FiganOptions
Options & Configuration · AiDotNet.Video.Options
Configuration for Figan.
Inheritance: ModelOptions → NeuralNetworkOptions → FiganOptions
For Beginners
These control how many levels of detail the motion estimate is refined through, how wide the small convolutional networks are, and how strongly the critic influences training.
How It Works
Defaults come from van Amersfoort et al., “Frame Interpolation with Multi-Scale Deep Loss Functions
and Generative Adversarial Networks” (arXiv:1711.06045): J = 3 scales, 32-channel six-layer
generator modules with 3x3 kernels, an eight-block discriminator starting at 32 filters, Adam at
1e-4, batch 128, 128x128 crops.
This replaces TDPNetOptions, whose NumDiffBlocks, NumHeads and
DifferenceThreshold configured a “difference-aware attention” mechanism with no basis in any
paper. They are gone rather than renamed.
Constructors
| Constructor | Summary |
|---|---|
FiganOptions() |
Properties
| Property | Summary |
|---|---|
OnnxModelOptions OnnxOptions { get; set; } |
Gets or sets ONNX runtime options. |
double DropoutRate { get; set; } |
Gets or sets the dropout rate. |
double LeakyReluSlope { get; set; } |
Gets or sets the negative slope of the discriminator’s leaky ReLU. |
double LearningRate { get; set; } |
Gets or sets the learning rate. |
int CropSize { get; set; } |
Gets or sets the training crop size. |
int DiscriminatorBlocks { get; set; } |
Gets or sets the number of discriminator blocks. |
int DiscriminatorFilters { get; set; } |
Gets or sets the discriminator’s initial filter count. |
int KernelSize { get; set; } |
Gets or sets the convolution kernel size. |
int LayersPerModule { get; set; } |
Gets or sets the number of convolutional layers per generator module. |
int NumFeatures { get; set; } |
Gets or sets the hidden channel width of every generator module. |
int NumScales { get; set; } |
Gets or sets the number of scales J in the coarse-to-fine flow recursion. |
string ModelPath { get; set; } |
Gets or sets an optional ONNX model path for inference-only use. |