Releases: ICB-DCM/pyABC
Releases · ICB-DCM/pyABC
pyABC 0.11.8
pyABC 0.11.7
- Decompose ABCSMC.run for easier outer loop (#510)
pyABC 0.11.6
pyABC 0.11.5
pyABC 0.11.4
pyABC 0.11.3
pyABC 0.11.2
- Remove codacy due to excessive permission requests
- Tidy up example titles
pyABC 0.11.1
pyABC 0.11.0
This release brings, besides many bug fixes and smaller improvements, in particular outlier-robust and flexible adaptive distance functions working with any inputs. Further, a preliminary framework for semi-automatic summary statistics learning is provided (fully documented implementation planned for 0.12.0).
Diverse:
- Shorten date-time log (#456)
- Add look-ahead example notebook (#461)
- Fix decoration of
plot_acceptance_rates_trajectory
(#465) - Hot-fix redis clean-up (#475)
Semi-automatic summary statistics and robust sample weighting (#429)
Breaking changes:
* API of the `(Adaptive)PNormDistance` was altered substantially to allow
cutom definition of update indices.
* Internal weighting of samples (should not affect users).
Semi-automatic summary statistics:
* Implement (Adaptive)PNormDistance with the ability to learn summary
statistics from simulations.
* Add `sumstat` submodule for generic mappings (id, trafos), and especially a
`PredictorSumstat` summary statistic that can make use of `Predictor` objects.
* Add subsetting routines that allow restricting predictor model training
samples.
* Add `predictor` submodule with generic `Predictor` class and concrete
implementations including linear regression, Lasso, Gaussian Process,
Neural Network.
* Add `InfoWeightedPNormDistance` that allows using predictor models to weight
data not only by scale, but also by information content.
Outlier-robust adaptive distances:
* Update documentation towards robust distances.
* Add section in the corresponding notebook.
* Implement PCMAD outlier correction scheme.
Changes to internal sample weighting:
* Do not normalize weights of in-memory particles by model; this allows to
more easily use the sampling weights and the list of particles for
adaptive components (e.g. distance functions)
* Normalization of population to 1 is applied on sample level in the
sampler wrapper function
* In the database, normalization is still by sample to not break old db
support; would be nicer to also there only normalize by total sum
-- requires a db update though.
Changes to internal object instruction from samples:
* Pass sample instead of weighted_sum_stats to distance function.
This is because thus the distance can choose on its own what it wants
-- all or only accepted particles; distances; weights; parameters;
summary statistics.
Visualization:
* Function to plot adaptive distance weights from log file.