Mixup: Blending Two Training Examples
Beginner
The problem, from first principles
Mixup (Zhang et al., 2018) trains on convex combinations of pairs of examples and their labels. It is a simple augmentation that smooths the decision boundary between classes.
From theory to code
Implement mixup_inputs(x, perm, lam), which blends each example with the example at its permuted index using weight lam.
Constraints
lamis in[0, 1].permis a permutation of the batch indices.
Hints
Hint 1
Index the batch with perm to get the partners, then take the weighted sum.
Keywords
research-papersaugmentationmixup
Python
Run your code to see results
No results yet
Run your code to check it against the examples, or submit to run every test.