SigmoidBoundedParameter value does not match given value

[GiM6114]

Hi !

When instantiating a SigmoidBoundedParameter object with value v and some bounds, the value attribute of the object initializes not at v but at random values, negative from my experience (which can cause issues later for optimization if the model is not robust to negative values for the parameter for instance). Comparing its implementation to BoundedParameter, maybe it just lacks super().init(value)?
Please find attached a notebook that shows this for RWW parameters and compares it to BoundedParameter which has the correct behavior.

test_RWW_sigmoidboundedparameter.ipynb

[mapi1]

Thanks for the write-up and MWE, that made this easy to dig into.

Mainly this is working as intended, not a missing super().init(). For SigmoidBoundedParameter (and its parent TransformedParameter) .value deliberately holds the unconstrained leaf, the pre-image in logit space, because that's what the optimizer differentiates in unbounded space. The bounded value is recovered on the fly via __jax_array__(), so the model only ever sees a value inside [low, high]. The negative number for tau_s is just logit(100/300), not a corrupted init.

It's the same contract across all types, just less "different" elsewhere: BoundedParameter.value is the unclipped value, NormalizedParameter.value is normalized to ones.

Your confusion was fair though, since .value being the raw leaf was not easily discoverable. In an attempt to make it clearer I:

• Added a .constrained_value property returning the value used in computation (same as collect_parameters, or just passing the parameter into any jnp/np op (calling __jax_array__() implicitly)).
• show_parameters / PrintParameterCallback now lead with the constrained value and bounds.
• Documented .value vs .constrained_value in the docstrings and parameters doc.

Currently, the changes live on main, I plan to release this as v0.3.2. Let me know if this sufficiently solves the issue for you or you think we can make this behavior even more clear or easier to understand.
Here is an updated script: test_RWW_sigmoidboundedparameter.py