MGDA

class torchjd.aggregation.MGDA(epsilon=0.001, max_iters=100)

Aggregator performing the gradient aggregation step of Multiple-gradient descent algorithm (MGDA) for multiobjective optimization. The implementation is based on Algorithm 2 of Multi-Task Learning as Multi-Objective Optimization.

Parameters:

• epsilon (float) – The value of \(\hat{\gamma}\) below which we stop the optimization.

• max_iters (int) – The maximum number of iterations of the optimization loop.

class torchjd.aggregation.MGDAWeighting(epsilon=0.001, max_iters=100)

Weighting giving the weights of MGDA.

Parameters:

• epsilon (float) – The value of \(\hat{\gamma}\) below which we stop the optimization.

• max_iters (int) – The maximum number of iterations of the optimization loop.