Open Set Domain Adaptation by Backpropagation (OSBP)

class dalib.adaptation.osbp.UnknownClassBinaryCrossEntropy(t=0.5)

Binary cross entropy loss to make a boundary for unknown samples, proposed by Open Set Domain Adaptation by Backpropagation (ECCV 2018).

Given a sample on target domain \(x_t\) and its classifcation outputs \(y\), the binary cross entropy loss is defined as

\[L_{adv}(x_t) = -t log(p(y=C+1|x_t)) - (1-t)log(1-p(y=C+1|x_t))\]

where t is a hyper-parameter and C is the number of known classes.

Parameters

t (float) – Predefined hyper-parameter. Default: 0.5

Inputs:

• y (tensor): classification outputs (before softmax).

Shape:

• y: \((minibatch, C+1)\) where C is the number of known classes.

• Outputs: scalar