wenmin-wu/cv-test-time-augmentation

Test-Time Augmentation (TTA)

Overview

Generate multiple augmented copies of each input image at inference (flips, 90-degree rotations, brightness/contrast jitter), run the model on all copies, then average the predictions. Reduces prediction variance and improves robustness without retraining.

Quick Start

import numpy as np

def tta_predict(model, image, n_augments=8):
    preds = []
    for _ in range(n_augments):
        aug = apply_random_augment(image)  # flip, rotate, color jitter
        preds.append(model.predict(aug[np.newaxis]))
    return np.mean(preds, axis=0)

def apply_random_augment(img):
    if np.random.rand() > 0.5:
        img = np.fliplr(img)
    if np.random.rand() > 0.5:
        img = np.flipud(img)
    k = np.random.randint(0, 4)
    img = np.rot90(img, k)
    return img

Workflow

1. Define augmentation set (flips, rotations, color transforms)
2. For each test image, generate N augmented copies
3. Run model inference on all copies
4. Average (classification) or geometrically reverse + average (segmentation) predictions
5. Use averaged prediction as final output

Key Decisions

• N augments: 4-8 is typical; diminishing returns beyond 16
• Augment types: Match training augmentations for best effect
• Segmentation TTA: Must reverse spatial transforms before averaging masks
• Cost: Inference time scales linearly with N; batch augmented copies for GPU efficiency

References

• HPA Single Cell Classification competition (Kaggle)
• Source: hpa-cellwise-classification-inference