Development ideas

Some ideas for future development of MDT.

• to discuss: add ‘scale_factor’ to the parameters as a simple scaling factor

  • for use in MLE and FIM
  • replacing ScaleTransform and scale_factor in numdiff_info
  • perhaps add it to MCMC as well

• MPT (Microstructure Prototol Toolbox)

  • Perhaps a different package using MDT and MOT for protocol optimization

• More optimization routines:

  • L-BFGS-B
  • BOBYQA or NEWUOA from Powell
  • Particle Swarm
  • See the book “Introduction to derivative free optimization” or similar books

• More sampling routines:

  • emcee

• Automatic derivatives

  • This is only doable with OpenCL 2.2

• Python 3.7:

  • removed ordereddict
  • use f-strings

• OpenCL 2.0:

  • remove priors and merge with the constraints

• remove the JohnsonNoise model

  • have the OffsetGaussian an separate input for the offset

• protocol options using a selection matrix.

  • i.e. per protocol row a 0 or 1 if that protocol is supposed to be used
  • Allows protocol options per voxel.

• remove mot_float_type, replace with either float or double depending on insight/tests
• move g to gx, gy, gz
• Move to using an OpenCL Context instead of the devices
• Add more model documentation, in particular what all the output maps are

• DynamicGlobal kernel data element

  • as a substitute for Local where needed

• Add more documentation on how to change devices