Importance Weighted Auto Encoder

A tensorflow implementation of Importance Weighted Auto Encoder ^[1]

Requirements

• tensorflow
• numpy
• matplotlib

Usage

python main.py  --dataset {mnist,omniglot} \
                --k <# of particles for training> \
                --test_k <# number of particles for testing> \
                --n_steps <# of steps> \
                --batch_size <batch size>

Datasets

• MNIST - automatically downloaded by tensorflow
• OMNIGLOT - run download_omniglot.sh

Results

The following are the log-likelihood values after training for 400,000 steps with a batch size of 100 for different number of particles (k) and test_k = 5000.

k	NLL (MNIST)	NLL (OMNIGLOT)
1	90.26	114.68
5	88.49	112.25
50	87.34	110.31

References

[1] Burda, Y., Grosse, R. and Salakhutdinov, R., 2015. Importance Weighted Autoencoders. arXiv preprint arXiv:1509.00519.