ShiftNetOptions
Options & Configuration · AiDotNet.Video.Options
Configuration options for Shift-Net, the grouped spatial-temporal shift video restorer (Li et al., CVPR 2023), in its denoising configuration.
Inheritance: ModelOptions → NeuralNetworkOptions → ShiftNetOptions
For Beginners
Shift-Net cleans a video by sliding slices of each frame’s features in time (to the neighbouring frames) and in space (by a few pixels in many directions), letting the following convolutions pick whichever slide lines up. Sliding is free, so it gets the benefit of alignment without computing motion.
How It Works
Defaults follow the authors’ release (github.com/dasongli1/Shift-Net, basicsr/models/archs/gshift_denoise1.py and options/davis_denoise.yml): 24-channel full-resolution features refined by five U-Nets before and after the shift encoder, an 80-channel shift encoder at half resolution, spatial shifts of 4 and 8 pixels, and Adam at 4e-4 (betas 0.9, 0.99) cosine-annealed to 1e-7 over 450k iterations on BasicSR’s PSNR loss with weight 0.5.
Constructors
| Constructor | Summary |
|---|---|
ShiftNetOptions() |
Initializes a new instance with default values. |
ShiftNetOptions(ShiftNetOptions other) |
Initializes a new instance by copying from another instance. |
Properties
| Property | Summary |
|---|---|
OnnxModelOptions OnnxOptions { get; set; } |
Gets or sets the ONNX runtime options. |
double LearningRate { get; set; } |
Gets or sets Adam’s initial learning rate (4e-4). |
double MinLearningRate { get; set; } |
Gets or sets the learning rate the cosine schedule anneals to (1e-7). |
double PsnrLossWeight { get; set; } |
Gets or sets the weight of BasicSR’s PSNR loss (0.5). |
double? NoiseLevel { get; set; } |
Gets or sets the noise standard deviation (in the frames’ own units) that inference’s noise map carries; null estimates it from the frames. |
int CosineIterations { get; set; } |
Gets or sets the iterations of the cosine schedule (450000). |
int NumFeatures { get; set; } |
Gets or sets the shift encoder’s channels (80); a multiple of 16, since half of them are split into the eight shift groups of the spatial shift. |
int NumUNets { get; set; } |
Gets or sets the refinement U-Nets before the shift encoder, and again after it (5). |
int RefinementFeatures { get; set; } |
Gets or sets the full-resolution channels of the feature extractor and refinement U-Nets (24). |
int ShiftStride { get; set; } |
Gets or sets the spatial shift s: the shift groups move by s and 2s pixels (4). |
int UNetGrowth { get; set; } |
Gets or sets the channels each refinement U-Net level adds when it halves the resolution (12). |
string ModelPath { get; set; } |
Gets or sets the path to the ONNX model file. |