-
Notifications
You must be signed in to change notification settings - Fork 24
Bumps architecture
this is an overview, glossing over most details
This is the central class defined and used in bumps
. Users define their problem by
initializing FitProblem
with a list of models and optional constraints.
FitProblem
is then used in the fitters, where at each step the parameters are updated then
the negative-log-likelihood function or nllf()
is evaluated.
This module defines a number of fitting engines, each of which takes a FitProblem
as input and can then
take steps to minimize the nllf
by trying different combinations of parameters.
Two of the most commonly-used fitters are:
Uses the so-called "amoeba" or "Nelder-Mead" algorithm to quickly descend
into the closest minimum. This is the default fitter in bumps
, and it can be very effective.
The most advanced fitting engine in bumps
.
Once converged, it samples the space of possible Parameter combinations near the "best fit" and records the
history of Parameter
vectors along with the associated nllf()
values. This history can be used
to extract Parameter
uncertainties and correlations statistically.
- For running fits in batch mode or non-interactively, exporting all the fit results to an output folder on completion.
- Extremely useful for fitting on HPC infrastructure (e.g. SLURM jobs)
- The original
WxPython
GUI forbumps
- allows interactive modification of the loaded
FitProblem
- select/deselect fit parameters
- modify parameter values
- realtime feedback in
data
panel
- controls for starting and monitoring fits
- convergence plot
- plots of the statistical outputs of
DREAM
fits.
A recently-developed second GUI for bumps
, based on a web server (with bidirectional socket.io
server included).
- replicates all the panels in
bumps.gui
- uses
plotly
for interactive plots on the web - enables interacting with fit sessions running remotely on HPC (by forwarding the web ports)