Richard Hayes
Richard Hayes
Make the factor optimiser's optimise method take a factor approximation and return a new model distribution. This is a prerequisite for parallelisation at the factor optimisation level.
https://pypi.org/project/autofit/ Images etc. don't seem to be interpretted correctly
Ensure log evidence is non-zero by including the self message in the computation for #544
Some of the values in search_direction become infinite resulting in bad values being passed to the fit. Here's the offending search_direction: ```python VariableData({: -inf, : -inf, : 14.556109976510278}) ``` This...
../../../../autofit/graphical/declarative.py:114: in optimise self.mean_field_approximation() ../../../../autofit/graphical/expectation_propagation.py:352: in run model_approx, status = optimiser.optimise(factor, model_approx) ../../../../autofit/graphical/optimise.py:360: in optimise res.full_hess_inv) ../../../../autofit/graphical/factor_graphs/transform.py:196: in from_dense return cls(cho_factor(hess)) ../../../../../../anaconda3/envs/autolens/lib/python3.7/site-packages/scipy/linalg/decomp_cholesky.py:153: in cho_factor check_finite=check_finite) _ _ _ _ _...
The model uses recursion extensively to self-inspect, determining dimensionality and populating priors for example. However, some attributes associated with the model are there for its function as oppose to representing...
Analysis now compute a Samples object called latent_samples which comprises latent variables and corresponding sample data such as likelihood and weight. Latent sample data is output in the same way...
Visualise model and prior relationships using a graph. Also visualise factor and variable relationships using a graph.
During the workshop some objects did not print a user friendly representation.