pydantic-sweep

pydantic_sweep is a library to programmatically, safely and flexibly define complex parameter sweeps over pydantic models in Python.

Highlights:

• Specify parameter sweeps in Python
• Flexibility: specify complex parameter combinations by chaining simple functional operations
• Safety checks for parameter combinations (get meaningful errors early)
• pydantic field validation

For example, the following code will instantiate models with (x=5, sub=Sub1(s=1) and (x=6, sub=Sub1(s=2) and try each of those with seed values of seed=43 and seed=44, leading to four different configurations:

import pydantic_sweep as ps

class Sub1(ps.BaseModel):
    s: int = 5

class Sub2(ps.BaseModel):
    y: str = "hi"

class Model(ps.BaseModel):
    seed: int = 42
    x: int = 5
    sub: Sub1 | Sub2

# We want to test across two seeds
configs = ps.config_product(
    ps.field("seed", [43, 44]),
    # And two specific Sub1 and `x` combinations
    ps.config_zip(
        ps.field("x", [ps.DefaultValue, 6]),
        ps.field("sub.s", [1, 2]),
    )
)
# This includes safety checks that Sub1 / Sub2 are uniquely specified
models = ps.initialize(Model, configs)

# The code above is equivalent to
models_manual = [
    Model(seed=43, sub=Sub1(s=1)),
    Model(seed=43, x=6, sub=Sub1(s=2)), 
    Model(seed=44, sub=Sub1(s=1)),
    Model(seed=44, x=6, sub=Sub1(s=2)), 
]
assert models == models_manual

# We can also check that we didn't accidentally duplicate a setting
ps.check_unique(models)

While in this toy example, manually specifying the combinations may still be viable, the library allows infinitely combining different configs and sub-models, making it a powerful tool for large-scale experiment definition. To learn mode about the capabilities of the library please visit the documentation page.

Installation

You can install the library by checking out the repo and running

pip install 'pydantic-sweep'

License

The main code-base is licensed under MPL-2.0 a weak copy-left license that allows commercial use. See the license file for the exact clauses and this FAQ for a high-level description.

An exception from this are the documentation in the docs and example folders as well as this README file, which are licensed under the 0BSD: a highly permissive license that does not require attribution. That way, you are free to copy & paste example code into your use-case. See here for a high-level description.