Class ResNetConfiguration

Namespace

AiDotNet.Configuration

Assembly

AiDotNet.dll

Configuration options for ResNet (Residual Network) neural network architectures.

public class ResNetConfiguration

Inheritance

object

ResNetConfiguration

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.MemberwiseClone()

object.ReferenceEquals(object, object)

object.ToString()

Remarks

This configuration class provides all the settings needed to instantiate a ResNet network. It follows the AiDotNet pattern where users provide minimal configuration and the library supplies sensible defaults.

For Beginners: ResNet networks are deep convolutional neural networks that use skip connections to enable training of very deep architectures. This configuration lets you choose which ResNet variant to use (ResNet18, ResNet34, ResNet50, ResNet101, or ResNet152), set the number of output classes for your classification task, and optionally customize the input image dimensions and other parameters.

Typical Usage:

var config = new ResNetConfiguration(ResNetVariant.ResNet50, numClasses: 1000);
var architecture = config.ToArchitecture<float>();
var network = new ResNetNetwork<float>(architecture);

Constructors

ResNetConfiguration(ResNetVariant, int, int, int, int, bool, bool)

Initializes a new instance of the ResNetConfiguration class.

public ResNetConfiguration(ResNetVariant variant, int numClasses, int inputHeight = 224, int inputWidth = 224, int inputChannels = 3, bool includeClassifier = true, bool zeroInitResidual = true)

Parameters

variant ResNetVariant

The ResNet variant to use.

numClasses int

The number of output classes for classification.

inputHeight int

The height of input images (default: 224).

inputWidth int

The width of input images (default: 224).

inputChannels int

The number of input channels (default: 3 for RGB).

includeClassifier bool

Whether to include the classifier layers (default: true).

zeroInitResidual bool

Whether to use zero-initialization for residual branches (default: true).

Exceptions

ArgumentOutOfRangeException

Thrown when numClasses is less than or equal to 0, or when image dimensions are invalid.

Properties

BaseChannels

Gets the base channel counts for each stage.

public int[] BaseChannels { get; }

Property Value

int[]

Remarks

For BasicBlock: actual channels = base channels For BottleneckBlock: actual channels = base channels * 4 (due to expansion)

BlockCounts

Gets the block counts for each of the 4 stages based on the variant.

public int[] BlockCounts { get; }

Property Value

int[]

Remarks

ResNet architectures have 4 stages after the initial convolution:

• Stage 1 (conv2_x): 56x56 spatial, 64/256 channels
• Stage 2 (conv3_x): 28x28 spatial, 128/512 channels
• Stage 3 (conv4_x): 14x14 spatial, 256/1024 channels
• Stage 4 (conv5_x): 7x7 spatial, 512/2048 channels

Expansion

Gets the expansion factor for the blocks.

public int Expansion { get; }

Property Value

int

Remarks

BasicBlock has expansion 1, BottleneckBlock has expansion 4.

IncludeClassifier

Gets whether to include the fully connected classifier layer.

public bool IncludeClassifier { get; }

Property Value

bool

Remarks

Default is true. Set to false when using ResNet as a feature extractor for transfer learning.

For Beginners: The classifier is the final part of the network that produces class predictions. When using ResNet for transfer learning or as a feature extractor, you may want to remove the classifier and add your own custom layers.

InputChannels

Gets the number of input channels.

public int InputChannels { get; }

Property Value

int

Remarks

Default is 3 for RGB images. Use 1 for grayscale images.

For Beginners: Color images typically have 3 channels (Red, Green, Blue). Grayscale images have 1 channel. Most pre-trained ResNet models expect 3-channel inputs.

InputHeight

Gets the height of input images in pixels.

public int InputHeight { get; }

Property Value

int

Remarks

Default is 224, which is the standard ImageNet input size.

For Beginners: ResNet networks were designed for 224x224 images. While you can use different sizes, 224x224 is recommended for best results, especially when using pre-trained weights.

InputShape

Gets the computed input shape as [channels, height, width].

public int[] InputShape { get; }

Property Value

int[]

InputWidth

Gets the width of input images in pixels.

public int InputWidth { get; }

Property Value

int

Remarks

Default is 224, which is the standard ImageNet input size.

