GitHub

Saving and Loading LikelihoodModels

Models can be saved easily using BSON.jl as they are defined using a Julia struct.

The only thing to be careful of is ensuring that empty dataframe rows within the model are removed prior to saving using trim_model_dfs! (as they include undefined references). 

using LikelihoodBasedProfileWiseAnalysis

# function definitions and initialisation...
model = 
trim_model_dfs!(model)
using BSON: @save

@save "mymodel.bson" model

To load model in a new session/file:

using LikelihoodBasedProfileWiseAnalysis
using BSON: @load

# function definitions....

@load "mymodel.bson" model

Potential Issues When Loading

There are a couple of things to watch out for if the model saved had functions defined in model.core:

  • The log-likelihood function (and if defined, the prediction and error functions) must be defined with the same name and be available in the scope we are loading model in. Functions used within model.core.optimizationsettings must also be defined in the scope we are loading model in (i.e. by loading LikelihoodBasedProfileWiseAnalysis).
  • The variable name model has when saved is the same name it needs to be loaded with.

Fixing Issues

The first of these issues we can get around by converting our CoreLikelihoodModel to a BaseLikelihoodModel before saving. The only difference between these two structs is that BaseLikelihoodModel doesn't contain fields for the log-likelihood, prediction and error functions. This means we can load a saved model without needing those functions defined in the local scope, which may be useful for workflows where the computation is performed in one file and plotting of outputs is performed in another file.

We can use remove_functions_from_core! to perform this task, pulling out the original model.core so we can put it back into model later if desired: 

core_original = remove_functions_from_core!(model)
@save "mymodel.bson" model

# restore the core that has the functions if desired
model.core = core_original

If we want to add the log-likelihood function to this loaded version of the model we can use add_loglikelihood_function! after loading LikelihoodBasedProfileWiseAnalysis.

using LikelihoodBasedProfileWiseAnalysis
# log-likelihood function function definition
function loglikefunction(θ, data); return ... end
add_loglikelihood_function!(model, loglikefunction)
Missing log-likelihood function

Trying to use a profile-related function will result in an error if the log-likelihood function is not defined in model.core.

The prediction function can be added in the same fashion using add_prediction_function!. However, the log-likelihood function must have been added first.

The error function can also be added in the same fashion using add_error_function!. However, the log-likelihood must have been added first.

Saving and Loading Functions

LikelihoodBasedProfileWiseAnalysis.add_loglikelihood_function! — Function
add_loglikelihood_function!(model::LikelihoodModel, loglikefunction::Function; 
    optimizationsettings::OptimizationSettings=default_OptimizationSettings())

Adds a log-likelihood function, loglikefunction, to model as well as optimization settings, optimizationsettings, using default_OptimizationSettings.

Requirements for loglikefunction: loglikelihood function which takes two arguments, θ and data, in that order, where θ is a vector containing the values of each parameter in θnames and data is a Tuple or NamedTuple containing any additional information required by the log-likelihood function, such as the time points to be evaluated at.

source