Skip to content

Data Augmentation

Artificially increase training data by applying random transformations. Reduces overfitting, improves generalization. Virtually free performance improvement for vision and NLP tasks.

Image Augmentation

Basic Transforms

from torchvision import transforms

train_transform = transforms.Compose([
    transforms.RandomResizedCrop(224),
    transforms.RandomHorizontalFlip(p=0.5),
    transforms.RandomRotation(15),
    transforms.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.1),
    transforms.RandomAffine(degrees=0, translate=(0.1, 0.1)),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225])
])

# Validation: NO augmentation (except resize + normalize)
val_transform = transforms.Compose([
    transforms.Resize(256),
    transforms.CenterCrop(224),
    transforms.ToTensor(),
    transforms.Normalize(mean=[0.485, 0.456, 0.406],
                         std=[0.229, 0.224, 0.225])
])

Advanced Techniques

  • Cutout/Random Erasing: randomly mask rectangular regions
  • Mixup: blend two images and their labels: x = lambdax1 + (1-lambda)x2
  • CutMix: paste patch from one image onto another, mix labels by area ratio
  • AutoAugment: learned augmentation policies

Text Augmentation

  • Synonym replacement: replace words with synonyms
  • Back-translation: translate to another language and back
  • Random insertion/deletion/swap: perturb word order
  • Contextual augmentation: use language model to replace words

Tabular Data Augmentation

  • SMOTE: generate synthetic minority samples for imbalanced classes
  • Noise injection: add Gaussian noise to features
  • Mixup: blend feature vectors and labels
from imblearn.over_sampling import SMOTE
sm = SMOTE(random_state=42)
X_resampled, y_resampled = sm.fit_resample(X_train, y_train)

Principles

  • Apply augmentation only to training data, never validation/test
  • Augmentations should be label-preserving (flipping a "6" may create a "9")
  • Domain-appropriate: don't flip text horizontally, don't rotate medical images if orientation matters
  • More aggressive augmentation for smaller datasets
  • Combine multiple augmentations for maximum effect

Gotchas

  • Over-aggressive augmentation can harm performance (model learns to handle noise instead of signal)
  • Vertical flips rarely appropriate (most natural images have gravity)
  • Color augmentation can destroy important color information (medical imaging)
  • SMOTE generates synthetic points in feature space - may create unrealistic samples
  • Augmentation adds training time - balance benefit vs cost

See Also