NumClasses

Gets the number of output classes for classification.

public int NumClasses { get; }

Property Value

int

Remarks

For Beginners: This is the number of categories your model will classify images into. For example, if you're classifying cats vs dogs, this would be 2. For ImageNet, it's 1000. For CIFAR-10, it's 10.

NumConvLayers

Gets the total number of convolutional layers in the network.

public int NumConvLayers { get; }

Property Value

int

NumWeightLayers

Gets the total number of weight layers (conv + FC).

public int NumWeightLayers { get; }

Property Value

int

TotalInputSize

Gets the total number of input features (channels * height * width).

public int TotalInputSize { get; }

Property Value

int

UseAutodiff

Gets or sets whether to use automatic differentiation for backpropagation.

public bool UseAutodiff { get; set; }

Property Value

bool

Remarks

Default is false, which uses the optimized manual backward implementation.

For Beginners: Backpropagation is how neural networks learn from their mistakes. Autodiff automatically computes gradients, which is more flexible but slightly slower. The manual implementation is optimized for ResNet's specific architecture.

UsesBottleneck

Gets whether this variant uses BasicBlock (ResNet18/34) or BottleneckBlock (ResNet50/101/152).

public bool UsesBottleneck { get; }

Property Value

bool

Variant

Gets the ResNet variant to use.

public ResNetVariant Variant { get; }

Property Value

ResNetVariant

Remarks

For Beginners: The variant determines how deep the network is. ResNet50 is the most commonly used variant. Deeper variants (ResNet101, ResNet152) have more capacity but require more computational resources.

ZeroInitResidual

Gets whether to use zero-initialization for the last batch normalization in each residual block.

public bool ZeroInitResidual { get; }

Property Value

bool

Remarks

Default is true. This technique (from "Bag of Tricks" paper) improves training stability by initializing the last BN layer in each residual branch to zero, making the initial residual branch act as an identity.

For Beginners: This is a training trick that helps the network start learning more effectively. When enabled, each residual block initially acts like an identity function, making the network easier to optimize.

Methods

CreateForCIFAR(ResNetVariant, int)

Creates a configuration optimized for CIFAR-10/CIFAR-100 datasets.

public static ResNetConfiguration CreateForCIFAR(ResNetVariant variant, int numClasses)

Parameters

variant ResNetVariant

The ResNet variant to use.

numClasses int

The number of output classes (10 for CIFAR-10, 100 for CIFAR-100).

Returns

ResNetConfiguration

A ResNetConfiguration optimized for CIFAR datasets.

Remarks

For Beginners: CIFAR images are 32x32 pixels, which is smaller than ImageNet's 224x224. This configuration adjusts the input size accordingly. Note that for CIFAR-sized images, a modified ResNet architecture is typically used that removes the initial 7x7 conv and pooling.

CreateForTesting(int)

Creates a minimal ResNet configuration optimized for fast test execution.

public static ResNetConfiguration CreateForTesting(int numClasses)

Parameters

numClasses int

The number of output classes.

Returns

ResNetConfiguration

A minimal ResNet configuration for testing.

Remarks

Uses ResNet18 (smallest variant) with 32x32 input resolution, resulting in minimal network construction time suitable for unit tests. Construction time is typically under 50ms.

CreateLightweight(int)

Creates a lightweight configuration using ResNet18.

public static ResNetConfiguration CreateLightweight(int numClasses)

Parameters

numClasses int

The number of output classes.

Returns

ResNetConfiguration

A ResNetConfiguration for ResNet18.

Remarks

For Beginners: Use this when you need faster training and inference, or when working with smaller datasets where larger models might overfit.

CreateResNet50(int)

Creates a new configuration for ResNet50.

public static ResNetConfiguration CreateResNet50(int numClasses)

Parameters

numClasses int

The number of output classes.

Returns

ResNetConfiguration

A ResNetConfiguration for ResNet50.

Remarks

For Beginners: This is the recommended ResNet configuration for most use cases. ResNet50 provides a good balance of accuracy and computational efficiency.