Normalizing constant for q(z|ψ)

mingzhang-yin / SIVI

A variational inference method with accurate uncertainty estimation. It uses a new semi-implicit variational family built on neural networks and hierarchical distribution (ICML 2018).

http://proceedings.mlr.press/v80/yin18b/yin18b.pdf

MIT License

52 stars 14 forks source link

Normalizing constant for q(z|ψ) #1

Closed artsobolev closed 5 years ago

artsobolev commented 5 years ago

Hello

Thank you for posting the source code for your experiments! While viewing it I noticed that computation of the density q(z|ψ) at SIVAE.py#L113 seemingly lacks the normalizing constant that includes the changing variance term.

Am I missing something? Is it cancelled out somehow?

mingzhang-yin commented 5 years ago

Hi,

Thanks for your interest in SIVI. The normalizing constant with covariance is on #L114 log_H_iw = tf.log(tf.reduce_mean(ker,axis=2))-0.5*tf.reduce_sum(z_logv_iw,2)