BalancedEpisodicDataLoader<T, TInput, TOutput>
Models & Types · AiDotNet.Data.Loaders
Provides balanced episodic task sampling that ensures equal class representation across multiple tasks.
Inheritance: DataLoaderBase<T> → EpisodicDataLoaderBase<T, TInput, TOutput> → BalancedEpisodicDataLoader<T, TInput, TOutput>
For Beginners
Standard random sampling might pick some classes more often than others by chance. This can cause problems in meta-learning:
- The model might learn some classes better than others
- Training could be biased toward frequently-sampled classes
- Evaluation metrics might be skewed
The balanced loader solves this by:
- Tracking how many times each class has been selected
- Preferring classes that haven’t been used as much
- Ensuring fair representation across all classes over time
When to use this:
- Long meta-training runs where balanced class exposure matters
- When your dataset has many classes and you want uniform coverage
- When evaluating meta-learning algorithms fairly across all classes
Trade-off: Less random than uniform sampling, but more balanced. Good for training, but you might want standard EpisodicDataLoader for final evaluation to match real-world randomness.
How It Works
The BalancedEpisodicDataLoader extends the standard episodic loader by tracking class usage across multiple tasks and preferentially sampling under-represented classes. This ensures that over many episodes, all classes appear roughly the same number of times, preventing bias toward frequently-sampled classes.
Thread Safety: This class is not thread-safe due to internal state tracking. Create separate instances for concurrent task generation.
Performance: Slightly slower than standard EpisodicDataLoader due to usage tracking and weighted sampling, but still O(nWay × (kShot + queryShots)) for task creation.
Example
// Load a dataset with imbalanced class distribution
var features = new Matrix<double>(1000, 784);
var labels = new Vector<double>(1000); // 20 classes with varying frequencies
// Create balanced loader - all classes will be sampled equally over time
var loader = new BalancedEpisodicDataLoader<double, Matrix<double>, Vector<double>>(
datasetX: features,
datasetY: labels,
nWay: 5,
kShot: 3,
queryShots: 10,
seed: 42
);
// Over 1000 episodes, each class will appear in roughly the same number of tasks
for (int episode = 0; episode < 1000; episode++)
{
var task = loader.GetNextTask();
// Train on balanced distribution of classes
}
Constructors
| Constructor | Summary |
|---|---|
BalancedEpisodicDataLoader(Matrix<T> datasetX, Vector<T> datasetY, int nWay, int kShot, int queryShots, int? seed) |
Initializes a new instance of the BalancedEpisodicDataLoader for balanced N-way K-shot task sampling. |