Configuration

Gryffin provides a flexible configuration interface, it accepts either a path to a config.json file or a python dict.

gryffin = Gryffin(config_file='/path/to/your/config.json')

gryffin = Gryffin(config_dict={})

Gryffin exposes 5 configurable modules, general, database, model, parameters and objectives.

{
    "general": {},
    "database": {},
    "model": {},
    "parameters": [],
    "objectives": []
}

config = {
    "general": {},
    "database": {},
    "model": {},
    "parameters": [],
    "objectives": []

}

General Configuration

Parameter [type]	Definition
num_cpus [int | string]	Number of CPUs to use, options are a number or ‘all’ (default: 1)
boosted [bool]	Use kernel boosting (default: True)
caching [bool]	Use kernel caching (default: True)
auto_desc_gen [bool]	Use automatic descriptor generation (default: False)
batches [int]	(default: 1)
sampling_strategies [int]	(default: 2)
softness [float]	Softness of Chimera for multiobj optimizations (default: 0.001)
feas_approach [string]	Approach to unknown feasibility constraints, options are ‘fwa’ (feasibility-weighted acquisition), ‘fca’ (feasibility-constrained acquisition) or ‘fia’ (feasibility-interpolated acquisition). (default: ‘fwa’)
feas_param [int]	Sensitivity to feasibility constraints (default: 1)
dist_param [float]	Factor modulating density-based penalty in sample selector (default: 0.5)
random_seed [None | int]	Set random seed (default: None)
save_database [bool]	(default: False)
aquisition_optimizer [string]	Set aquisition optimization method, options are ‘adam’ or ‘genetic’ (default: ‘adam’)
obj_transform [None | string]	Set objective transform, options are None, ‘sqrt’, ‘cbrt’ or ‘square’ (default: ‘sqrt’)
num_random_samples [int]	Number of samples per dimension to sample when optimizing acquisition function (default: 200)
reject_tol [int]	Tolerance in rejection sampling, relevant when known constraints or fca used (default: 1000)
vebosity [int]	Set verbosity level, from 0 to 5. 0: FATAL, 1: ERROR, 2: WARNING, 3: STATS, 4: INFO, 5: DEBUG (default: 4)

Database Configuration

Parameter [type]	Definition
format [string]	(default: ‘sqlite’)
path [int]	(default: ‘./SearchProgress’)
log_observations [bool]	(default: True)
log_runtimes [bool]	(default: True)

Model Configuration

Parameter	Definition
num_epochs [int]	Number of training epochs (default: 2e3)
learning_rate [float]	Model learning rate (default: 5e-2)
num_draws [int]	(default: 1e3)
num_layers [int]	Set the number of hidden layers in the model (default: 3)
hidden_shape [int]	Set the dimensionality of the hidden layers (default: 6)
weight_loc [float]	(default: 0.0)
weight_scale [float]	(default: 1.0)
bias_loc [float]	(default: 0.0)
bias_scale [float]	(default: 1.0)

Parameters Configuration

Gryffin supports 3 parameter types, continuous, discrete and categorical. Each parameter is configured as elements of the root level parameters list:

{
    "parameters": [
            {}
    ]
}

Continuous Parameters:

Parameter [type]	Definition
name [string]	Human-readable parameter name
type [string]	Selects parameter type, either ‘continuous’, ‘discrete’ or ‘categorical’
low [float]	Lower bound of continuous parameter
high [float]	Upper bound of continuous parameter. Note: high must be larger than low.
periodic [bool]	Boolean flag indicating that the parameter is periodic

Discrete Parameters:

Parameter [type]	Definition
name [string]	Human-readable parameter name
type [string]	Selects parameter type, either ‘continuous’, ‘discrete’ or ‘categorical’
low [float]	Lower bound of discrete parameter
high [float]	Upper bound of continuous parameter. Note: high must be larger than low.
options [List[]]	ToDo: Need explanation of options
descriptors [List[]]	ToDo: Need explanation of descriptors

Categorical Parameters:

Parameter [type]	Definition
name [string]	Human-readable parameter name
type [string]	Selects parameter type, either ‘continuous’, ‘discrete’ or ‘categorical’
options [List[]]	ToDo: Need explanation of options
descriptors [List[]]	ToDo: Need explanation of descriptors
category_details [List[]]	ToDo: Need explanation of category_details

Objective Configuration

Each objective is configured as elements of the root level objective list:

{
    "objectives": [
            {},
    ]
}

Parameter [type]	Definition
name [string]	Human-readable parameter name
goal [string]	Optimization objective, options are ‘min’ or ‘max’
tolerance [float]	Termination tolerance on parameter changes
absolute [bool]	Boolean flag indicating if objective is absolute