Data Augmentation Strategies for Deep Learning

Data augmentation synthetically increases dataset size by creating transformed copies.
Each augmented example should preserve the semantic label (a rotated X-ray still shows the same condition).

Effect on training:
  1. More training examples → harder to memorise, forces generalisation
  2. Invariance: the model learns that predictions should be stable under transforms
  3. Regularisation: equivalent to penalising predictions that are sensitive to augmentations

Rule of thumb:
  If you can see the label is preserved after the transform, it's safe.
  If the transform could change the clinical finding, don't use it.

import torch
import torchvision.transforms as T
import torchvision.transforms.functional as TF

# ── Safe for chest X-ray ──
xray_train_transform = T.Compose([
    T.Resize(256),
    T.RandomResizedCrop(224, scale=(0.8, 1.0)),   # mild zoom
    T.RandomHorizontalFlip(p=0.5),                 # OK: L/R symmetric
    T.RandomRotation(degrees=10),                  # small rotation only
    T.ColorJitter(brightness=0.1, contrast=0.1),  # mild intensity
    T.ToTensor(),
    T.Normalize(mean=[0.5], std=[0.5]),
])

# ── More aggressive: histopathology (rotation-invariant) ──
pathology_train_transform = T.Compose([
    T.Resize(256),
    T.RandomCrop(224),
    T.RandomHorizontalFlip(),
    T.RandomVerticalFlip(),            # valid: tissue is orientation-invariant
    T.RandomRotation(degrees=90),      # full 90° steps
    T.ColorJitter(brightness=0.2, contrast=0.2, saturation=0.2, hue=0.05),
    T.ToTensor(),
    T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

# ── Natural image (ImageNet-style) ──
imagenet_train_transform = T.Compose([
    T.RandomResizedCrop(224),
    T.RandomHorizontalFlip(),
    T.ColorJitter(brightness=0.4, contrast=0.4, saturation=0.4, hue=0.1),
    T.ToTensor(),
    T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
    T.RandomErasing(p=0.5),
])

# Validation: NO random augmentation — only deterministic resizing
val_transform = T.Compose([
    T.Resize(256),
    T.CenterCrop(224),
    T.ToTensor(),
    T.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]),
])

import torch
import numpy as np

def augment_ecg(
    signal: torch.Tensor,   # (C, T) — channels × timesteps
    p: float = 0.5,
) -> torch.Tensor:
    """
    Augmentation for ECG signals.
    All transforms preserve clinical interpretation.
    """
    signal = signal.clone()
    
    # 1. Gaussian noise (simulate measurement noise)
    if torch.rand(1).item() < p:
        noise_level = 0.01 * signal.std()
        signal += torch.randn_like(signal) * noise_level
    
    # 2. Baseline wander (simulate electrode movement)
    if torch.rand(1).item() < p:
        t = torch.linspace(0, 1, signal.shape[-1])
        freq = torch.rand(1).item() * 0.5   # low frequency wander
        wander = 0.05 * torch.sin(2 * np.pi * freq * t)
        signal += wander.unsqueeze(0)
    
    # 3. Time scaling (small stretch/compression)
    if torch.rand(1).item() < p:
        scale = 0.9 + 0.2 * torch.rand(1).item()   # [0.9, 1.1]
        new_len = int(signal.shape[-1] * scale)
        signal = torch.nn.functional.interpolate(
            signal.unsqueeze(0), size=new_len, mode="linear", align_corners=False
        ).squeeze(0)
        # Crop or pad to original length
        if signal.shape[-1] >= 500:
            signal = signal[:, :500]
        else:
            signal = torch.nn.functional.pad(signal, (0, 500 - signal.shape[-1]))
    
    # 4. Random channel permutation (not for 12-lead ECG — leads have specific meaning)
    # Only safe for unlabelled channel orders (e.g., accelerometer axes)
    
    return signal

# Augmentation for tabular clinical data
def augment_tabular(
    X: torch.Tensor,      # (n_features,)
    y: float,
    noise_std: float = 0.02,
    dropout_prob: float = 0.05,
) -> tuple[torch.Tensor, float]:
    """Add mild noise and feature dropout for tabular augmentation."""
    X_aug = X.clone()
    
