Codes for PAC-Bayesian Contrastive Unsupervised Representation Learning

Experimental environment

• Ubuntu 18.04.1 LTS
• Cuda: 10.2
• Cudnn: 7.6.3
• conda

---

git clone [email protected]:nzw0301/pb-contrastive.git
cd pb-contrastive

Optional: Install miniconda3-latest via pyenv

# cd pb-contrastive

pyenv install miniconda3-latest
pyenv local miniconda3-latest
conda create --name pac-bayes --file conde/requirements.txt -y
pyenv local miniconda3-latest/envs/pac-bayes

Install GPUs supproted PyTorch

See also the latest PyTorch.

Note that PyTorch was 1.2.0.

git clone --recursive https://github.com/pytorch/pytorch
cd pytorch
git submodule sync
git submodule update --init --recursive

export CMAKE_PREFIX_PATH=${CONDA_PREFIX:-"$(dirname $(which conda))/../"}
python setup.py install

Run codes on CNN-README.md and MLP-README.md under code dir. Then run create-tables.ipynb to create tables in the main paper.

Option: Run codes on non-iid-README.md, then run create-tables-in-Appendix.ipynb to create tables in the appendix.

Optional: Install parts of experimental dependencies on CPU via Dockerfile to run the jupyter notebook

We provide a docker environment to run notebooks on your local machine without GPUs.

For bash/zsh:

# cd code
docker build . -t pb-contrastive:latest
docker run -i -p 8888:8888 -v $(pwd):/pb-contrastive/code -w="/pb-contrastive/code" -t pb-contrastive /bin/bash

jupyter notebook --ip=0.0.0.0 --allow-root

For fish:

# cd code
docker build . -t pb-contrastive:latest
docker run -i -p 8888:8888 -v (pwd):/pb-contrastive/code -w="/pb-contrastive/code" -t pb-contrastive /bin/bash

jupyter notebook --ip=0.0.0.0 --allow-root

---

Related resources

• Presentation slides
• Presentation on Youtube

Reference

@inproceedings{NGG2020,
    title = {PAC-Bayesian Contrastive Unsupervised Representation Learning},
    author = {Kento Nozawa, Pascal Germain, Benjamin Guedj},
    year = {2020},
    booktitle = {UAI},
    pages = {21--30}
}