skada.deep.DANLoss

class skada.deep.DANLoss(sigmas=None, eps=1e-07)[source]

Loss DAN

This loss reduces the MMD distance between source features and target features. From [14].

Parameters:

sigmasarray-like, optional (default=None): The sigmas for the Gaussian kernel.
epsfloat, default=1e-7: Small constant added to median distance calculation for numerical stability.

References

[14]

Mingsheng Long et. al. Learning Transferable Features with Deep Adaptation Networks. In ICML, 2015.

forward(features_s, features_t, **kwargs)[source]: Compute the domain adaptation loss