MCL-NF

Official PyTorch implementation of Meta-Continual Learning of Neural Fields (arXiv 2025).

This repository provides a clean, minimal implementation to train and evaluate MCL-NF across three different data modalities:

• Images: CelebA, FFHQ, Imagenette
• Video: VoxCeleb
• Audio: LibriSpeech (1-second and 3-second variants)

---

🛠 Setup & Dependencies

We recommend using Conda to manage your environment.

1. Create and activate a new environment:

conda create -n mcl_nf python=3.10 -y
conda activate mcl_nf

2. Install PyTorch: (Note: Adjust the CUDA version to match your system. Here we use CUDA 11.8 as an example.)

pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu118

3. Install remaining dependencies:

pip install numpy einops Pillow PyYAML tensorboard

(Depending on your environment, you may also need pytorchvideo for certain video operations.)

---

📂 Data Configuration

To maintain privacy and prevent local paths from leaking, dataset paths are dynamically configured. By default, the code looks for datasets in the ./data directory relative to the project root.

You can override the default data root directory using one of the following methods:

Method 1: Environment Variable (Recommended)

export MCL_NF_DATA_ROOT=/absolute/path/to/your/datasets

Method 2: Command-Line Argument Pass the --data_root argument when running the scripts (see Usage below).

Expected Dataset Structure

If using the default ./data folder, your directory structure should look like this:

data/
  ├── celeba/          # For CelebA
  ├── ffhq/            # For FFHQ
  ├── imagenette/      # For Imagenette
  ├── voxceleb/        # For VoxCeleb
  └── librispeech/     # For LibriSpeech

---

🚀 Usage

You can launch experiments by specifying a dataset and optionally passing a YAML configuration file.

Training

Run the main.py script to start training.

Example: Training on VoxCeleb

python main.py --dataset voxceleb --data_root /path/to/your/data

Example: Training on LibriSpeech (using a config file) If you have a configuration file located at configs/maml_librispeech3.yaml, you can pass it to automatically load hyperparameters:

python main.py --configs configs/maml_librispeech3.yaml --data_root /path/to/your/data

Customizing Hyperparameters via CLI: Command-line arguments will override configuration files.

python main.py \
    --dataset celeba \
    --configs configs/maml_celeba.yaml \
    --inner_step 4 \
    --inner_iter 1 \
    --lr 1e-5 \
    --batch_size 32

Evaluation

To evaluate a saved model checkpoint, use eval.py:

python eval.py \
    --dataset voxceleb \
    --data_root /path/to/your/data \
    --load_path logs/<your_experiment_folder>/best.model

(Make sure to adjust the <your_experiment_folder> to point to your actual log directory).

---

📝 Citation

If you find this code or our paper useful in your research, please cite our work:

@inproceedings{woo2025mclnf,
  title={Meta-Continual Learning of Neural Fields},
  author={Woo, Seungyoon and Yun, Junhyeog and Kim, Gunhee},
  booktitle={The Thirteenth International Conference on Learning Representations},
  year={2025}
}