DeepEMDOptions<T, TInput, TOutput>
Options & Configuration · AiDotNet.MetaLearning.Options
Configuration options for DeepEMD (Zhang et al., CVPR 2020) few-shot learning.
Inheritance: ModelOptions → DeepEMDOptions<T, TInput, TOutput>
Implements: IMetaLearnerOptions<T>
For Beginners
DeepEMD uses a clever way to compare examples:
The Problem with simple metrics: ProtoNets averages all features into one vector. This loses local details like “the left part of image A looks like the right part of image B.”
DeepEMD’s solution: Earth Mover’s Distance Think of features as piles of dirt. EMD measures the minimum “work” needed to reshape one pile into another. This captures structural alignment:
- Each local feature (patch) from one image can match any patch from another
- The cost of matching = distance between the patches
- EMD finds the optimal matching that minimizes total distance
Three modes:
- FCN: Feature maps from the last conv layer (spatial patches)
- Grid: Divide feature maps into a regular grid
- Sampling: Sample representative local features
How It Works
DeepEMD uses the Earth Mover’s Distance (optimal transport) as a metric for comparing image representations in few-shot learning. Unlike ProtoNets which compares single vectors, DeepEMD compares sets of local features using optimal transport, capturing fine-grained structural similarity.
Reference: Zhang, C., Cai, Y., Lin, G., & Shen, C. (2020). DeepEMD: Few-Shot Image Classification with Differentiable Earth Mover’s Distance and Structured Classifiers. CVPR 2020.
Constructors
| Constructor | Summary |
|---|---|
DeepEMDOptions(IFullModel<T, TInput, TOutput> metaModel) |
Initializes a new instance of DeepEMDOptions. |
Properties
| Property | Summary |
|---|---|
IEpisodicDataLoader<T, TInput, TOutput> DataLoader { get; set; } |
Gets or sets the episodic data loader. |
IFullModel<T, TInput, TOutput> MetaModel { get; set; } |
Gets or sets the feature extractor model. |
IGradientBasedOptimizer<T, TInput, TOutput> InnerOptimizer { get; set; } |
Must stay null: this algorithm takes no plain gradient inner step (it adapts without one, or by its own update rule), so a supplied optimizer is rejected when the learner is built. |
IGradientBasedOptimizer<T, TInput, TOutput> MetaOptimizer { get; set; } |
Optimizer for the meta-update. |
ILossFunction<T> LossFunction { get; set; } |
Loss function for meta-training; null uses the default multi-class classification loss. |
bool EnableCheckpointing { get; set; } |
|
bool UseFirstOrder { get; set; } |
|
double InnerLearningRate { get; set; } |
Not used by DeepEMD: its adaptation takes no gradient steps (see AdaptationSteps). |
double OuterLearningRate { get; set; } |
|
double SinkhornRegularization { get; set; } |
Gets or sets the regularization parameter for Sinkhorn algorithm. |
double Temperature { get; set; } |
Gets or sets the temperature for the EMD-based classification logits. |
double? GradientClipThreshold { get; set; } |
|
int AdaptationSteps { get; set; } |
Number of gradient steps taken on each task’s support set. |
int CheckpointFrequency { get; set; } |
|
int EvaluationFrequency { get; set; } |
|
int EvaluationTasks { get; set; } |
|
int MetaBatchSize { get; set; } |
|
int NumMetaIterations { get; set; } |
|
int NumNodes { get; set; } |
Gets or sets the number of local features (nodes) per example for EMD computation. |
int SinkhornIterations { get; set; } |
Gets or sets the number of Sinkhorn iterations for approximate EMD. |
int? RandomSeed { get; set; } |
Gets or sets the random seed. |
string EMDMode { get; set; } |
Gets or sets the EMD mode: “fcn”, “grid”, or “sampling”. |
Methods
| Method | Summary |
|---|---|
IMetaLearnerOptions<T> Clone() |
|
bool IsValid() |