resnetNetwork

2-D residual neural network

Since R2024a

Syntax

net = resnetNetwork(inputSize,numClasses)

net = resnetNetwork(inputSize,numClasses,Name=Value)

Description

net = resnetNetwork(inputSize,numClasses) creates a 2-D residual neural network with the specified image input size and number of classes.

To create a 3-D residual network, use resnet3dNetwork.

net = resnetNetwork(inputSize,numClasses,Name=Value) also specifies additional options using one or more name-value arguments.

Tip

To load a pretrained ResNet neural network, use the imagePretrainedNetwork function.

Examples

Create Residual Network

Create a residual network with a bottleneck architecture.

imageSize = [224 224 3];
numClasses = 10;

net = resnetNetwork(imageSize,numClasses)

net = 
  dlnetwork with properties:

         Layers: [176x1 nnet.cnn.layer.Layer]
    Connections: [191x2 table]
     Learnables: [214x3 table]
          State: [106x3 table]
     InputNames: {'input'}
    OutputNames: {'softmax'}
    Initialized: 1

  View summary with summary.

Analyze the network using the analyzeNetwork function.

analyzeNetwork(net)

This network is equivalent to a ResNet-50 residual network.

Residual Network with Custom Stack Depth

Create a ResNet-101 network using a custom stack depth.

imageSize = [224 224 3];
numClasses = 10;

stackDepth = [3 4 23 3];
numFilters = [64 128 256 512];

net = resnetNetwork(imageSize,numClasses, ...
    StackDepth=stackDepth, ...
    NumFilters=numFilters)

net = 
  dlnetwork with properties:

         Layers: [346x1 nnet.cnn.layer.Layer]
    Connections: [378x2 table]
     Learnables: [418x3 table]
          State: [208x3 table]
     InputNames: {'input'}
    OutputNames: {'softmax'}
    Initialized: 1

  View summary with summary.

Analyze the network.

analyzeNetwork(net)

Input Arguments

inputSize — Network image input size

vector of positive integers

Network image input size, specified as one of these values:

• Vector of positive integers [h w] — Input has a height and width of h and w, respectively.

• Vector of positive integers [h w c] — Input has a height, width, and number of channels of h, w, and c, respectively. For RGB images, c is 3, for grayscale images, c to 1.

The values of inputSize depend on the InitialPoolingLayer option:

• If InitialPoolingLayer is "max" or "average", then h and w must be greater than or equal to the corresponding entries in 2*InitialStride*2^D, where D is the number of downsampling blocks.

• If InitialPoolingLayer is "none", then h and w must be greater than or equal to the corresponding entries in InitialStride*2^D.

Data Types: single | double | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32 | uint64

numClasses — Number of classes for classification tasks

positive integer

Number of classes for classification tasks, specified as a positive integer.

The function returns a neural network for classification tasks with the specified number of classes by setting the output size of the last fully connected layer to numClasses.

Data Types: single | double | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32 | uint64

Name-Value Arguments

Specify optional pairs of arguments as Name1=Value1,...,NameN=ValueN, where Name is the argument name and Value is the corresponding value. Name-value arguments must appear after other arguments, but the order of the pairs does not matter.

Example: net = resnetNetwork(inputSize,numClasses,BottleneckType="none") returns a 2-D residual neural network without bottleneck components.

Initial Layers

InitialFilterSize — Filter size in first convolutional layer

7 (default) | positive integer | vector of two positive integers

Filter size in the first convolutional layer, specified as one of these values:

• Positive integer — First convolutional layer has filters with a height and width of the specified value.

• Vector of positive integers [h w] — First convolutional layer has a filters with the height and width of h and w, respectively.

Data Types: single | double | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32 | uint64

InitialNumFilters — Number of filters in first convolutional layer

64 (default) | positive integer

Number of filters in the first convolutional layer, specified as a positive integer. The number of initial filters determines the number of channels (feature maps) in the output of the first convolutional layer in the residual network.

Data Types: single | double | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32 | uint64

InitialStride — Stride in first convolutional layer

2 (default) | positive integer | vector of two positive integers

Stride in the first convolutional layer, specified as one of these values:

• Positive integer — First convolutional layer has a vertical and horizontal stride of the specified value.

• Vector of positive integers [h w] — First convolutional layer has a vertical and horizontal stride of h and w, respectively.

The stride defines the step size for traversing the input vertically and horizontally.

Data Types: single | double | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32 | uint64

InitialPoolingLayer — First pooling layer

"max" (default) | "average" | "none"

First pooling layer before the initial residual block, specified as one of these values:

• "max" — Use a max pooling layer before the initial residual block. For more information, see maxPooling2dLayer.

• "average" — Use an average pooling layer before the initial residual block. For more information, see averagePooling2dLayer.

• "none"— Do not use a pooling layer before the initial residual block.

Network Architecture

ResidualBlockType — Residual block type

"batchnorm-before-add" (default) | "batchnorm-after-add"

Residual block type, specified as one of these values:

• "batchnorm-before-add" — Include the batch normalization layer before the addition layer in the residual blocks [1].

