use functions such as addLayers (Deep Learning Toolbox) and connectLayers (Deep Learning Toolbox) to build the network. Do not include output layers in the network.

• Train the network using the trainnet (Deep Learning Toolbox) function and specify the loss using the focalCrossEntropy function. For example, this code trains a dlnetwork network called net using the training data images.

  netTrained = trainnet(images,net,@focalCrossEntropy,options);