Parameter plot w/ hier. pars, noise estimation for splines

[Doresic]

Extended the parameter plots to include hierarchical parameters. Not all, but the interpretable ones: scaling, offset, and noise. So excluding the very numerous spline/optimal scaling inner parameters.

• The amici objective has an attribute self.inner_parameters in which it will save the best last inner parameters of the current optimization run (start point).
• Then at the end of the optimization run, it will be saved into the optimization_result dictionary.
• The parameter plotting routine will check for inner parameters in parameters.py/handle_inputs and concatenate them to the outputs accordingly.

Implemented the noise estimation for observables with non-linear monotone data.

• The noise is calculated analytically, basically in the same way as with relative data.
• To specify that the noise parameters should be optimized hierarchically, as with relative data, the parameterType column entry of the noise parameter should be sigma. Then the SplineInnerProblem will extract the noise parameters.

[Doresic]

Example of what the plot looks like for the Boehm model with hierarchically estimated noise:

Since the scale of the hierarchical parameters is lin, it will be harder to see the kinetic parameters nicely. So I've added a flag plot_inner_parameters to the parameters plotting routine. The default is True.

[codecov-commenter]

Codecov Report

Merging #1061 (05fa654) into develop (fad8041) will decrease coverage by 0.10%.
The diff coverage is 86.56%.

❗ Your organization is not using the GitHub App Integration. As a result you may experience degraded service beginning May 15th. Please install the Github App Integration for your organization. Read more.

@@             Coverage Diff             @@
##           develop    #1061      +/-   ##
===========================================
- Coverage    84.88%   84.78%   -0.10%     
===========================================
  Files          147      147              
  Lines        11252    11384     +132     
===========================================
+ Hits          9551     9652     +101     
- Misses        1701     1732      +31     

Impacted Files	Coverage Δ
pypesto/visualize/ordinal_categories.py	8.63% <0.00%> (-0.04%)	⬇️
pypesto/visualize/spline_approximation.py	15.00% <0.00%> (-0.26%)	⬇️
pypesto/hierarchical/optimal_scaling/solver.py	93.03% <50.00%> (-1.44%)	⬇️
pypesto/optimize/optimizer.py	88.29% <66.66%> (-0.16%)	⬇️
pypesto/hierarchical/optimal_scaling/problem.py	96.74% <80.00%> (-0.40%)	⬇️
...sto/hierarchical/spline_approximation/parameter.py	81.81% <80.00%> (+1.81%)	⬆️
...to/hierarchical/spline_approximation/calculator.py	76.66% <83.33%> (+0.74%)	⬆️
...pesto/hierarchical/spline_approximation/problem.py	89.18% <85.36%> (-1.72%)	⬇️
...ypesto/hierarchical/spline_approximation/solver.py	83.68% <85.89%> (-0.82%)	⬇️
pypesto/visualize/parameters.py	83.73% <94.11%> (+1.18%)	⬆️
... and 7 more

... and 3 files with indirect coverage changes

[FFroehlich]

Graph looks nice! Only looked at changes is pypesto/objective/amici/amici.py and pypesto/petab/importer.py

[dweindl]

Thanks. Looks good as far as I can tell.

[Doresic]

Base test is failing very often. Specifically test_objective.py/test_finite_difference_checks():

# Test the single step size `check_grad` method.
        eps = 1e-5
        result_single_eps = objective.check_grad(np.array([theta]), eps=eps)
>       assert result_single_eps['rel_err'].squeeze() == rel_err(eps)
E       assert 1.1987971168495947e-10 == 1.1987977630586244e-10
E        +  where 1.1987971168495947e-10 = <bound method NDFrame.squeeze of x_0    1.198797e-10\nName: rel_err, dtype: float64>()
E        +    where <bound method NDFrame.squeeze of x_0    1.198797e-10\nName: rel_err, dtype: float64> = x_0    1.198797e-10\nName: rel_err, dtype:

It seems that the assert is checking whether very small numbers are equal.
Is this failing in other PRs?

[dweindl]

Is this failing in other PRs?

Dunno, but it definitely makes sense to change that to something like numpy.testing.assert_array_almost_equal_nulp

[PaulJonasJost]

Looks good to me. Do we have a test for the inner parameter visualisation? Could be useful just to ensure its running later on (as other visualization tests)

[PaulJonasJost]

Is this failing in other PRs?

Dunno, but it definitely makes sense to change that to something like numpy.testing.assert_array_almost_equal_nulp

should be fixed in #1064

[stephanmg]

Looks fine. I agree on Paul's comment, a test for inner parameter visualization would be great.

[Doresic]

I've just added a test for the parameter plot with inner hierarchical parameters