Stylized Synthetic Augmentation Further Improves Corruption Robustness

This repository accompanies the paper “Stylized Synthetic Augmentation Further Improves Corruption Robustness”, available here:
https://arxiv.org/abs/2512.15675

We train image classification models using additional synthetic data and stylization. The repository provides flexible configuration of multiple data augmentation strategies and experiment setups.

---

🧭 Overview

• Training of corruption-robust image classifiers
• Support for synthetic data augmentation
• Support for stylization-based augmentation
• Flexible experiment configuration
• Works with multiple datasets and repository structures

---

📂 Repository Structure

• run_exp.py – main experiment launcher
• experiments/
  • train.py – training script
  • eval.py – evaluation script
  • configs/config_{ID}.py – experiment configuration files
• experiments/models/ – model definitions
• paths.json – configuration for dataset and checkpoint paths
• data/ – contains information for c and c-bar datasets

---

▶️ Running Experiments

run_exp.py runs one or multiple experiment IDs.

Each experiment setup must be defined in experiments/configs/config_{ID}.py

Internally, the launcher calls:

• experiments/train.py
• experiments/eval.py

---

🛠 Path Configuration

Use paths.json to specify directories for:

• datasets
• pretrained or trained models
• external storage layouts (e.g., Kaggle, custom structures)

Default expectation:

project_root/

├── repository/

├── data/

└── trained_models/

The data/ folder inside this repository only contains information for c and c-bar datasets; full datasets must be placed in the external data/ directory referenced in paths.json.

---

📚 Datasets

Automatically downloaded

• CIFAR-10
• CIFAR-100

Both are placed automatically into data/.

Must be added manually

• ImageNet
• TinyImageNet
• Corrupted variants:
  • -c
  • -c-bar

---

🧪 Synthetic Data Usage

To enable generated data (generate_ratio > 0.0), place .npz files in data/ with the naming pattern: {dataset}-add-1m-dm.npz

They can be obtained from:

• https://github.com/wzekai99/DM-Improves-AT
  or generated via:

• https://github.com/NVlabs/edm

---

🎨 Stylization Features

Stylization requires encoded image features from Painter-by-Numbers.

Required file in data/: style_feats_adain_1000.npy

For exact reproduction, download the 1000 features used here:

• https://zenodo.org/records/16279015

---

🧭 Model Architectures

Models are located in: experiments/models/

Key characteristics:

• include parameter factor for TinyImageNet (64×64)
• same base architecture as CIFAR (32×32)
• first convolution uses stride = factor = 2 for TinyImageNet
• all models inherit forward pass from ct_model.py, enabling:
  • normalization
  • noise injection
  • mixup
  • deeper-layer augmentations

---

✅ Capabilities Summary

• corruption-robust training
• integration of synthetic and stylized data
• configurable experiment setups
• flexible path handling
• unified augmentation control inside the forward pass

---