Class NASOptions<T>

Namespace

AiDotNet.Models.Options

Assembly

AiDotNet.dll

Configuration options for Neural Architecture Search (NAS).

public class NASOptions<T>

Type Parameters

T

The numeric type for calculations.

Inheritance

object

NASOptions<T>

Inherited Members

object.Equals(object)

object.Equals(object, object)

object.GetHashCode()

object.GetType()

object.MemberwiseClone()

object.ReferenceEquals(object, object)

object.ToString()

Remarks

For Beginners: NAS automatically discovers optimal neural network architectures. Instead of manually designing layers and connections, NAS explores the design space to find architectures that maximize accuracy while meeting hardware constraints.

Quick Start Example:

var options = new NASOptions<double>
{
    Strategy = AutoMLSearchStrategy.DARTS,
    TargetPlatform = HardwarePlatform.Mobile,
    MaxSearchTime = TimeSpan.FromHours(2),
    HardwareConstraints = new HardwareConstraints<double>
    {
        MaxLatency = 10.0,    // 10ms max inference time
        MaxMemory = 50.0      // 50MB max model size
    }
};

Available Strategies:

• DARTS: Fast gradient-based search (~1-2 GPU days)
• GDAS: Improved DARTS with better discretization
• OnceForAll: Train once, deploy anywhere with elastic networks
• NeuralArchitectureSearch: Auto-selects best algorithm

Properties

ArchitectureLearningRate

Gets or sets the learning rate for architecture parameters.

public double ArchitectureLearningRate { get; set; }

Property Value

double

Remarks

For Advanced Users: Controls how fast the architecture search converges. Lower values are more stable but slower.

CheckpointDirectory

Gets or sets the checkpoint directory path.

public string? CheckpointDirectory { get; set; }

Property Value

string

ElasticDepths

Gets or sets the elastic depth values for OFA networks.

public List<int>? ElasticDepths { get; set; }

Property Value

List<int>

Remarks

For OFA: List of possible layer counts that the network can use. Example: [2, 3, 4] means network can have 2, 3, or 4 layers per block.

ElasticExpansionRatios

Gets or sets the elastic expansion ratios for OFA inverted residuals.

public List<int>? ElasticExpansionRatios { get; set; }

Property Value

List<int>

Remarks

For OFA: List of expansion ratios for inverted residual blocks. Example: [3, 4, 6] for MobileNetV2-style blocks.

ElasticKernelSizes

Gets or sets the elastic kernel sizes for OFA networks.

public List<int>? ElasticKernelSizes { get; set; }

Property Value

List<int>

Remarks

For OFA: List of possible kernel sizes. Example: [3, 5, 7] means 3x3, 5x5, or 7x7 kernels.

ElasticWidths

Gets or sets the elastic width multipliers for OFA networks.

public List<double>? ElasticWidths { get; set; }

Property Value

List<double>

Remarks

For OFA: List of channel width multipliers. Example: [0.75, 1.0, 1.25] means 75%, 100%, or 125% of base channels.

Generations

Gets or sets the number of generations for evolutionary search.

public int Generations { get; set; }

Property Value

int

Remarks

For Evolutionary/OFA Specialization: Number of evolution cycles. More generations allow better convergence.

HardwareConstraints

Gets or sets the hardware constraints for the architecture search.

public HardwareConstraints<T>? HardwareConstraints { get; set; }

Property Value

HardwareConstraints<T>

Remarks

For Beginners: Set limits that your target device can handle:

HardwareConstraints = new HardwareConstraints<double>
{
    MaxLatency = 10.0,    // Max inference time in milliseconds
    MaxMemory = 50.0,     // Max model size in MB
    MaxEnergy = 100.0     // Max energy per inference in mJ
};

InputChannels

Gets or sets the number of input channels.

public int InputChannels { get; set; }

Property Value

int

Remarks

For Beginners: Number of channels in input data:

• 1: Grayscale images
• 3: RGB images
• 4: RGBA images

MaxEpochs

Gets or sets the maximum number of search epochs.

public int MaxEpochs { get; set; }

Property Value

int

Remarks

For Beginners: Number of training iterations for the architecture search. Higher values explore more thoroughly but take longer.