• "batchnorm-after-add" — Include the batch normalization layer after the addition layer in the residual blocks [2].

The ResidualBlockType argument specifies the location of the batch normalization layer in the standard and downsampling residual blocks. For more information, see More About.

BottleneckType — Block bottleneck type

"downsample-first-conv" (default) | "none"

Block bottleneck type, specified as one of these values:

• "downsample-first-conv" — Use bottleneck residual blocks that perform downsampling in the first convolutional layer of the downsampling residual blocks, using a stride of 2. A bottleneck residual block consists of three layers: a convolutional layer with filters of size 1 for downsampling the channel dimension, a convolutional layer with filters of size 3, and a convolutional layer with filters of size 1 for upsampling the channel dimension.

  The number of filters in the final convolutional layer is four times that in the first two convolutional layers.

• "none" — Do not use bottleneck residual blocks. The residual blocks consist of two convolutional layers with filters of size 3.

A bottleneck block reduces the number of channels by a factor of four by performing a convolution with filters of size 1 before performing convolution with filters of size 3. Networks with and without bottleneck blocks have a similar level of computational complexity, but the total number of features propagating in the residual connections is four times larger when you use bottleneck units. Therefore, using a bottleneck increases the efficiency of the network [1].

For more information on the layers in each residual block, see Residual Network.

StackDepth — Number of residual blocks in each stack

[3 4 6 3] (default) | vector of positive integers

Number of residual blocks in each stack, specified as a vector of positive integers.

For example, if the stack depth is [3 4 6 3], the network has four stacks, with three blocks, four blocks, six blocks, and three blocks.

Specify the number of filters in the convolutional layers of each stack using the NumFilters argument. StackDepth must have the same number of elements as NumFilters.

Data Types: single | double | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32 | uint64

NumFilters — Number of filters in convolutional layers of each stack

[64 128 256 512] (default) | vector of positive integers

Number of filters in the convolutional layers of each stack, specified as a vector of positive integers.

• If BottleneckType is "downsample-first-conv", then the number of filters in the first two convolutional layers in each block of each stack is NumFilters. The final convolutional layer has four times the number of filters in the first two convolutional layers.

  For example, if NumFilters is [4 5] and BottleneckType is "downsample-first-conv", then in the first stack, the first two convolutional layers in each block have 4 filters and the final convolutional layer in each block has 16 filters. In the second stack, the first two convolutional layers in each block have 5 filters and the final convolutional layer has 20 filters.

• If BottleneckType is "none", then the number of filters in the convolutional layers in each stack is NumFilters.

NumFilters must have the same number of elements as the StackDepth value.

The NumFilters value determines the layers on the residual connection in the initial residual block. There is a convolutional layer on the residual connection when one of these conditions is met:

• BottleneckType is "downsample-first-conv" and InitialNumFilters is not equal to four times the first element of NumFilters.

• BottleneckType is "none" and InitialNumFilters is not equal to the first element of NumFilters.

For more information about the layers in each residual block, see Residual Network.

Data Types: single | double | int8 | int16 | int32 | int64 | uint8 | uint16 | uint32 | uint64

Normalization — Data normalization

"zerocenter" (default) | "zscore"

Data normalization to apply every time data is forward-propagated through the input layer, specified as one of the following:

• "zerocenter" — Subtract the mean of the training data.

• "zscore" — Subtract the mean and then divide by the standard deviation of the training data.

The trainnet function automatically calculates the mean and standard deviation of the training data.

Initialize — Flag to initialize learnable parameters

1 (true) (default) | 0 (false)

Flag to initialize learnable parameters, specified as 1 (true) or 0 (false).

Output Arguments

net — Residual neural network

dlnetwork object

Residual network, returned as a dlnetwork object.

More About

Residual Network

Residual networks (ResNets) are a type of deep network that consist of building blocks that have residual connections (also known as skip or shortcut connections). These connections allow the input to skip the convolutional units of the main branch, thus providing a simpler path through the network. By allowing the parameter gradients to flow more easily from the final layers to the earlier layers of the network, residual connections help mitigate the problem of vanishing gradients during early training.

The structure of a residual network is flexible. The key component is the inclusion of the residual connections within residual blocks. A group of residual blocks is called a stack. A ResNet architecture consists of initial layers, followed by stacks containing residual blocks, and then the final layers. A network has three types of residual blocks:

• Initial residual block — This block occurs at the start of the first stack. The layers in the residual connection of the initial residual block determine if the block preserves the activation sizes or performs downsampling.

• Standard residual block — This block occurs multiple times in each stack, after the first downsampling residual block. The standard residual block preserves the activation sizes.

• Downsampling residual block — This block occurs once, at the start of each stack. The first convolutional unit in the downsampling block downsamples the spatial dimensions by a factor of two.

A typical stack has a downsampling residual block, followed by m standard residual blocks, where m is a positive integer. The first stack is the only stack that begins with an initial residual block.

The initial, standard, and downsampling residual blocks can be bottleneck or nonbottleneck blocks.

