test-time augmentation

**TTA** (Test-Time Augmentation) is an **inference technique that applies multiple augmentations to the test input, runs inference on each, and averages the predictions** — effectively ensembling over augmented views of the same input to improve prediction quality. **How Does TTA Work?** - **Augment**: Apply $K$ augmentations to the test input (e.g., flips, crops, rotations, scales). - **Infer**: Run the model on each of the $K$ augmented versions. - **Aggregate**: Average (or majority vote) the predictions: $hat{y} = frac{1}{K}sum_k f( ext{Aug}_k(x))$. - **Un-augment**: For spatial outputs (segmentation, detection), apply the inverse augmentation before averaging. **Why It Matters** - **Free Accuracy**: Typically 0.5-1.0% accuracy improvement with no model changes or retraining. - **Cost**: $K imes$ inference time — trades compute for accuracy. - **Standard Practice**: Routinely used in competitions, medical imaging, and safety-critical applications. **TTA** is **the inference ensemble** — running the model multiple times on augmented versions of the input for more reliable predictions.