[Obsolete] Develop sampling output views

[magland]

No description provided.

[magland]

Ported over some plots from MCMC Monitor

And the table now has a column for Chain and a column for Draw

And the summary splits out the means by chain

[WardBrian]

Some thoughts just looking at it/playing around, before looking at the code:

• Traceplots are super useful for real-time monitoring like in mcmc-monitor, but are less useful after the fact. This is in part because you lose the sense of "time" - in mcmc-monitor, you can view both how well the chains are mixing and if one of them is significantly behind the others. But in this regime where you're doing analysis after the fact, I believe they're mostly used as a diagnostic, and in my experience usually only after one of the summary statistics like $\hat{R}$ or ESS indicates that there was a problem.
  If you want to show one plot for each variable, the first thing I would start with is a histogram. Then probably a 2-d scatter between pairs of variables, and traceplots only if the user wants them
• Adding the Chain and Draw columns to the display is nice, but if the idea behind the download is to be able to use other tools in the Stan ecosystem, we should avoid adding any columns that Stan wouldn't have. Creating num_chains different csv files is still probably the best thing on that front
• I think the mean of all draws is more meaningful than the individual chain means, if you're picking one. Showing it for each chain maybe can give you a sense of if the chains didn't mix, but Rhat is better for that purpose anyway.

Here's the output of Stan's stansummary command, for reference:

Inference for Stan model: bernoulli_model
1 chains: each with iter=(1000); warmup=(0); thin=(1); 1000 iterations saved.

Warmup took 0.0080 seconds
Sampling took 0.048 seconds

                Mean     MCSE   StdDev     5%   50%   95%  N_Eff  N_Eff/s  R_hat

lp__            -7.3  3.8e-02     0.76   -8.8  -7.0  -6.7    399     8313   1.00
accept_stat__   0.92  3.5e-03  1.2e-01   0.64  0.97   1.0   1222    25458   1.00
stepsize__      0.96      nan  6.7e-16   0.96  0.96  0.96    nan      nan    nan
treedepth__      1.3  1.6e-02  4.7e-01    1.0   1.0   2.0    874    18204   1.00
n_leapfrog__     2.4  3.6e-02  1.1e+00    1.0   3.0   3.0    926    19286   1.00
divergent__     0.00      nan  0.0e+00   0.00  0.00  0.00    nan      nan    nan
energy__         7.8  5.2e-02  9.7e-01    6.8   7.6   9.8    353     7347   1.00

theta           0.25  6.1e-03     0.12  0.079  0.24  0.48    418     8702    1.0

Samples were drawn using hmc with nuts.
For each parameter, N_Eff is a crude measure of effective sample size,
and R_hat is the potential scale reduction factor on split chains (at 
convergence, R_hat=1).

All of these are considering all chains at once, but N_Eff through R_hat need to know which chain the draws originated from to do what they do correctly

[magland]

Okay thanks! I'll work on that.

[magland]

I made some changes based on your comments Brian.

I tried to get the summary table to match the output you pasted.

That involved porting code over from MCMC Monitor for calculating ESS Rhat, etc

There is a new "Histograms" tab that shows the parameter histograms. Ported over from MCMC Monitor.

I left in the trace plots tab for now, coming after the histograms tab. I think it's useful to have that available... but I do see what you are saying about how it's more useful for RT monitoring.

Regarding export to CSV. I propose we provide two options. (1) Downloading as single CSV with chain/draw columns (2) Downloading as a .Zip file of CSVs, one for each chain, in the format of Stan output. For the second, I suggest a separate PR. Thoughts?

@jsoules @WardBrian

[WardBrian]

The summary visually looks great.
The rhat/ess calculations are a bit tricky and require some verification.

For histograms, if you add histnorm: "probability", to the data object, you get normalized histograms, which are a bit more typical in this context compared to the raw counts

[magland]

For histograms, if you add histnorm: "probability", to the data object, you get normalized histograms, which are a bit more typical in this context compared to the raw counts

I have added histnorm: "probability"

I have also changed the title for this PR

[magland]

Replaced by #32