Imbalanced learning is a machine learning paradigm whereby a classifier has to learn from a dataset that has a skewed class distribution. An imbalanced dataset may have a detrimental impact on the classifier's performance.
Rebalancing a dataset is one way to deal with class imbalance. This can be done by:
- under-sampling common classes.
- over-sampling rare classes.
- doing a mix of both.
PyTorch provides some utilities for rebalancing a dataset, but they are limited to batch datasets of known length (i.e., they require a dataset to have a __len__
method). Community contributions such as ufoym/imbalanced-dataset-sampler are cute, but they also only work with batch datasets (also called map-style datasets in PyTorch jargon). There's also a GitHub issue opened on the PyTorch GitHub repository, but it doesn't seem very active.
This repository implements data resamplers that wrap an IterableDataset
. Each data resampler also inherits from IterableDataset
. The latter was added to PyTorch in this pull request. In particular, the provided methods do not require you to have to know the size of your dataset in advance. Each methods works for both binary and multi-class classification.
☝️ If you're looking to sample your data completely at random, without taking into consideration the class distribution, then we recommend that you do it yourself in your IterableDataset
implementation. Indeed, you just have to generate a random number between 0 and 1 and keep a sample if the sampled number is under a given threshold. This library is meant to be used when you want to use resampling to balance your class distribution.
$ pip install pytorch_resample
As a running example, we'll define an IterableDataset
that iterates over the output of scikit-learn's make_classification
function.
>>> from sklearn import datasets
>>> import torch
>>> class MakeClassificationStream(torch.utils.data.IterableDataset):
...
... def __init__(self, *args, **kwargs):
... self.X, self.y = datasets.make_classification(*args, **kwargs)
...
... def __iter__(self):
... yield from iter(zip(self.X, self.y))
The above dataset can be provided to a DataLoader
in order to iterate over Tensor
batches. For the sake of example, we'll generate 10.000 samples, with 50% of 0s, 40% of 1s, and 10% of 2s. We can use a collections.Counter
to measure the effective class distribution.
>>> import collections
>>> dataset = MakeClassificationStream(
... n_samples=10_000,
... n_classes=3,
... n_informative=6,
... weights=[.5, .4, .1],
... random_state=42
... )
>>> y_dist = collections.Counter()
>>> batches = torch.utils.data.DataLoader(dataset, batch_size=16)
>>> for xb, yb in batches:
... y_dist.update(yb.numpy())
>>> for label in sorted(y_dist):
... frequency = y_dist[label] / sum(y_dist.values())
... print(f'• {label}: {frequency:.2%} ({y_dist[label]})')
• 0: 49.95% (4995)
• 1: 39.88% (3988)
• 2: 10.17% (1017)
The data stream can be under-sampled with the pytorch_resample.UnderSampler
class. The latter is a wrapper that has to be provided with an IterableDataset
and a desired class distribution. It inherits from IterableDataset
, and may thus be used instead of the wrapped dataset. As an example, let's make it so that the classes are equally represented.
>>> import pytorch_resample
>>> import torch
>>> sample = pytorch_resample.UnderSampler(
... dataset=dataset,
... desired_dist={0: .33, 1: .33, 2: .33},
... seed=42
... )
>>> isinstance(sample, torch.utils.data.IterableDataset)
True
>>> y_dist = collections.Counter()
>>> batches = torch.utils.data.DataLoader(sample, batch_size=16)
>>> for xb, yb in batches:
... y_dist.update(yb.numpy())
>>> for label in sorted(y_dist):
... frequency = y_dist[label] / sum(y_dist.values())
... print(f'• {label}: {frequency:.2%} ({y_dist[label]})')
• 0: 33.30% (1007)
• 1: 33.10% (1001)
• 2: 33.60% (1016)
As shown, the observed class distribution is close to the specified distribution. Indeed, there are less 0s and 1s than above. Note that the values of the desired_dist
parameter are not required to sum up to 1. Indeed, the distribution is normalized automatically.
