blei-lab / edward

A probabilistic programming language in TensorFlow. Deep generative models, variational inference.
http://edwardlib.org
Other
4.83k stars 759 forks source link

SGD optimizer #878

Open BelhalK opened 6 years ago

BelhalK commented 6 years ago

I believe the ELBO is maximized using SGD in KLQP.py First: By default only one data is used to compute the estimator of the full gradient. Does n_samples argument allows to fix the number of gradients to calculate in order to estimate this noisy gradient? Second: Are you considering using SAG (see Le Roux et al, 2013) or SAGA (see Defazio et al, 2014) updates instead of SGD?