mixup is an R package for data-augmentation inspired by mixup: Beyond Empirical Risk Minimization
If you like mixup, give it a star, or fork it and contribute!
Create additional training data for toy dataset:
library(mixup)
# Use builtin mtcars dataset with mtcars$am (automatic/manual) as binary target
data(mtcars)
str(mtcars)
summary(mtcars[, -9])
summary(mtcars$am)
# Strictly speaking this is 'input mixup' (see Details section below)
set.seed(42)
mtcars.mix <- mixup(mtcars[, -9], mtcars$am)
summary(mtcars.mix$x)
summary(mtcars.mix$y)
# Further info
?mixupRequires R version 3.2.0 and higher.
install.packages('devtools') # Install devtools package if necessary
library(devtools)
devtools::install_github('makeyourownmaker/mixup')The mixup function enlarges training sets using linear interpolations of features and associated labels as described in https://arxiv.org/abs/1710.09412.
Virtual feature-target pairs are produced from randomly drawn
feature-target pairs in the training data.
The method is straight-forward and data-agnostic. It should
result in a reduction of generalisation error.
Mixup constructs additional training examples:
x' = λ * x_i + (1 - λ) * x_j, where x_i, x_j are raw input vectors
y' = λ * y_i + (1 - λ) * y_j, where y_i, y_j are one-hot label encodings
(x_i, y_i) and (x_j ,y_j) are two examples drawn at random from the training data, and λ ∈ [0, 1] with λ ∼ Beta(α, α) for α ∈ (0, ∞). The mixup hyper-parameter α controls the strength of interpolation between feature-target pairs.
| Parameter | Description | Notes |
|---|---|---|
| x1 | Original features | Required parameter |
| y1 | Original labels | Required parameter |
| alpha | Hyperparameter specifying strength of interpolation | Defaults to 1 |
| concat | Concatenate mixup data with original data | Defaults to FALSE |
| batch_size | How many mixup values to produce | Defaults to number of examples |
The x1 and y1 parameters must be numeric and must have equal numbers of examples. Non-finite values are not permitted. Factors should be one-hot encoded.
For now, only binary classification is supported. Meaning y1 must contain only numeric 0 and 1 values.
Alpha values must be greater than or equal to zero. Alpha equal to zero specifies no interpolation.
The mixup function returns a two-element list containing interpolated x and y values. Optionally, the original values can be concatenated with the new values.
It is worthwhile distinguishing between mixup usage with deep learning and other learning methods. Mixup with deep learning can improve generalisation when a new mixed dataset is generated every epoch or even better for every minibatch. This level of granularity may not be possible with other learning methods. For example, simple linear modeling may not benefit much from training on a single (potentially greatly expanded) pre-mixed dataset. This single pre-mixed dataset approach is sometimes referred to as 'input mixup'.
In certain situations, under-fitting can occur when conflicts between synthetic labels of the mixed-up examples and labels of the original training data are present. Some learning methods may be more prone to this under-fitting than others.
Data augmentation is occasionally referred to as a regularisation technique. Regularisation decreases a model's variance by adding prior knowledge (sometimes using shrinkage). Increasing training data (using augmentation) also decreases a model's variance. Data augmentation is also a form of adding prior knowledge to a model.
If you use mixup in a scientific publication, then consider citing the original paper:
mixup: Beyond Empirical Risk Minimization
By Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, David Lopez-Paz
https://arxiv.org/abs/1710.09412
I have no affiliation with MIT, FAIR or any of the authors.
- Improve docs
- Add more detailed examples
- Different data types e.g. tabular, image etc
- Different parameters
- Different learning methods
- Add more detailed examples
- Add my time series mixup variant
- Applies mixup technique to two time series separated by 'time_diff' period
- Implemented and tested in this encoder decoder Jupyter notebook
- Lint package with goodpractice
- Add tests
- Add support for one-hot encoded labels
- Add label preserving option
- Add support for mixing within the same class
- Usually doesn't perform as well as mixing within all classes
- May still have some utility e.g. unbalanced data sets
- Generalise to regression problems
Other implementations:
- pytorch from hongyi-zhang
- pytorch from facebookresearch
- keras from yu4u
- mxnet from unsky
- Python package for data augmentation inspired by Mixup: Beyond Empirical Risk Minimization
Discussion:
Closely related research:
- Manifold Mixup: Better Representations by Interpolating Hidden States
- MixUp as Locally Linear Out-Of-Manifold Regularization
Loosely related research:
Pull requests are welcome. For major changes, please open an issue first to discuss what you would like to change.