Variation Inference output not read properly if eta (stepsize) != 1

[ro-joshi]

Summary:

The output csv from variational inference is not properly read if eta (stepsize) is not equal to 1.

Description:

The line here: https://github.com/stan-dev/cmdstanpy/blob/master/cmdstanpy/utils.py#L564 requires that eta = 1 in order to read the output csv. However, eta can take on different values (e.g. 0.1, etc.), that can be set by the user or adaptively determined by the ADVI algorithm. See https://mc-stan.org/rstan/reference/stanmodel-method-vb.html for comparison.

Current Version:

0.9.5

[mitzimorris]

thanks for reporting - do you have an example CSV file or snippet you could share?

[ro-joshi]

Thanks for taking a look, snippet below. I cut off columns and rows in the brackets <> for clarity

# stan_version_major = 2
# stan_version_minor = 23
# stan_version_patch = 0
# model = factor_model
# method = variational
#   variational
#     algorithm = meanfield (Default)
#       meanfield
#     iter = 10000 (Default)
#     grad_samples = 1 (Default)
#     elbo_samples = 100 (Default)
#     eta = 1 (Default)
#     adapt
#       engaged = 1 (Default)
#       iter = 50 (Default)
#     tol_rel_obj = 0.01 (Default)
#     eval_elbo = 100 (Default)
#     output_samples = 1000 (Default)
# id = 0 (Default)
# data
#   file = /tmp/tmpkx56q518/3b3icq7l.json
# init = 2 (Default)
# random
#   seed = 41822
# output
#   file = /tmp/tmpkx56q518/factor-202008052035-1-y1qwu9eq.csv
#   diagnostic_file =  (Default)
#   refresh = 100 (Default)
# num_threads = 1
lp__,log_p__,log_g__,phi,<...further columns...>
# Stepsize adaptation complete.
# eta = 0.1
0,0,0,2.84783,1.50599,1.0663,<...further columns...>
0,-1.45135e+06,-6556.27,2.62732,<...further columns...>
<...further rows...>

[mitzimorris]

fixed by https://github.com/stan-dev/cmdstanpy/pull/283, closing