pytorch-multilabel-balanced-sampler 0.1.3

PyTorch samplers that output roughly balanced batches with support for multilabel datasets.

PyTorch Multilabel Balanced Samplers

This package provides samplers to fetch data samples from multilabel datasets in a balanced manner. Balanced sampling from multilabel datasets can be especially useful to handle class imbalance issues.

Samplers

• BaseMultilabelBalancedRandomSampler: This is the base class for all the provided samplers. It initializes the basic structure required for sampling, such as class indices.

• RandomClassSampler: This sampler randomly chooses a class and then picks a random example from that class.

• ClassCycleSampler: As the name suggests, it cycles through each class and fetches a random example from the current class.

• LeastSampledClassSampler: Chooses the class with the least number of samples fetched so far and retrieves a random example from that class.

Usage

Installation:

This package is installable via pip:

pip install pytorch-multilabel-balanced-sampler

Initialization:

For all samplers, the initialization arguments are:

• labels: A 2D tensor of shape (n_examples, n_classes) containing the one-hot encoded labels for the dataset.
• indices: A sequence of integers representing the indices of the dataset. Default is the range of the dataset size.

from pytorch_multilabel_balanced_sampler.samplers import RandomClassSampler, ClassCycleSampler, LeastSampledClassSampler

sampler1 = RandomClassSampler(labels=my_labels, indices=my_indices)
sampler2 = ClassCycleSampler(labels=my_labels)
sampler3 = LeastSampledClassSampler(labels=my_labels, indices=my_indices)

Fetching samples:

Iterate over the sampler object to fetch samples:

for sample in sampler1:
    print(sample)

Note:

All samplers are inherited from BaseMultilabelBalancedRandomSampler, which in turn inherits from PyTorch's Sampler class. This ensures compatibility with PyTorch's data loading utilities.

License

The MIT License (MIT). License

Feedback & Issues

For feedback, issues, or feature requests, please raise an issue on the GitHub repository.