wenmin-wu/cv-mixup-label-smoothing

Mixup + Label Smoothing

Overview

Mixup and label smoothing both soften training targets to reduce overfitting, but through different mechanisms. Mixup creates virtual training examples by linearly blending two images and their labels: img = λ·img1 + (1-λ)·img2, label = λ·y1 + (1-λ)·y2. Label smoothing shifts hard labels toward uniform: y_smooth = (1-ε)·y + ε/K. Combining both provides complementary regularization — mixup smooths the input space while label smoothing prevents overconfident predictions. Together they typically improve generalization by 0.5–1.5% accuracy.

Quick Start

import numpy as np
import torch

class MixupLabelSmoothingDataset(torch.utils.data.Dataset):
    def __init__(self, dataset, mixup_prob=0.5, mixup_alpha=0.2,
                 label_smoothing=0.05, num_classes=2):
        self.dataset = dataset
        self.mixup_prob = mixup_prob
        self.alpha = mixup_alpha
        self.eps = label_smoothing
        self.K = num_classes

    def smooth_label(self, label):
        """Apply label smoothing: shift toward uniform."""
        one_hot = torch.zeros(self.K)
        one_hot[label] = 1.0
        return one_hot * (1 - self.eps) + self.eps / self.K

    def __getitem__(self, idx):
        img1, label1 = self.dataset[idx]
        target1 = self.smooth_label(label1)

        if np.random.random() < self.mixup_prob:
            j = np.random.randint(len(self.dataset))
            img2, label2 = self.dataset[j]
            target2 = self.smooth_label(label2)

            lam = np.random.beta(self.alpha, self.alpha)
            img = lam * img1 + (1 - lam) * img2
            target = lam * target1 + (1 - lam) * target2
        else:
            img, target = img1, target1

        return img, target

Workflow

1. Apply label smoothing to convert hard labels to soft targets
2. With probability mixup_prob, sample a second example and blend
3. Blend both images and soft targets with the same lambda
4. Train with soft cross-entropy (KL divergence) loss, not hard CE

Key Decisions

• Mixup alpha: 0.2 is standard; higher (0.4) for stronger regularization
• Label smoothing ε: 0.05–0.1; higher for noisy labels or small datasets
• Loss function: Must use soft-label loss (KLDivLoss or manual BCE with soft targets)
• Order: Smooth first, then mixup — smoothing modifies the targets mixup will blend

References

• VinBigData 2-Class Classifier Complete Pipeline