Flow-based Deep Generative Models

This is the course final project for SDS 384 Statistical Maching Learning Optimization.

Models implemented (currently):

• Basic flows (planar & radial)
• RealNVP (adapted from the original TensorFlow implementation [link])

RealNVP Usage:

# mnist
python train.py --dataset mnist --hidden_dim 32 --num_levels 2 --num_residual_blocks 5
# cifar10
python train.py --dataset cifar10 --hidden_dim 64 --num_levels 2 --num_residual_blocks 8
# celeba
python train.py --dataset celeba --hidden_dim 32 --num_levels 5 --num_residual_blocks 2

Notes on CelebA dataset

It has been a well-known issue that first-time download of CelebA via torchvision API, i.e. torchvision.datasets.CelebA will fail almost surely with option download=True since Google imposed their restriction on data access and daily limit. One solution is to directly open the dataset's official Google Drive link [here] in a browser and download it. Alternatively, we may resort to the kaggle datasets API. To use this solution, please set FROM_KAGGLE in the datasets.py to True. Plus, you will need a Kaggle account and an authentication token.

Generated Celeb-A samples over 30 epochs

References:

• Rezende, D. & Mohamed, S.. (2015). Variational Inference with Normalizing Flows. Proceedings of the 32nd International Conference on Machine Learning, in Proceedings of Machine Learning Research 37:1530-1538 Available from https://proceedings.mlr.press/v37/rezende15.html.
• Dinh, L., Sohl-Dickstein, J., & Bengio, S. (2017). Density estimation using Real NVP. 5th International Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. Opgehaal van https://openreview.net/forum?id=HkpbnH9lx
• van den Berg, R., Hasenclever, L., Tomczak, J., & Welling, M. (2018). Sylvester normalizing flows for variational inference. proceedings of the Conference on Uncertainty in Artificial Intelligence (UAI). http://auai.org/uai2018/proceedings/papers/156.pdf
• Kingma, D. P., & Dhariwal, P. (2018). Glow: Generative Flow with Invertible 1x1 Convolutions. In S. Bengio, H. Wallach, H. Larochelle, K. Grauman, N. Cesa-Bianchi, & R. Garnett (Reds), Advances in Neural Information Processing Systems (Vol 31). Opgehaal van https://proceedings.neurips.cc/paper/2018/file/d139db6a236200b21cc7f752979132d0-Paper.pdf
• Grathwohl, W., Chen, R. T. Q., Bettencourt, J., Sutskever, I., & Duvenaud, D. (2019). FFJORD: Free-form Continuous Dynamics for Scalable Reversible Generative Models. International Conference on Learning Representations. https://openreview.net/forum?id=rJxgknCcK7