pythonsi.domain_adaptation

Domain adaptation methods with selective inference.

Classes

class pythonsi.domain_adaptation.OptimalTransportDA

Optimal Transport Domain Adaptation with selective inference support.

The optimal transport problem solved is:

\[\min_{T \in \mathcal{P}} \langle C, T \rangle\]

where \(\mathcal{P}\) is the set of transport plans with given marginals and \(C\) is the cost matrix between domains.

x_source_node

Source domain feature node

Type:

Data or None

y_source_node

Source domain label node

Type:

Data or None

x_target_node

Target domain feature node

Type:

Data or None

y_target_node

Target domain label node

Type:

Data or None

x_output_node

Adapted feature output node

Type:

Data

y_output_node

Adapted label output node

Type:

Data

interval

Feasible interval for the last inference call

Type:

list or None

x_output_data

Stored adapted features from last inference call

Type:

array-like or None

y_output_data

Stored adapted labels from last inference call

Type:

tuple or None

forward(xs: ndarray[tuple[Any, ...], dtype[floating]], ys: ndarray[tuple[Any, ...], dtype[floating]], xt: ndarray[tuple[Any, ...], dtype[floating]], yt: ndarray[tuple[Any, ...], dtype[floating]])

Solve optimal transport and construct adapted dataset.

Parameters:

• xs (array-like, shape (ns, d)) – Source domain features

• ys (array-like, shape (ns, 1)) – Source domain labels

• xt (array-like, shape (nt, d)) – Target domain features

• yt (array-like, shape (nt, 1)) – Target domain labels

Returns:

• x_tilde (array-like, shape (ns+nt, d)) – Adapted feature matrix

• y_tilde (array-like, shape (ns+nt, 1)) – Adapted label vector

• B (list of int) – Basic feasible solution indices

• c_features (array-like, shape (ns*nt, 1)) – Feature space cost matrix

• Omega (array-like, shape (ns+nt, ns+nt)) – Transformation matrix for adaptation

Notes

The adapted dataset is constructed as:

\[ \begin{align}\begin{aligned}\begin{split}\tilde{\mathbf{x}} = \Omega \begin{bmatrix} \mathbf{x}_s \\ \mathbf{x}_t \end{bmatrix}\end{split}\\\begin{split}\tilde{\mathbf{y}} = \Omega \begin{bmatrix} \mathbf{y}_s \\ \mathbf{y}_t \end{bmatrix}\end{split}\end{aligned}\end{align} \]

where \(\Omega\) incorporates the optimal transport plan.

inference(z: float) → Tuple[list, ndarray[tuple[Any, ...], dtype[floating]]]

Find feasible interval of the Optimal Transport for the parametrized data at z .

Parameters:

z (float) – Scalar parameter

Returns:

final_interval – Feasible interval [lower, upper] for z

Return type:

list

run(xs: Data, ys: Data, xt: Data, yt: Data) → Data

Configure domain adaptation with input data.

Parameters:

• xs (array-like, shape (ns, d)) – Source domain features

• ys (array-like, shape (ns, 1)) – Source domain labels

• xt (array-like, shape (nt, d)) – Target domain features

• yt (array-like, shape (nt, 1)) – Target domain labels

Returns:

• x_output_node (Data) – Node containing adapted features

• y_output_node (Data) – Node containing adapted labels

Examples

>>> ot_da = OptimalTransportDA()
>>> x_out, y_out = ot_da.run(xs, ys, xt, yt)
>>> adapted_x = x_out()

class pythonsi.domain_adaptation.RepresentationLearningDA(model: object, device: str = 'cpu')