Pipeline

from __future__ import annotations

from numpy import allclose
from numpy import linspace
from numpy import newaxis

from gemseo.mlearning.transformers.pipeline import Pipeline
from gemseo.mlearning.transformers.scaler.scaler import Scaler

To illustrate the pipeline, we consider very simple data:

data = linspace(0, 1, 100)[:, newaxis]

First, we create a pipeline of two transformers: the first one shifts the data while the second one reduces their amplitude.

pipeline = Pipeline(transformers=[Scaler(offset=1), Scaler(coefficient=2)])

Then, we fit this Pipeline to the data, transform them and compute the Jacobian:

transformed_data = pipeline.fit_transform(data)
transformed_jac_data = pipeline.compute_jacobian(data)

Lastly, we can do the same with two scalers:

shifter = Scaler(offset=1)
shifted_data = shifter.fit_transform(data)
scaler = Scaler(coefficient=2)
data_shifted_then_scaled = scaler.fit_transform(shifted_data)
jac_shifted_then_scaled = scaler.compute_jacobian(
    shifted_data
) @ shifter.compute_jacobian(data)

and verify that the results are identical:

assert allclose(transformed_data, data_shifted_then_scaled)
assert allclose(transformed_jac_data, jac_shifted_then_scaled)

Note that a Pipeline can compute the Jacobian as long as the Transformer instances that make it up can do so.