Fast core functions for navis
re-implemented in Cython.
The idea is that navis
will use fastcore
if installed and fall back to
the pure-Python / numpy implementation if not.
Currently implemented:
- vertex similarity (Jarrell et al., 2012)
- shortest path from source to target (~40x faster than iGraph)
- geodesic distance matrix (up to 100x faster than scipy)
See further down for details.
I'm still figuring out the best way for building and packaging pre-compiled binaries (i.e. wheels). For now, you will need to compile it yourself during setup. This requires a C-compiler to be present (see here for a very brief explanation).
$ pip3 install git+git://github.com/navis-org/fastcore@main
>>> import numpy as np
>>> import navis
>>> import fastcore
>>> # Grab an example skeleton
>>> n = navis.example_neurons(1)
>>> # Time navis' scipy-based function for all-by-all geodesic distances
>>> %time m1 = navis.geodesic_matrix(n, weights=None)
CPU times: user 4.58 s, sys: 153 ms, total: 4.73 s
Wall time: 4.73 s
>>> # Time the analogous function in fastcore
>>> %time m2 = fastcore.geodesic_matrix(n.nodes.node_id.values, n.nodes.parent_id.values)
CPU times: user 2.17 s, sys: 173 ms, total: 2.35 s
Wall time: 258 ms
>>> # Make sure results are the same
>>> np.all(m1 == m2)
True
We need openmp
for threaded processing. Without it, this is not much faster
than the pure numpy implementations. If your compiler does not support
openmp
, you will get an error along the lines of -fopenmp not supported
.
This happens e.g. on OSX if you use the clang bundled with XCode. In my case,
I was able to work around it by installing llvm
with homebrew and then
adding a couple flags to my ~/.bash_profile
to make sure the homebrew llvm
is actually used:
export PATH="/usr/local/opt/llvm/bin:$PATH"
export LDFLAGS="-L/usr/local/opt/llvm/lib -Wl,-rpath,/usr/local/opt/llvm/lib"
export CPPFLAGS="-I/usr/local/opt/llvm/include"
To compile the extensions in place:
$ python setup.py build_ext --inplace