    # 1. Gaussian noise (proportional to feature std)
    X_aug += torch.randn_like(X) * noise_std
    
    # 2. Feature dropout (simulate missing values)
    mask = torch.rand_like(X) > dropout_prob
    X_aug = X_aug * mask.float()   # zero out dropped features
    
    return X_aug, y

# Note: tabular augmentation is delicate — noise must be clinically plausible
# Do NOT augment features with natural constraints (age ≥ 0, INR ≥ 0.5)

import torch
import torch.nn as nn

def mixup_data(
    X: torch.Tensor,
    y: torch.Tensor,
    alpha: float = 0.4,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, float]:
    """
    Mixup: interpolate between two training examples and their labels.
    X_mixed = λ·X_a + (1-λ)·X_b
    y_mixed = λ·y_a + (1-λ)·y_b  (soft labels)
    λ ~ Beta(alpha, alpha)
    """
    if alpha > 0:
        lam = torch.distributions.Beta(alpha, alpha).sample().item()
    else:
        lam = 1.0
    
    batch_size = X.shape[0]
    idx = torch.randperm(batch_size)   # random permutation for pairing
    
    X_mixed = lam * X + (1 - lam) * X[idx]
    y_a, y_b = y, y[idx]
    
    return X_mixed, y_a, y_b, lam

def mixup_criterion(
    criterion: nn.Module,
    pred: torch.Tensor,
    y_a: torch.Tensor,
    y_b: torch.Tensor,
    lam: float,
) -> torch.Tensor:
    return lam * criterion(pred, y_a) + (1 - lam) * criterion(pred, y_b)

# Training with Mixup
model = nn.Sequential(nn.Linear(20, 64), nn.ReLU(), nn.Linear(64, 1))
optimizer = torch.optim.AdamW(model.parameters(), lr=3e-4)
criterion = nn.BCEWithLogitsLoss()

X = torch.randn(64, 20)
y = torch.randint(0, 2, (64,)).float()

# Mixup training step
X_mix, ya, yb, lam = mixup_data(X, y, alpha=0.4)
optimizer.zero_grad()
pred = model(X_mix).squeeze()
loss = mixup_criterion(criterion, pred, ya, yb, lam)
loss.backward()
optimizer.step()
print(f"Mixup loss: {loss.item():.4f}, λ={lam:.3f}")

import torch
import numpy as np

def cutmix_data(
    X: torch.Tensor,   # (B, C, H, W) images
    y: torch.Tensor,   # (B,) labels
    alpha: float = 1.0,
) -> tuple[torch.Tensor, torch.Tensor, torch.Tensor, float]:
    """
    CutMix: paste a rectangular patch from one image into another.
    More informative than Mixup for images (preserves spatial structure).
    """
    lam = torch.distributions.Beta(alpha, alpha).sample().item()
    
    B, C, H, W = X.shape
    idx = torch.randperm(B)
    
    # Sample cut box
    cut_ratio = np.sqrt(1 - lam)
    cut_w = int(W * cut_ratio)
    cut_h = int(H * cut_ratio)
    cx = np.random.randint(W)
    cy = np.random.randint(H)
    
    x1 = max(cx - cut_w // 2, 0)
    x2 = min(cx + cut_w // 2, W)
    y1 = max(cy - cut_h // 2, 0)
    y2 = min(cy + cut_h // 2, H)
    
    X_cut = X.clone()
    X_cut[:, :, y1:y2, x1:x2] = X[idx, :, y1:y2, x1:x2]
    
    # Adjust λ to match actual patch area
    lam = 1 - (x2 - x1) * (y2 - y1) / (W * H)
    
    return X_cut, y, y[idx], lam

# CutMix loss is same as Mixup loss

import torch
import torch.nn as nn

def predict_with_tta(
    model: nn.Module,
    X: torch.Tensor,   # single image (C, H, W)
    n_augmentations: int = 8,
) -> torch.Tensor:
    """
    Apply multiple augmentations at test time and average predictions.
    Improves reliability — important for clinical AI.
    """
    import torchvision.transforms.functional as TF
    
    model.eval()
    preds = []
    
    augmentations = [
        lambda x: x,                                 # original
        lambda x: TF.hflip(x),                       # horizontal flip
        TF.adjust_brightness(X, 1.1),  # not callable like this — shown for structure
        TF.center_crop(X, 200),
    ]
    
    # Simpler: generate n random augmented versions
    augment = torch.nn.Sequential(
        *[torch.nn.Identity()] * n_augmentations   # placeholder
    )
    
    for _ in range(n_augmentations):
        # Apply random augmentation
        x_aug = X + 0.01 * torch.randn_like(X)   # mild noise as example
        x_aug = x_aug.unsqueeze(0)                 # add batch dim
        
        with torch.no_grad():
            pred = torch.sigmoid(model(x_aug).squeeze())
            preds.append(pred)
    
    # Average predictions
    return torch.stack(preds).mean(dim=0)