Saving And loading Controllers Using Dill¶

Basic usage¶

Workflow managers can be saved to and loaded using dill.

For instance, you could generate a basic Controller object in an interactive python session as follows:

In [ ]:

# Uncomment the following line when running on Google Colab
# !pip install autora

# Uncomment the following line when running on Google Colab # !pip install autora

In [ ]:

import autora.workflow
import dill

controller = autora.workflow.Controller()

with open("default-controller.dill", "wb") as file:
    dill.dump(controller, file)

import autora.workflow import dill controller = autora.workflow.Controller() with open("default-controller.dill", "wb") as file: dill.dump(controller, file)

The same Controller state can be loaded in a separate session as follows:

In [ ]:

import dill

with open("default-controller.dill", "rb") as file:
    reloaded_controller = dill.load(file)

reloaded_controller

import dill with open("default-controller.dill", "rb") as file: reloaded_controller = dill.load(file) reloaded_controller

Out[ ]:

<autora.workflow.controller.Controller at 0x12a49b400>

Executors¶

A Controller for real use always includes at least one (and usually more than one) Executor.

In this example, the Controller includes:

• an experimentalist, which suggests a random sample across a particular problem domain,
• an experiment runner, which makes synthetic observations of a polynomial function,
• and a theorist, which is a scikit-learn regressor.

In [ ]:

import numpy as np
from sklearn.model_selection import GridSearchCV
from sklearn.pipeline import make_pipeline as make_theorist_pipeline
from sklearn.preprocessing import PolynomialFeatures
from sklearn.linear_model import LinearRegression

import autora.workflow
from autora.experimentalist.pipeline import make_pipeline as make_experimentalist_pipeline
from autora.variable import VariableCollection, Variable


rng = np.random.default_rng(180)

experimentalist = make_experimentalist_pipeline(
    [np.linspace, rng.choice],
    params={"linspace": {"start": [-10], "stop": [+10], "num":1001},
            "choice": {"size": 10}}
)

def experiment_runner(x, coefs=[2.0, 3.0, 1.0], noise_std=0.1):
    """Simple experiment."""
    x_ = np.array(x)
    y_ = (
        coefs[0] * x_**2.0
        + coefs[1] * x_
        + coefs[2]
        + rng.normal(0.0, noise_std, size=x_.shape)
    )
    return y_

theorist = GridSearchCV(make_theorist_pipeline(PolynomialFeatures(), LinearRegression()),
                        param_grid={"polynomialfeatures__degree": [0, 1, 2, 3, 4]}, scoring="r2")

controller = autora.workflow.Controller(
    variables=VariableCollection(independent_variables=[Variable("x")], dependent_variables=[Variable("y")]),
    experiment_runner=experiment_runner,
    experimentalist=experimentalist,
    theorist=theorist,
)

with open("simple-controller.dill", "wb") as file:
    dill.dump(controller, file)

import numpy as np from sklearn.model_selection import GridSearchCV from sklearn.pipeline import make_pipeline as make_theorist_pipeline from sklearn.preprocessing import PolynomialFeatures from sklearn.linear_model import LinearRegression import autora.workflow from autora.experimentalist.pipeline import make_pipeline as make_experimentalist_pipeline from autora.variable import VariableCollection, Variable rng = np.random.default_rng(180) experimentalist = make_experimentalist_pipeline( [np.linspace, rng.choice], params={"linspace": {"start": [-10], "stop": [+10], "num":1001}, "choice": {"size": 10}} ) def experiment_runner(x, coefs=[2.0, 3.0, 1.0], noise_std=0.1): """Simple experiment.""" x_ = np.array(x) y_ = ( coefs[0] * x_**2.0 + coefs[1] * x_ + coefs[2] + rng.normal(0.0, noise_std, size=x_.shape) ) return y_ theorist = GridSearchCV(make_theorist_pipeline(PolynomialFeatures(), LinearRegression()), param_grid={"polynomialfeatures__degree": [0, 1, 2, 3, 4]}, scoring="r2") controller = autora.workflow.Controller( variables=VariableCollection(independent_variables=[Variable("x")], dependent_variables=[Variable("y")]), experiment_runner=experiment_runner, experimentalist=experimentalist, theorist=theorist, ) with open("simple-controller.dill", "wb") as file: dill.dump(controller, file)

Later, we can reload the same Controller and run the cycle:

In [ ]:

with open("simple-controller.dill", "rb") as file:
    controller_loaded = dill.load(file)

controller_loaded.run(num_steps=10)

with open("simple-controller.dill", "rb") as file: controller_loaded = dill.load(file) controller_loaded.run(num_steps=10)

Out[ ]:

<autora.workflow.controller.Controller at 0x12a4e7430>

In [ ]:

controller_loaded.state.models[-1].best_params_

controller_loaded.state.models[-1].best_params_

Out[ ]:

{'polynomialfeatures__degree': 2}

In [ ]:

import matplotlib.pyplot as plt
x = np.linspace([-10], [10], 100)
plt.plot(x, experiment_runner(x, noise_std=0.),
         label="ground truth",
         lw=10, alpha=0.2, c="black")
plt.plot(x, controller_loaded.state.models[-1].best_estimator_.predict(x),
         label="prediction",
         lw=1, alpha=1, c="red")
plt.legend()

import matplotlib.pyplot as plt x = np.linspace([-10], [10], 100) plt.plot(x, experiment_runner(x, noise_std=0.), label="ground truth", lw=10, alpha=0.2, c="black") plt.plot(x, controller_loaded.state.models[-1].best_estimator_.predict(x), label="prediction", lw=1, alpha=1, c="red") plt.legend()

Out[ ]:

<matplotlib.legend.Legend at 0x12f5cdbd0>