Problem with using ODE solvers with autodifferentiation

#1

I am using the DifferentialEquations package to simulate Hodgkin Huxley neurons. This involves evaluating certain parameters inside the ODE function using dependent variables for which I need to make function calls from within the ODE function, e.g.
V = u[1];
[param1, param2] = my_function(V);
du[1] = (V - param1) / param2;

To speed up the code I have made my_function accept only Float64 by type assertion:

function my_function(V::Float64)
#operations
end

The problem is while using certain solvers that use autodiff (e.g. QNDF() or Rodas4()) I get an error when instead of passing a Float64 a Dual type number is passed. The solution here describes two workarounds: 1st is obvious (turn off autodiff). However i don’t understand the second. Could anyone please explain it to me? If I just turn off autodiff the accuracy of the solution will take a hit right?
Thanks

#2

Restricting the input types has no effect on performance, just remove the restriction and you’ll be fine :grinning:

https://docs.julialang.org/en/v1/manual/performance-tips/index.html

See “type declaratios”

5 Likes
#3

Actually it does seem to make a difference. Please see attached screenshot:
I used QNDF() from the DifferentialEquations package: without any type assertions on the input to the functions computing the parameters (hence with autodiff=true) vs. with Float64 type assertions (hence autodiff=false). However at the expense of increased time, the solution (neuron membrane voltage plot in the plot window) did not drift off with autodiff turned on (blue=autodiff, orange=no autodiff). So assuming autodiff gives the better solution, i guess I have to compromise on the time and memory effeciency.
Are there better ways to approach this? Please advise.
Thanks
Screenshot%20from%202019-10-12%2011-04-27

#4

Type assertions in function signatures don’t do anything for performance. Functions autospecialize on types. That’s the very essence of why Julia works.

#5

Thanks for clarifying. So is it using/not using autodiff that makes the difference here?

#6

Yes. It has to build double caches since it needs to use (m+1)*n sized caches for m partials

#7

Thanks a lot for the help!

#8

You can probably use StaticArrays.jl to reduce the difference in timing. Indeed, you seem to have a low dimensional ODE.

2 Likes