A bottleneck block reduces the number of channels by a factor of four by performing a convolution with filters of size 1 before performing convolution with filters of size 3. Networks with and without bottleneck blocks have a similar level of computational complexity, but the total number of features propagating in the residual connections is four times larger when you use bottleneck units. Therefore, using a bottleneck increases the efficiency of the network [1].

The options you set determine the layers inside each block.

Block Layers

Name	Initial Layers	Initial Residual Block	Standard Residual Block (BottleneckType="downsample-first-conv")	Standard Residual Block (BottleneckType="none")	Downsampling Residual Block	Final Layers
Description	A residual network starts with these layers, in order: imageInputLayer (or image3dInputLayer for 3-D networks) convolution2dLayer (or convolution3dLayer for 3-D networks) batchNormalizationLayer reluLayer (Optional) maxPooling2dLayer or averagePooling2dLayer (or maxPooling3dLayer or averagePooling3dLayer for 3-D networks) Set the optional pooling layer using the InitialPoolingLayer option.	The main branch of the initial residual block has the same layers as a standard residual block. The InitialNumFilters and NumFilters values determine the layers on the residual connection. The residual connection has a convolutional layer with filters and a stride of size 1 when one of these conditions is met: BottleneckType is "downsample-first-conv" and InitialNumFilters is not equal to four times the first element of NumFilters. BottleneckType is "none" and InitialNumFilters is not equal to the first element of NumFilters. If ResidualBlockType is "batchnorm-before-add", then the residual connection also has a batch normalization layer.	The standard residual block with bottleneck units has these layers, in order: convolution2dLayer (or convolution3dLayer for 3-D networks) with filters and a stride of size 1 batchNormalizationLayer reluLayer convolution2dLayer (or convolution3dLayer for 3-D networks) with filters of size 3 and a stride of size 1 batchNormalizationLayer reluLayer convolution2dLayer (or convolution3dLayer for 3-D networks) with filters and a stride of size 1 batchNormalizationLayer additionLayer reluLayer The standard block has a residual connection from the output of the previous block to the addition layer. Set the position of the addition layer using the ResidualBlockType argument.	The standard residual block without bottleneck units has these layers, in order: convolution2dLayer (or convolution3dLayer for 3-D networks) with filters of size 3 and a stride of size1 batchNormalizationLayer reluLayer convolution2dLayer (or convolution3dLayer for 3-D networks) with filters of size 3 and a stride of size 1 batchNormalizationLayer additionLayer reluLayer The standard block has a residual connection from the output of the previous block to the addition layer. Set the position of the addition layer using the ResidualBlockType argument.	The downsampling residual block is the same as the standard block (either with or without the bottleneck) but with a stride of size 2 in the first convolutional layer and additional layers on the residual connection. The layers on the residual connection depend on the value of ResidualBlockType. If ResidualBlockType is "batchnorm-before-add", then the second branch contains a convolution2dLayer (or convolution3dLayer for 3-D networks) with filters of size 1 and a stride of size 2, and a batchNormalizationLayer. If ResidualBlockType is "batchnorm-after-add", then the second branch contains a convolution2dLayer (or convolution3dLayer for 3-D networks) with filters of size 1 and a stride of 2. The downsampling block halves the height and width of the input, and increases the number of channels.	A residual network ends with these layers, in order: globalAveragePooling2dLayer (or globalAveragePooling3dLayer for 3-D networks) fullyConnectedLayer softmaxLayer
Example Visualization		Example of an initial residual block for a network without a bottleneck and with the batch normalization layer before the addition layer.	Example of the standard residual block for a network with a bottleneck and with the batch normalization layer before the addition layer.	Example of the standard residual block for a network without a bottleneck and with the batch normalization layer before the addition layer.	Example of a downsampling residual block for a network without a bottleneck and with the batch normalization layer before the addition layer.

The convolution and fully connected layer weights are initialized using the He weight initialization method [3].

Tips

• When working with small images, set the InitialPoolingLayer option to "none" to remove the initial pooling layer and reduce the amount of downsampling.

• Residual networks are usually named ResNet-X, where X is the depth of the network. The depth of a network is defined as the largest number of sequential convolutional or fully connected layers on a path from the network input to the network output. You can use the following formula to compute the depth of your network:

  $$\text{depth = }\{\begin{array}{c}1+2{\displaystyle \sum }_{i=1}^N{s}_i+1\text{ If no bottleneck}\\ 1+3{\displaystyle \sum }_{i=1}^N{s}_i+1\text{ If bottleneck}\end{array}\text{ ,}$$

  where s_i is the depth of stack i.

References

[1] He, Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. “Deep Residual Learning for Image Recognition.” Preprint, submitted December 10, 2015. https://arxiv.org/abs/1512.03385.

[2] He, Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. “Identity Mappings in Deep Residual Networks.” Preprint, submitted July 25, 2016. https://arxiv.org/abs/1603.05027.

[3] He, Kaiming, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. "Delving Deep into Rectifiers: Surpassing Human-Level Performance on ImageNet Classification." In Proceedings of the 2015 IEEE International Conference on Computer Vision, 1026–1034. Washington, DC: IEEE Computer Vision Society, 2015.

Version History

Introduced in R2024a