Create distance_array that can process AtomGroup

[orbeckst]

Expected behaviour

distance_array and friends are used often in user code; they should take AtomGroup instances as arguments

Actual behaviour

Exception is raise because only coordinate arrays are allows

Proposed change

Leave the distance_array in lib.distances as the bare bones version

Create a wrapper in analysis.distances that allows use of AtomGroup.

I'm not totally happy with keeping the same name but switching names in one either of the two places will break code.

[richardjgowers]

It'd be cool to have the function, but I don't think it should have the same name. Something like distance_matrix might work?

[orbeckst]

You could argue that we'd overload dist.distance_array so it might be ok to use the same name. In scipy the equivalent function to lib.distances.distances_arrayis called cdist (although no pbc handling).

distance_matrix sounds potatoe-potato... no offense to the English English speaking world ;-).

Maybe just distances(*groups) and roll self_distance_array and distance_array into one function? In particular we should have an optimized code path for distance_array(g, g) anyway. It is often convenient to get the self_distance_array back as a symmetric matrix but reformatting the 1D representation of self_distance_array can take longer than the calculation itself.

[kain88-de]

How about all functions imported from MDAnalysis.analysis.distances accept Atomgroups and all functions in MDAnalysis.lib.distances accept numpy arrays? That would be a clear distinction.

[jbarnoud]

It would break a lot of code, though.

[richardjgowers]

@jbarnoud it would have to have a different name. We can't have distance_array having different arguments depending on where it's imported from. analysis.distances.distances() can be something smart like @orbeckst suggested, lib.distances.distance_array() is for np arrays

[richardjgowers]

Here's a weird idea, make a distance_array method... AtomGroup.distance_array(other)...

[jbarnoud]

@richardjgowers I do not really have a rational about it, but it does not sound right. It is most likely a matter of habit, though. Trying to come up with a rational anyway, I do prefer the symmetry between the behaviour on array and the one on atom groups; also the method that works on atom groups could also work on position arrays if at least one of the argument is an atom group (I often compute the distance between an atom group and a centre of mass for instance); finally the method space of the atom groups is already quite cluttered.

Just some thoughts, though.