You may use pytorch_resample.OverSampler
to instead oversample the data. It has the same signature as pytorch_resample.UnderSampler
, and can thus be used in the exact same manner.
>>> sample = pytorch_resample.OverSampler(
... dataset=dataset,
... desired_dist={0: .33, 1: .33, 2: .33},
... seed=42
... )
>>> isinstance(sample, torch.utils.data.IterableDataset)
True
>>> y_dist = collections.Counter()
>>> batches = torch.utils.data.DataLoader(sample, batch_size=16)
>>> for xb, yb in batches:
... y_dist.update(yb.numpy())
>>> for label in sorted(y_dist):
... frequency = y_dist[label] / sum(y_dist.values())
... print(f'• {label}: {frequency:.2%} ({y_dist[label]})')
• 0: 33.21% (4995)
• 1: 33.01% (4965)
• 2: 33.78% (5080)
In this case, the 1s and 2s have been oversampled.
The pytorch_resample.HybridSampler
class can be used to compromise between under-sampling and over-sampling. It accepts an extra parameter called sampling_rate
, which determines the percentage of data to use. This allows to control how much data is used for training, whilst ensuring that the class distribution follows the desired distribution.
>>> sample = pytorch_resample.HybridSampler(
... dataset=dataset,
... desired_dist={0: .33, 1: .33, 2: .33},
... sampling_rate=.5, # use 50% of the dataset
... seed=42
... )
>>> isinstance(sample, torch.utils.data.IterableDataset)
True
>>> y_dist = collections.Counter()
>>> batches = torch.utils.data.DataLoader(sample, batch_size=16)
>>> for xb, yb in batches:
... y_dist.update(yb.numpy())
>>> for label in sorted(y_dist):
... frequency = y_dist[label] / sum(y_dist.values())
... print(f'• {label}: {frequency:.2%} ({y_dist[label]})')
• 0: 33.01% (1672)
• 1: 32.91% (1667)
• 2: 34.08% (1726)
As can be observed, the amount of streamed samples is close to 5000, which is half the size of the dataset.
It's possible to determine the exact number of samples each resampler will stream back in advance, provided the class distribution of the data is known.
>>> n = 10_000
>>> desired = {'cat': 1 / 3, 'mouse': 1 / 3, 'dog': 1 / 3}
>>> actual = {'cat': .5, 'mouse': .4, 'dog': .1}
>>> pytorch_resample.UnderSampler.expected_size(n, desired, actual)
3000
>>> pytorch_resample.OverSampler.expected_size(n, desired, actual)
15000
>>> pytorch_resample.HybridSampler.expected_size(n, .5)
5000
By design OverSampler
and HybridSampler
yield repeated samples one after the other. This might not be ideal, as it is usually desirable to diversify the samples within each batch. We therefore recommend that you use a shuffling buffer, such as the ShuffleDataset
class proposed here.
I've written a simple benchmark to verify that resampling brings a performance boost and can reduce computation time. It works, but take it with a grain of salt, as it is far from being exhaustive. Feel free to contribute more sophisticated benchmarks.
As far as I know, the methods that are implemented in this package do not exist in the litterature per se. I first stumbled on the under-sampling method by myself, which turned out to be equivalent to rejection sampling. I then worked out the necessary formulas for over-sampling and the hybrid method. Both of the latter are based on the idea of sampling from a Poisson distribution, which I took from the Online Bagging and Boosting paper by Nikunj Oza and Stuart Russell. The innovation lies in the determination of the rate that satisfies the desired class distribution.
$ git clone https://github.com/MaxHalford/pytorch-resample
$ cd pytorch-resample
$ python -m venv .env
$ source .env/bin/activate
$ pip install poetry
$ poetry install
$ poetry shell
$ pytest
The MIT License (MIT). Please see the license file for more information.