randaugment

**RandAugment** is a **dramatically simplified data augmentation strategy that achieves state-of-the-art results by randomly selecting N transformations from a pool and applying each with a fixed magnitude M** — replacing AutoAugment's expensive 5,000-GPU-hour search with just two tunable hyperparameters that can be optimized with a simple grid search, making automated augmentation accessible to any practitioner without massive computational resources. **What Is RandAugment?** - **Definition**: An augmentation strategy that randomly selects N transformations from a fixed pool of 14 operations and applies each with the same magnitude M — requiring no dataset-specific search, no reinforcement learning, and no proxy task, while matching or exceeding the performance of learned augmentation policies. - **The Insight**: AutoAugment's expensive search finds optimal per-operation magnitudes (rotate at magnitude 7, shear at magnitude 5). RandAugment shows that using the same magnitude M for all operations works nearly as well — reducing the search space from thousands of parameters to just 2. - **Philosophy**: "Simple baselines are often underrated" — sometimes the optimal solution is not the most complex one. **How RandAugment Works** | Step | Process | Example | |------|---------|---------| | 1. Define pool of K transforms | 14 standard transforms | Rotate, Shear, Translate, Brightness, etc. | | 2. For each training image | Randomly select N transforms from the pool | N=2: select Rotate and Contrast | | 3. Apply each with magnitude M | Same M for all selected transforms | M=9: moderate-to-strong transforms | | 4. Feed augmented image to model | Standard training pipeline | Model trains on varied augmentations | **The Transform Pool (14 Operations)** | Operation | Description | Magnitude Example (M=9) | |-----------|-------------|------------------------| | Identity | No change | — | | Rotate | Rotate by angle | ±13.5° | | ShearX/Y | Shear horizontally/vertically | 0.3 shear factor | | TranslateX/Y | Shift pixels | 14 pixels | | AutoContrast | Maximize contrast | — (binary) | | Equalize | Histogram equalization | — (binary) | | Solarize | Invert pixels above threshold | Threshold 178 | | Posterize | Reduce bits per color channel | 5 bits | | Brightness | Adjust brightness | Factor 1.9 | | Contrast | Adjust contrast | Factor 1.9 | | Color | Adjust saturation | Factor 1.9 | | Sharpness | Adjust sharpness | Factor 1.9 | **Hyperparameter Tuning** | Hyperparameter | Typical Values | Effect | |---------------|---------------|--------| | **N** (number of ops) | 1-3 | More ops = stronger augmentation | | **M** (magnitude) | 5-15 (out of 30) | Higher = more distortion | Typical grid: N ∈ {1, 2, 3} × M ∈ {5, 7, 9, 11, 13, 15} = 18 experiments. **RandAugment vs Alternatives** | Method | Search Cost | Hyperparameters | Key Advantage | |--------|-----------|-----------------|-------------| | **Hand-designed** | Human time | Many per-transform params | Domain knowledge | | **AutoAugment** | 5,000 GPU hours | Policy per dataset | Optimal (but expensive) | | **RandAugment** | ~18 grid search runs | Just N and M | Simple, effective, practical | | **TrivialAugment** | 0 | Zero hyperparameters | Even simpler (1 random op) | **Results** | Dataset | Model | Without Aug | RandAugment | AutoAugment | |---------|-------|------------|-------------|-------------| | CIFAR-10 | WRN-28-10 | 3.87% | 2.70% | 2.68% | | ImageNet | ResNet-50 | 23.7% | 22.4% | 22.4% | | SVHN | WRN-28-2 | 1.88% | 1.36% | 1.30% | RandAugment matches AutoAugment within ~0.1% on all benchmarks — at a fraction of the computational cost. **RandAugment is the practical standard for automated data augmentation** — proving that randomly selecting N operations at a fixed magnitude M rivals expensive learned policies, making strong augmentation accessible to any practitioner through a simple 2-parameter grid search instead of thousands of GPU hours.

Go deeper with CFSGPT

Get AI-powered deep-dives, save terms, and run advanced simulations — free account.

Create Free Account