Struct SamplerDataset

pub struct SamplerDataset<D, I> { /* private fields */ }

Sample items from a dataset.

This is a convenient way of modeling a dataset as a probability distribution of a fixed size. You have multiple options to instantiate the dataset sampler.

• With replacement (Default): This is the most efficient way of using the sampler because no state is required to keep indices that have been selected.

• Without replacement: This has a similar effect to using a shuffled dataset, but with more flexibility since you can set the dataset to an arbitrary size. Once every item has been used, a new cycle is created with a new random suffle.

Implementations

impl<D, I> SamplerDataset<D, I>

where D: Dataset<I>, I: Send + Sync,

pub fn new<O>(dataset: D, options: O) -> SamplerDataset<D, I>

where O: Into<SamplerDatasetOptions>,

Creates a new sampler dataset with replacement.

When the sample size is less than or equal to the source dataset size, data will be sampled without replacement from the source dataset in a uniformly shuffled order.

When the sample size is greater than the source dataset size, the entire source dataset will be sampled once for every multiple of the size ratios; with the remaining samples taken without replacement uniformly from the source. All samples will be returned uniformly shuffled.

Arguments

• dataset: the dataset to wrap.
• options: the options to configure the sampler dataset.

Examples

use burn_dataset::transform::{
  SamplerDataset,
  SamplerDatasetOptions,
};

// Examples below assuming `dataset.len()` = `10`.

// sample size: 5
// WithReplacement
// rng: StdRng::from_os_rng()
SamplerDataset::new(dataset, 5);

// sample size: 10 (source)
// WithReplacement
// rng: StdRng::from_os_rng()
SamplerDataset::new(dataset, SamplerDatasetOptions::default());

// sample size: 15
// WithoutReplacement
// rng: StdRng::seed_from_u64(42)
SamplerDataset::new(
  dataset,
  SamplerDatasetOptions::default()
    .with_size(1.5)
    .without_replacement()
    .with_rng(42),
);

pub fn with_replacement(dataset: D, size: usize) -> SamplerDataset<D, I>

Creates a new sampler dataset with replacement.

Arguments

• dataset: the dataset to wrap.
• size: the effective size of the sampled dataset.

pub fn without_replacement(dataset: D, size: usize) -> SamplerDataset<D, I>

Creates a new sampler dataset without replacement.

When the sample size is less than or equal to the source dataset size, data will be sampled without replacement from the source dataset in a uniformly shuffled order.

When the sample size is greater than the source dataset size, the entire source dataset will be sampled once for every multiple of the size ratios; with the remaining samples taken without replacement uniformly from the source. All samples will be returned uniformly shuffled.

Arguments

• dataset: the dataset to wrap.
• size: the effective size of the sampled dataset.

pub fn is_with_replacement(&self) -> bool

Determines if the sampler is using the “with replacement” strategy.

Returns

• true: If the sampler is configured to sample with replacement.
• false: If the sampler is configured to sample without replacement.

Trait Implementations

impl<D, I> Dataset<I> for SamplerDataset<D, I>

where D: Dataset<I>, I: Send + Sync,

fn get(&self, index: usize) -> Option<I>

Gets the item at the given index.

fn len(&self) -> usize

Gets the number of items in the dataset.

fn is_empty(&self) -> bool

Checks if the dataset is empty.

fn iter(&self) -> DatasetIterator<'_, I>

where Self: Sized,

Returns an iterator over the dataset.