SigOpt-Lite is an open source tool for locally running a lightweight version of SigOpt. Sigopt-Lite contains all the computation elements of SigOpt-Server, but bypasses the need to set up servers and Docker. SigOpt-Lite gives access to the SigOpt Core Module functionality; to learn more about that, visit the SigOpt documentation. To learn about how to host your own SigOpt server, visit the SigOpt-Server repository or the official SigOpt OSS site (sigopt.org).
Executing the following command will install both the sigopt client (to interact with SigOpt) and the sigoptlite driver (to run the computations locally).
pip install 'sigopt[lite]'After that is installed, a SigOpt connection object can be created in Python with the following commands.
from sigopt import Connection
conn = Connection(driver="lite")From this point, SigOpt Core module functionality can be accessed through conn, following the same patterns as detailed in the SigOpt documentation.
Here is an example of using SigOpt-Lite.
from sigopt import Connection
conn = Connection(driver="lite")
experiment_meta = dict(
parameters=[
dict(name="x0", type="double", bounds=dict(min=0, max=1)),
dict(name="x1", type="int", bounds=dict(min=0, max=10)),
],
metrics=[
dict(
name="f",
objective="minimize",
)
],
observation_budget=20,
)
e = conn.experiments().create(**experiment_meta)
suggestion = conn.experiments(e.id).suggestions().create()
conn.experiments(e.id).observations().create(
suggestion=suggestion.id,
values=[{"name": "f", "value": 2}],
)More SigOpt examples can be found at the Examples GitHub Repo. Note, some of those examples will not be compatible with SigOpt-Lite, including those using the AI Module.
SigOpt-Lite supports an alternative computation mode that can generate suggestions with reduced computation burden. You can activate this compute mode at Connection object instantiation.
from sigopt import Connection
conn = Connection(driver="lite", compute_mode="kde_only")To support easy use of SigOpt-Lite, the calling sequences involving the Connection object are meant to match those for connecting with api.sigopt.com (as is directed in the SigOpt documentation). Still, in pursuit of simplicitly and a minimalistic installation, SigOpt-Lite does have some limitations.
Sigopt-Lite functionality matches the core module of our hosted SigOpt platform, but only certain API endpoints are available.
| Objects | Endpoints |
|---|---|
| Experiment | Create, Detail |
| Suggestion | Create |
| Observation | Create, List |
| Best Assignments | Detail |
The following behavior limitations are strictly enforced in SigOpt-Lite.
- SigOpt-Lite can only run one Experiment at a time: the Experiment ID for the sole experiment is
"-1". - SigOpt-Lite supports no parallelism: an observation must be reported for the open suggestion before creating another unique suggestion. Parallel suggestions can be accessed through an account at sigopt.com or through the open source SigOpt-Server.
- SigOpt-Lite does not support Grid Search.
The following set of soft limitations are recommended by the SigOpt research team to ensure a pleasant experience using SigOpt-Lite.
- Maximum number of parameters <= 100
- Maximum number of observations <= 10,000
- Maximum categorical breadth (total number of categorical values across all categorical parameters <= 10
- Maximum number of parameter constraints <= 10
- Maximum number of unique integer constrained variables <= 10
- Maximum number of metric constraints <= 4
- Maximum conditionals breadth (total number of conditional values across all conditional parameters <= 6