Vectorise particle filter #365
Conversation
Sometimes the addition of noise to the last measurement call led to a target being out of the fov of the radar.
sdhiscocks
requested review from
a team,
sdhiscocks,
oharrald-Dstl and
jmbarr
and removed request for
a team
sdhiscocks
added
breaking
I took a look at this and agree that it's harder to read, and also would be better not too pass weights orders around. Out of interest, I tried using the log values to sort, to avoid overhead of the custom type, and then pure NumPy C optimised code could do the sort. I found applying the following to f4a1966 (commit before diff --git a/stonesoup/resampler/particle.py b/stonesoup/resampler/particle.py
index bb275766..77c11eb1 100644
--- a/stonesoup/resampler/particle.py
+++ b/stonesoup/resampler/particle.py
@@ -26,7 +26,8 @@ class SystematicResampler(Resampler):
particles = Particles(particle_list=particles)
n_particles = len(particles)
weight = Probability(1/n_particles)
- weight_order = np.argsort(particles.weight, kind='stable')
+ weights = np.array([weight.log_value for weight in particles.weight])
+ weight_order = np.argsort(weights, kind='stable')
cdf = [v.log_value for v in np.cumsum([particles.weight[i] for i in weight_order])]
# Pick random starting point |
| state_vector: StateVectors = Property(default=None, doc="State vectors of particles") | ||
| weight: MutableSequence[Probability] = Property(default=None, doc='Weights of particles') | ||
| parent: 'Particles' = Property(default=None, doc='Parent particles') | ||
| particle_list: MutableSequence[Particle] = Property(default=None, |
There was a problem hiding this comment.
Why do you need a collection of StateVectors and a list of Particles? Could we do away with Particle altogether?
There was a problem hiding this comment.
The list of particles was for backwards compatibility. In particular many of the unit tests loop over lists of particles so this was a good sanity check. I guess now I have finished vectorising and am confident I've not changed the output we could get rid of Particle and edit all the unit tests to use the new class?
There was a problem hiding this comment.
So little on the fence either way with this one. The only thought I had, is there may be cases where you can't easily vectorise a function or operation, and therefore being able to loop over particles (as Particle states) to run a function/calculation, and use that list to repopulate the Particles, could be useful.
We probably should encourage people not to do this however, so should update our code/tests.
Codecov Report
@@ Coverage Diff @@
## master #365 +/- ##
==========================================
+ Coverage 94.09% 94.18% +0.08%
==========================================
Files 133 133
Lines 6168 6222 +54
Branches 894 911 +17
==========================================
+ Hits 5804 5860 +56
+ Misses 263 262 -1
+ Partials 101 100 -1
Flags with carried forward coverage won't be shown. Click here to find out more.
Continue to review full report at Codecov.
|
idorrington-dstl
force-pushed
the
vectorise_particle_filter
branch
from
f4e1751
to
cacf463
Compare
idorrington-dstl
force-pushed
the
vectorise_particle_filter
branch
from
96dac02
to
2532a8e
Compare
…r each particle. Particles behaves like Particle when there is only one particle, so that we only need Particles. Made the particles weights a list of probablilities to make it more like particle
…ist of particles. This makes it backwards compatible
…aster. Removed unneccesary deep copies
Test making Particles from a list of particles
…ility above the CDF. Is faster and simpler.
Co-authored-by: Steven Hiscocks <sdhiscocks@dstl.gov.uk>
…an update weights inplace, which is slightly faster and neater.
…articles state as input
Tidying up
idorrington-dstl
force-pushed
the
vectorise_particle_filter
branch
from
3f58ca3
to
8931fd9
Compare
…rised particles state
…rtesian2DToBearingRange models.
…This method doesn't get called often, so doesn't affect run time significantly.
…n ParticleState (the weights are made as an array when the Particles object is made, so this never does anything)
…rticles class. Check that either all particles have parents or none do and return ValueError otherwise.
…pler method to allow numpy to do its thing.
Added tests to check that the Particles object will give ValueErrors when input is wrong.
idorrington-dstl
force-pushed
the
vectorise_particle_filter
branch
from
8931fd9
to
0c42bd9
Compare
Vectorised and optimised the particle filter.
Did this by creating a new particles object where the state_vector, weight, and parent attributes are vectors/arrays corresponding to the same attributes of the individual particles.
The most significant of the other optimisations made is that the pdf function now optionally outputs the indexes to sort the particles by size order. This saves us running np.argsort on the particle weights in the resample step leading to a ~35% decrease in run time (though it does make the code slightly harder to follow as the order of the particles is calculated in a different function to where it is used).