Allow simulation output to be cached #1404

Activity

1. MichaelClerx

   added

   feature

   on Oct 12, 2021
2. MichaelClerx commented on Aug 13, 2022

   

   MichaelClerx

   on Aug 13, 2022

   Last edited by MichaelClerx

   MemberAuthor

   More actions

   We might only need it for Error and LogPdf ?

   Here's an example for an error.
   It uses the lru_cache decorator.

   The only complication is that, because it needs to place x in a dict, x needs to be hashable. Numpy arrays are not hashable, so here it converts it to a tuple first.

   class CachingError(pints.ErrorMeasure):
       """
       Wraps around another error measure and provides a ``sensitivities()``
       method for use with e.g. ``scipy``.
   
       All calls to the error and to :meth:`sensitivities` are mapped to
       :meth:`evaluateS1`. To reduce redundant calculations, up to 32 results of
       calling ``evaluateS1`` will be kept in cache.
   
       Note: Using this wrapper for methods that don't require sensitivities will
       result in very poor performance.
       """
   
       def __init__(self, error):
           self._e = error
   
       def __call__(self, x):
           return self._both(tuple(x))[0]
   
       @functools.lru_cache(maxsize=32)
       def _both(self, x):
           return self._e.evaluateS1(x)
   
       def evaluateS1(self, x):
           return self._both(tuple(x))
   
       def n_parameters(self):
           return self._e.n_parameters()
   
       def sensitivities(self, x):
           return self._both(tuple(x))[1]
3. MichaelClerx commented on Aug 13, 2022

   

   MichaelClerx

   on Aug 13, 2022

   MemberAuthor

   More actions

   Just noticed that this is not required for scipy, you should set jac=True instead and have it call evaluateS1 :D
4. pints-team

   locked and limited conversation to collaborators on Jun 23, 2024
5. MichaelClerx

   converted this issue into a discussion #1685 on Jun 23, 2024