bambi model management (bammm)

Estimating complex models can take time. bammm allows one to save estimated models and to later load them as required (without the need to re-estimate them).

Currently it is build to only work with (generalized) linear models in python that have been esimtated using bambi but in essence it can be used to save samples from any PyMC model.

Functionality

1. Creation and mangement of database of local models (simple JSON file)
2. Model (dependent variable, fixed and random effects) and estimation parameters (no chains, samples/chain, cores) specification.
3. Automatic build of regression equation (input to bambi) using mm.geerate_equation()
4. Model estimation using mm.estimate_lmm()
5. Saving (mm.save_model_info()) and updating (mm.update_model_entry()) of estimated model information.
6. Loading previously esimtated model (mm.estimate_lmm())

Usage

Setup

Install using pip:

pip install bammm-ozika

Load

import bammm.bammm as mm

Full example script can be found here.

Generate your local JSON database

Before first use on a project a local database needs to be generated. Specify a path for the database:

db_path = os.path.join(root_dir, "demo", "my_project", "databse_name.json")

Initialize database:

mm.models_init(db_path)

Fitting a new linear model

Load database

models = json.load(open(db_path, "r"))

Specify linear model. The idea is to have a specific model name and a "group/family" for each model. bammm will create a folder (model_family) and save data of individual models in there using pickle.

# load data that come with bambi
data = bmb.load_data("sleepstudy")

model_family = "sleepstudy" # this will be used as a folder name to host the models
model_identifier = "maximum_model"
# dependent variable
mod["lmm"]["dep_var"] = "Reaction" # Reaction time
# fixed effects
mod["lmm"]["fxeff"] = ["Days"] # longitudinal data set
# random effects
mod["lmm"]["rneff"] = ["Days|Subject"]
# build equation
mod["lmm"]["eq"] = mm.generate_equation(mod["lmm"]["dep_var"], mod["lmm"]["fxeff"], mod["lmm"]["rneff"])

Specify estimation parameters.

# fitting information
mod["est"]["nchains"] = 2
mod["est"]["nsamples"] = 4000
mod["est"]["ncores"] = 2 # number of cores to be useds in fitting

Specify paths and create relevant strings.

mod = mm.prepare_fit(mod, model_family, model_identifier, models_path)
# save model (it's a good idea to load/save the DB often, especially if one runs multiple models at the same time)
models[mod["name"]] = mod
mm.save_model_info(models, db_path)

Estimate model. This also automatically saves the data to the location specified in models_path

mod, results, m = mm.estimate_lmm(mod, data, override=0)
mm.update_model_entry(models, mod, db_path)

Load previously fitted model

It is rather simple

# load database
models = json.load(open(db_path, "r"))
model_name = "sleepstudy_maximum_model_2_4000" # model estimated above
mod = models[model_name]
# load model
mod, results, m = mm.estimate_lmm(mod, [], override=0)

I am not saying that his is the best way to achieve model saving/loading, but it seems to work well (at least for the poor practitioner of statistics that is me).