Hi,
I should also comment the matrices, in addition to the other comments.
I think the code in lib/dispersion/ns_2site_star.py is now excellently
documented. A novice NMR user can now learn everything from it! I
don't think that anything else is needed, though see what you think.
Note that this code is currently evolving.
Edward, I also have a question to line 75: where does the factor 2 come
from? that was not in the original version.
I don't see what it does. I think it should not be there.
I shifted it from the inside of the for loop. Specifically the
original expression "sci.expm(cR*2.*tcp))" has been changed to
"expm(cR2*tcp))", where cR2 is cR pre-multiplied by 2. This decreases
the number of cR*2. multiplications from the number of dispersion
points to 1. It is just one of a number of small speed ups.
I plan on identifying more places in the code where calculations can
be shifted up the code chain to absolutely minimise the total number
of mathematical operations. For example I am shifting some
calculations outside of the
lib.dispersion.ns_2site_star.r2eff_ns_2site_star() function to be
inside the func_ns_2site_star() target function of the
target_functions.relax_disp.Dispersion class. This is to minimise
calculations for each spin in the cluster and magnetic field strength,
i.e. the number of calculations drops from N_spins*N_B0 to 1. For
large spin clusters this is significant.
Another speed up used within relax target functions is the matrix
creation. Currently this involves first automatic memory allocation
at the start of each call followed by memory deallocation. The way
this is handled in relax is that when the target function class is
initialised, then all of the needed matrices are initialised. Then
the target function calls just update the pre-existing matrices. This
often has a significant effect on optimisation speed by minimising the
calls to the Python garbage collection and memory allocation routines.
In some cases, scalar multiplication and dot products for example, a
new matrix is created automatically hence this trick cannot always be
used (though faster alternatives can be created if desired).
Cheers,
Edward
Re: numerical cpmg fit