IWAE

chainer implementation of IWAE. This code is inspired code1 and code2.

Abstract

By utilizing k sampling, IWAE reduces the ELBO. If k is 1, it corresponds to the standard VAE.

python iwae.py --sampling_number=k　　

I check the value of ELBO after training.

k=1	k=5	k=10
100.14	98.88	98.34

By increasing the value of k, you can reduce ELBO.

The detail is described in iwae_analysis.ipynb.

Name		Name	Last commit message	Last commit date
Latest commit History 6 Commits
IWAE_analysis.ipynb		IWAE_analysis.ipynb
README.md		README.md
iwae.py		iwae.py
sampled_sampling_num_10.png		sampled_sampling_num_10.png
test.png		test.png
test_reconstructed_sampling_num_10.png		test_reconstructed_sampling_num_10.png
train.png		train.png
train_reconstructed_sampling_num_10.png		train_reconstructed_sampling_num_10.png
training_error_sampling_number_10.csv		training_error_sampling_number_10.csv