Aligned-MTL

class torchjd.aggregation.AlignedMTL(pref_vector=None)

Aggregator as defined in Algorithm 1 of Independent Component Alignment for Multi-Task Learning.

Parameters:

pref_vector (Tensor | None) – The preference vector to use. If not provided, defaults to \(\begin{bmatrix} \frac{1}{m} & \dots & \frac{1}{m} \end{bmatrix}^T \in \mathbb{R}^m\).

Note

This implementation was adapted from the official implementation.

class torchjd.aggregation.AlignedMTLWeighting(pref_vector=None)

Weighting giving the weights of AlignedMTL.

Parameters:

pref_vector (Tensor | None) – The preference vector to use. If not provided, defaults to \(\begin{bmatrix} \frac{1}{m} & \dots & \frac{1}{m} \end{bmatrix}^T \in \mathbb{R}^m\).