EMAVFIOptions
Options & Configuration · AiDotNet.Video.Options
Configuration options for EMAVFI, frame interpolation by
extracting motion and appearance via inter-frame attention (Zhang et al., CVPR 2023).
Inheritance: ModelOptions → NeuralNetworkOptions → EMAVFIOptions
For Beginners
EMA-VFI compares small windows of the two frames with attention: what each window attends to in the other frame describes its appearance, and where it attends describes its motion. Both feed a coarse-to-fine motion estimator and a refinement network that produce the in-between frame.
How It Works
The defaults are the release’s main model (MCG-NJU/EMA-VFI, config.py with F = 32, W = 7 and depths
2, 2, 2, 4, 4); Small returns EMA-VFI-small’s. Training follows train.py and Trainer.py.
Constructors
| Constructor | Summary |
|---|---|
EMAVFIOptions() |
Initializes a new instance with default values. |
EMAVFIOptions(EMAVFIOptions 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 AdamW’s peak learning rate (2e-4). |
double MergedWeight { get; set; } |
Gets or sets the weight of each stage’s warped blend in the loss (0.5). |
double MinLearningRate { get; set; } |
Gets or sets the rate the cosine schedule ends at (2e-5). |
double MlpRatio { get; set; } |
Gets or sets the attention blocks’ MLP expansion (4). |
double WeightDecay { get; set; } |
Gets or sets AdamW’s weight decay (1e-4). |
int FeatureChannels { get; set; } |
Gets or sets F, the first stage’s width; the five stages are F, 2F, 4F, 8F and 16F wide and the two attention stages have F/4 and F/2 heads (32). |
int LaplacianLevels { get; set; } |
Gets or sets the Laplacian loss’s pyramid levels (5). |
int TotalIterations { get; set; } |
Gets or sets the schedule’s length: 300 epochs of Vimeo90K’s 51,312 triplets at a total batch of 32 (four processes of 8). |
int WarmupSteps { get; set; } |
Gets or sets the linear warm-up from 0 (2000 steps). |
int WindowSize { get; set; } |
Gets or sets the attention window’s side W; odd blocks shift it by W / 2 (7). |
int[] Depths { get; set; } |
Gets or sets the five stages’ depths: convolutions in the first three, inter-frame attention blocks in the last two (2, 2, 2, 4, 4). |
string ModelPath { get; set; } |
Gets or sets the path to the ONNX model file. |
Methods
| Method | Summary |
|---|---|
EMAVFIOptions Small() |
EMA-VFI-small (config.py, F = 16, depths 2, 2, 2, 2, 2). |