import_model
Import surrogate model from archive file
Specification
Alias: None
Arguments: None
Child Keywords:
Required/Optional |
Description of Group |
Dakota Keyword |
Dakota Keyword Description |
---|---|---|---|
Optional |
User-customizable portion of exported/imported surrogate model filenames |
||
Required (Choose One) |
Surrogate Import Format |
Surrogate model plain-text archive file format |
|
Surrogate model binary archive file format |
Description
Imports a surrogate model from a file in binary_archive
or
text_archive
format, typically instead of constructing it from
generated or imported data. Importing can result in significant time
savings with some models, such as Gaussian processes. The file from
which to import is further specified with the child keywords
filename_prefix
and binary_archive
or text_archive
.
Default Behavior
When used in the context of an iteratively adapted or rebuilt
surrogate, the imported model will only replace the initial surrogate
model. Subsequent surrogate model builds require imported build points
or a design of experiments method specified through
dace_method_pointer
.
Usage Tips
When importing a surrogate model, it is crucial that the
global
surrogate model part of the Dakota input file
be identical for export and import, except for changing
export_model
and its child keywords to those needed for
import_model
. Any other keywords such as specifying a dace_iterator or
imported points must remain intact to satisfy internal surrogate
constructor requirements.
All variables over which the surrogate model was built (typically the active variables) when exported must be present in the Dakota study that imports it. The variables in the importing study will be matched by descriptor to those in the surrogate model. This allows the order and type of variables to change between export and import, for example to fix unimportant parameters as state variables even if they were previously active design variables.
For example, suppose the original Dakota study for the surrogate
build had active uniform uncertain variables: a, b, c, d, and inactive
state variables s, t, and the surrogate gets built over a, b, c,
d. The follow-on study could perform design optimization over active
design variables b and d, with inactive state variables a, c, s, t,
held at fixed initial_state
values. In this case the values of a,
b, c, d will get mapped to the surrogate and s, t dropped when
evaluating it. So it is important the state values for s and t haven’t
changed in some way that invalidates the surrogate.
While it might be possible to import surrogates externally-generated
using the surfpack
binary or the experimental dakota
.surrogates
Python module, it is untested and not the primary use case for this
capability.
Examples
The following model block exports a Gaussian process after it’s built:
model
id_model = 'SURR_M'
surrogate global
dace_method_pointer = 'DACE'
experimental_gaussian_process
metrics 'root_mean_squared'
cross_validation folds 5
export_model
filename_prefix = 'gp_export'
formats = binary_archive
export_approx_points_file 'gp_values.dat'
export_approx_variance_file 'gp_variance.dat'
And this model block imports it using a nearly identical model specification:
model
id_model = 'SURR_M'
surrogate global
dace_method_pointer = 'DACE'
experimental_gaussian_process
metrics 'root_mean_squared'
cross_validation folds 5
import_model
filename_prefix = 'gp_export'
binary_archive
export_approx_points_file 'gp_values.dat'
export_approx_variance_file 'gp_variance.dat'