KMaXDeepLabOptions

Options & Configuration · AiDotNet.ComputerVision.Segmentation.Panoptic

Configuration options for kMaX-DeepLab panoptic segmentation.

Inheritance: ModelOptions → NeuralNetworkOptions → ModelHyperparameterOptions → SegmentationModelOptions → KMaXDeepLabOptions

For Beginners

These options configure the KMaXDeepLab model. Default values follow the original paper settings.

Constructors

Constructor Summary
KMaXDeepLabOptions() Initializes a new instance with default values.
KMaXDeepLabOptions(KMaXDeepLabOptions other) Initializes a new instance by copying from another instance.

Properties

Property Summary
KMaXDeepLabModelSize ModelSize { get; set; } Gets or sets which published size variant of the model to build.
double ClassThresholdStuff { get; set; } Gets or sets the minimum class probability of a stuff cluster in the panoptic merge.
double ClassThresholdThing { get; set; } Gets or sets the minimum class probability of a thing cluster in the panoptic merge.
double InstanceDiscriminationLossWeight { get; set; } Gets or sets the weight of the pixel-wise instance discrimination loss (MaX-DeepLab’s auxiliary loss, kept by kMaX-DeepLab).
double InstanceDiscriminationSampleTemperature { get; set; } Gets or sets the temperature of the instance discrimination pixel sampling (pixels are drawn with probability proportional to their segment’s inverse area raised to this power).
double InstanceDiscriminationTemperature { get; set; } Gets or sets the instance discrimination similarity temperature.
double MaskIdLossWeight { get; set; } Gets or sets the weight of the mask-ID cross-entropy.
double OverlapThreshold { get; set; } Gets or sets the share of a cluster’s mask that must still be unclaimed for it to become a segment.
double PanopticQualityLossWeight { get; set; } Gets or sets the weight of the PQ-style loss.
double PixelConfidenceThreshold { get; set; } Gets or sets the mask probability a pixel needs to belong to a cluster’s mask.
double ReorderClassWeight { get; set; } Gets or sets the class-score exponent of the merge order.
double ReorderMaskWeight { get; set; } Gets or sets the mask-score exponent of the merge order.
double SemanticAsppDropRate { get; set; } Gets or sets the dropout after the ASPP projection.
double SemanticLossWeight { get; set; } Gets or sets the weight of the auxiliary semantic segmentation loss.
double SemanticSampleTemperature { get; set; } Gets or sets the temperature of the auxiliary semantic loss’s pixel sampling.
double VoidPriorProbability { get; set; } Gets or sets the void-class prior of the class head: a fixed bias log((C - 1) * p / (1 - p)) on the void logit (C the classes plus void) makes an uninformative class head predict void with probability p.
int FeedForwardDimension { get; set; } Gets or sets the hidden width of the cluster centers’ FFN.
int InstanceDiscriminationSampleCount { get; set; } Gets or sets how many pixels per image the instance discrimination loss samples.
int NumAttentionHeads { get; set; } Gets or sets the self-attention heads among the cluster centers.
int NumClusterCenters { get; set; } Gets or sets the number of cluster centers (mask slots).
int NumKMeansIterations { get; set; } Gets or sets the number of kMaX transformer decoder stages (each a k-means cross-attention, a self-attention among the cluster centers and an FFN).
int PixelEmbeddingDimension { get; set; } Gets or sets the width of the pixel embedding the cluster centers are compared with.
int PixelHeadDimension { get; set; } Gets or sets the width of the pixel-space head’s 1x1 convolution, between its 5x5 depthwise convolution and the pixel embedding.
int QueryDimension { get; set; } Gets or sets the width of the cluster centers (the decoder’s query and bottleneck width).
int SemanticAsppDimension { get; set; } Gets or sets the width of the auxiliary semantic decoder’s ASPP branches and projection.
int SemanticDecoderDimension { get; set; } Gets or sets the width of the semantic decoder’s fusion convolutions.
int SemanticHeadDimension { get; set; } Gets or sets the width of the semantic head’s 5x5 depthwise-separable convolution.
int SemanticSampleCount { get; set; } Gets or sets how many pixels per image the auxiliary semantic loss samples.
int[] SemanticAtrousRates { get; set; } Gets or sets the dilation rates of the ASPP’s 3x3 branches, on the stride-32 backbone features.
int[] SemanticLowLevelProjections { get; set; } Gets or sets the widths the semantic decoder projects its low-level backbone features to, coarsest first.

Methods

Method Summary
void Validate() Throws when a value on this instance cannot produce a working model.