MaxSearchTime

Gets or sets the maximum time for the architecture search.

public TimeSpan MaxSearchTime { get; set; }

Property Value

TimeSpan

Remarks

For Beginners: Longer search times generally find better architectures. Typical values:

• Quick test: 30 minutes - 1 hour
• Standard: 2-8 hours
• Production: 24-48 hours

MutationProbability

Gets or sets the mutation probability for evolutionary search.

public double MutationProbability { get; set; }

Property Value

double

Remarks

For Evolutionary: Probability of mutating each gene in offspring. Typical values: 0.05-0.2

NumClasses

Gets or sets the number of output classes.

public int NumClasses { get; set; }

Property Value

int

Remarks

For Classification: Number of classes to predict. Example: 10 for CIFAR-10, 1000 for ImageNet.

OnEpochComplete

Gets or sets the callback invoked after each search epoch.

public Action<int, T>? OnEpochComplete { get; set; }

Property Value

Action<int, T>

Remarks

For Monitoring: Use to log progress or save checkpoints. Parameters are (epoch, current best score).

PopulationSize

Gets or sets the population size for evolutionary search.

public int PopulationSize { get; set; }

Property Value

int

Remarks

For Evolutionary/OFA Specialization: Number of candidate architectures to maintain in each generation. Larger populations explore more diversity.

QuantizationAware

Gets or sets whether to use quantization-aware training during NAS.

public bool QuantizationAware { get; set; }

Property Value

bool

Remarks

For Beginners: When enabled, architectures are searched with quantization in mind, resulting in models that maintain accuracy after quantization.

QuantizationMode

Gets or sets the quantization mode for quantization-aware NAS.

public QuantizationMode QuantizationMode { get; set; }

Property Value

QuantizationMode

Remarks

For Quantization-Aware NAS: The type of quantization to simulate:

• Int8: Most common for mobile/edge deployment
• Float16: Good balance of speed and accuracy
• Dynamic: Quantize weights, compute in float
• Mixed: Different precision for different layers

RandomSeed

Gets or sets the random seed for reproducibility.

public int? RandomSeed { get; set; }

Property Value

int?

Remarks

For Reproducibility: Set a seed to get repeatable results.

SaveCheckpoints

Gets or sets whether to save architecture checkpoints during search.

public bool SaveCheckpoints { get; set; }

Property Value

bool

SearchSpace

Gets or sets the search space configuration.

public SearchSpaceBase<T>? SearchSpace { get; set; }

Property Value

SearchSpaceBase<T>

Remarks

For Advanced Users: Customize the operations and structure that NAS can explore. If null, a default search space with standard operations is used.

SpatialSize

Gets or sets the spatial size of the input (assuming square inputs).

public int SpatialSize { get; set; }

Property Value

int

Remarks

For Beginners: Width/height of input images. Common values: 32 (CIFAR), 224 (ImageNet), 299 (Inception).

Strategy

Gets or sets the NAS strategy to use.

public AutoMLSearchStrategy Strategy { get; set; }

Property Value

AutoMLSearchStrategy

Remarks

For Beginners: Different strategies have different trade-offs:

• DARTS: Fast but may produce suboptimal architectures
• GDAS: More stable than DARTS, slightly slower
• OnceForAll: Best for multi-device deployment

TargetPlatform

Gets or sets the target hardware platform for optimization.

public HardwarePlatform TargetPlatform { get; set; }

Property Value

HardwarePlatform

Remarks

For Beginners: Choose where your model will run:

• CPU: Standard desktop/server deployment
• GPU: High-performance with CUDA/OpenCL
• Mobile: Smartphones and tablets
• EdgeTPU: Google Edge TPU accelerators

Verbose

Gets or sets whether to enable verbose logging during search.

public bool Verbose { get; set; }

Property Value

bool

WeightLearningRate

Gets or sets the learning rate for network weights.

public double WeightLearningRate { get; set; }

Property Value

double

Remarks

For Advanced Users: Controls how fast the network weights update. Typically higher than architecture learning rate.

Methods

Validate()

Validates the options and throws if any are invalid.

public void Validate()

Exceptions

ArgumentException

Thrown when options are invalid.

With(Action<NASOptions<T>>)

Creates a copy of these options with the specified modifications.

public NASOptions<T> With(Action<NASOptions<T>> configure)

Parameters

configure Action<NASOptions<T>>

Action to configure the copy.

Returns

NASOptions<T>

A new NASOptions instance with modifications applied.