PyT-ImbalancedDatasetSampler

Introduction

An imbalanced dataset sampler for PyTorch. This an alternative for ufoym's ImbalancedDatasetSampler.

It does not randomly draw elements by a multinomial distribution, which may unwantetly draw some elements multiple times while skipping over other elements, but instead wraps the dataset around when oversampling.

Additionally, a finer control over the over- and undersampling procedure is allowed, by manually setting the number of data-points per-class or specifing per-class over- and undersampling factors. We do not require external packages other than pyTorch.

For a nice explanation with great visuals of over- and undersampling visit ufomy's github project.

Description

Create an instance of the ImbalancedDatasetSampler and pass it to the pyTorch DataLoader

from imbalanced_sampler import ImbalancedDatasetSampler

sampler = ImbalancedDatasetSampler(dataset=dataset, sampling_factor=s_f, shuffle=True, ....)

train_loader = torch.utils.data.DataLoader(
    train_dataset,
    sampler=sampler,
    ... )

The ImbalancedDatasetSampler takes multiple arguments allowing for different behaviour (also take a look at the docstring):

• dataset:torch.Dataset - the dataset
• sampling_factor:[str,float,int] - Sets the degree of over- or under-sampling for the classes. If set to an int, each class will be scaled to that many data points. If set to a float between 0 and 1, the class size will be interpolated between the smallest and largest class. A float between -1.0 and 0.0 will undersample the largest class to class_size(largest)*-sampling_factor. A value smaller than -1.0 will oversample the smallest class by a factor of -sampling_factor. All classes will always have the same size.
• num_classes: Number of classes. If set to None will be automatically determined.
• shuffle: If set to true, will shuffle the dataset at each epoch. Otherwise will always return the same order. However, if a class is undersampled, it is still non-deterministc with each iterator instanciation.
• labels: You can pass the labels directly here.
• callback_get_label: A Callable that will be called to generate the labels. Will default to lambda x: __getitem__(idx)[1] on the dataset if not provided.
• callback_type: If set to "single" the function will be called with the index of the dataset callback_get_label(idx). If set to multi, no arguments will be passed and it is expected to return a list of labels with length len(dataset).

How to install

You can directly install the imbalanced_sampler package from the provided setup.py file

pip install .

Alternativly install from the provided tar.gz file

pip install dist/imbalanced_sampler-0.1.tar.gz

Finally, you could just copy the ImbalancedDatasetSampler.py to your projects location.

Future

I want to add support for manually setting the classes to different sizes.