Brushing over notebook

[PaulJonasJost]

Mainly:

• shortening many optimizations
• Replacing Boehm with conversion reaction where possible
• Thermodynamic more robust.

Hopefully closes #1460 and #1396

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

All modified and coverable lines are covered by tests ✅

Project coverage is 82.89%. Comparing base (c5a46d4) to head (db1168a).

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@             Coverage Diff             @@
##           develop    #1470      +/-   ##
===========================================
- Coverage    82.92%   82.89%   -0.03%     
===========================================
  Files          163      163              
  Lines        13786    13786              
===========================================
- Hits         11432    11428       -4     
- Misses        2354     2358       +4     

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[Doresic]

Mostly looks good! Thanks.

I added a few comments of the same type about the reduction of starts of optimization and samples of sampling. I think it's good to reduce as much as we can so it takes slower. But I would not reduce so low that the figures look less presentable or nice. If the histograms start being just random bars it's not much use showing them. So if you didn't already, check whether the figures stay the same in character -- especially if there's no random seed that controls startpoints and how well the model is fit.

[Doresic]

I would just check that these 10 starts with the set random seed give a good model fit. It's kinda important to have fast and good sampling and profiling.

Also, is 10 enough to have a parameter histogram and scatter plot that looks like something nice?

[PaulJonasJost]

think problem would here be more that we do find optima veeery well and hence the correlation plot looks hideous. You think we should switch back to boehm? Histogram looks soso

[Doresic]

Yeah, if it makes the notebook, the figures and conclusions taken from them make more sense, I would go for that rather than it being a bit faster

[Doresic]

Remove all outputs from the notebook

[dilpath]

Fine for me, thanks!

I would not reduce so low that the figures look less presentable or nice.

re: this balance between computation time of CI vs. sufficient computation for nice figures, we could use some environment variable (e.g. CI and GITHUB_ACTIONS are always set to true during GitHub actions CI) to determine the computation effort in each notebook. This could look like this at the top of the notebooks.

import os
n_starts = 100
n_samples = 10000
if os.environ.get("CI", None) is not None:
    n_starts = 10
    n_samples = 1000

[Doresic]

Looks good! Thanks

[Doresic]

Ohh I just remembered.
At the retreat, we did talk about visualization again, and Jan again mentioned how it's always good to know in which order one should look at visualizations: Firstly, it's important to see that the model actually fits the data. If that's not true, waterfall plot and parameter plot are not important. Only when the data is somewhat fit, it's nice to look at the waterfall plot, to see convergence and parameter plot to see a fast-less-informative indication of identifiability and whether the bounds are good enough (actually changing the parameter bounds might be necessary to fit the data in the first place). Only then it makes sense to go into profiling and sampling -- to get some more accurate insight into parameter identifiability.

It would be nice if this was implemented in the getting_started notebook as a somewhat guide to estimating models. This might take some restructuring of the notebook, so I'm also ok with doing it myself in a separate PR. Wanted to mention it here so it's not forgotten :D