Mdt

====================

Changed
-------
- Changed the maximum bounds to real numbers, anticipating bounded optimization methods. It used to be -inf and +inf, it is now -1e20 and 1e20.
- The initialization dictionary now also recognizes MDT models with a suffix. Like "BinghamNODDI_r1_MySuffix", which will load the initialization from "BinghamNODDI_r1".

====================

Added
-----
- Adds the ``use_cascaded_inits`` flag to the model fit command to simplify the new initialization interface.
- Adds reduced Ramani QMT models to MDT.

Other
-----
- Made the generate_wm_mask function more general.
- Made all mot_float_type4 vectors float4 and moved some other parts from mot_float_type to either float or double depending on the situation.
- Removed the building pattern from the composite models.

====================
This version marks the deprecation of the Cascade models in MDT.
Instead, MDT now strikes a balance between customizability and ease of use.
For ease of use, using the GUI and command line MDT now automatically selects a good starting point based on pre-set rules.
For customizability you can use the Python API in which you first compute the initialization point and then manually provide it to the model fitting.
By providing it manually you can have a finer grained control over the initialization settings.

For more details on how to properly initialize in newer newer versions see :ref:`model_fitting`.
Even though this functionality is now deprecated, it will still be available for the near future to allow users the chance to move to the new workflow.

Changed
-------
- Deprecated the Cascade interface in favor of a more direct control of the initialization point.
- Refactored the documentation and added sampling documentation.

====================

Changed
-------
- Changed the specification of the volume selection in the composite models to allow selection based on all protocol columns.

Other
-----
- Small update to the BallStick cascade inits to make sure they are within bounds.
- Small update to the weight numerical differentiation to not use the upper bound anymore. Provides for slightly better uncertainty computations.

====================

Changed
-------
- Updated the Kurtosis initial parameter positions to try to prevent out of bounds problems within the optimization routines.

====================
The way the boundary constraints of the parameters are enforced is changed.
Previously MDT used parameter transformations to enforce boundary conditions, in this new version we use the new support of MOT for the boundary conditions.
In the new MOT version, boundary conditions (box-constraints) are handled by returning INFINITY if the bound is violated.
While this is a crude way of enforcing boundary conditions, it does relieve us of the parameter transformations in MDT.

As a result, fits looks slightly less noisy overall, and it seems to fit better in relaxometry models.


Added
-----
- Adds additional parameters types to link the likelihood functions better with the models.
- Adds support for the special parameter 'noise_std' to inject the current noise standard deviation into a compartment model.

Changed
-------
- Changed the way the bounds are handed in the optimization.
- Changed the likelihood function to always include the constant terms.

Fixed
-----
- Fixed bug in the mdt-estimate-noise-std method.

Other
-----
- Small update to the GDRCylinder bounds.
- Following the support in MOT for boundary constraints, changed the parameter transformations of the parameters.
- Small fix to the batch profiles, it did not pick up the noise_std.txt files.
- Documentation updates.
- Small update to the mdt-math-img command. Better way of handling multiple outputs.